2020-10-30 16:25:06 +08:00
|
|
|
# Copyright (C) 2019-2020 Zilliz. All rights reserved.
|
|
|
|
#
|
|
|
|
# Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance
|
|
|
|
# with the License. You may obtain a copy of the License at
|
|
|
|
#
|
|
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
#
|
|
|
|
# Unless required by applicable law or agreed to in writing, software distributed under the License
|
|
|
|
# is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express
|
|
|
|
# or implied. See the License for the specific language governing permissions and limitations under the License.
|
|
|
|
|
2022-03-16 16:51:22 +08:00
|
|
|
GO ?= go
|
|
|
|
PWD := $(shell pwd)
|
2022-06-28 19:58:25 +08:00
|
|
|
GOPATH := $(shell $(GO) env GOPATH)
|
|
|
|
SHELL := /bin/bash
|
2022-03-16 16:51:22 +08:00
|
|
|
OBJPREFIX := "github.com/milvus-io/milvus/cmd/milvus"
|
2020-10-30 16:25:06 +08:00
|
|
|
|
|
|
|
INSTALL_PATH := $(PWD)/bin
|
|
|
|
LIBRARY_PATH := $(PWD)/lib
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
PGO_PATH := $(PWD)/configs/pgo
|
2022-02-09 14:27:46 +08:00
|
|
|
OS := $(shell uname -s)
|
|
|
|
mode = Release
|
2023-10-11 20:45:35 +08:00
|
|
|
|
|
|
|
use_disk_index = OFF
|
2023-10-12 19:25:35 +08:00
|
|
|
ifdef disk_index
|
|
|
|
use_disk_index = ${disk_index}
|
2023-10-11 20:45:35 +08:00
|
|
|
endif
|
|
|
|
|
|
|
|
use_asan = OFF
|
|
|
|
ifdef USE_ASAN
|
|
|
|
use_asan =${USE_ASAN}
|
2023-01-29 10:59:49 +08:00
|
|
|
endif
|
2023-10-11 20:45:35 +08:00
|
|
|
|
2023-12-18 12:04:42 +08:00
|
|
|
use_dynamic_simd = ON
|
2023-07-13 16:22:30 +08:00
|
|
|
ifdef USE_DYNAMIC_SIMD
|
|
|
|
use_dynamic_simd = ${USE_DYNAMIC_SIMD}
|
|
|
|
endif
|
2024-02-20 19:16:52 +08:00
|
|
|
|
|
|
|
use_opendal = OFF
|
|
|
|
ifdef USE_OPENDAL
|
|
|
|
use_opendal = ${USE_OPENDAL}
|
|
|
|
endif
|
2023-09-05 17:39:48 +08:00
|
|
|
# golangci-lint
|
2023-11-21 10:04:21 +08:00
|
|
|
GOLANGCI_LINT_VERSION := 1.55.2
|
2023-09-05 17:39:48 +08:00
|
|
|
GOLANGCI_LINT_OUTPUT := $(shell $(INSTALL_PATH)/golangci-lint --version 2>/dev/null)
|
|
|
|
INSTALL_GOLANGCI_LINT := $(findstring $(GOLANGCI_LINT_VERSION), $(GOLANGCI_LINT_OUTPUT))
|
|
|
|
# mockery
|
|
|
|
MOCKERY_VERSION := 2.32.4
|
|
|
|
MOCKERY_OUTPUT := $(shell $(INSTALL_PATH)/mockery --version 2>/dev/null)
|
|
|
|
INSTALL_MOCKERY := $(findstring $(MOCKERY_VERSION),$(MOCKERY_OUTPUT))
|
2023-09-21 09:45:27 +08:00
|
|
|
# gci
|
|
|
|
GCI_VERSION := 0.11.2
|
|
|
|
GCI_OUTPUT := $(shell $(INSTALL_PATH)/gci --version 2>/dev/null)
|
|
|
|
INSTALL_GCI := $(findstring $(GCI_VERSION),$(GCI_OUTPUT))
|
|
|
|
# gofumpt
|
|
|
|
GOFUMPT_VERSION := 0.5.0
|
|
|
|
GOFUMPT_OUTPUT := $(shell $(INSTALL_PATH)/gofumpt --version 2>/dev/null)
|
|
|
|
INSTALL_GOFUMPT := $(findstring $(GOFUMPT_VERSION),$(GOFUMPT_OUTPUT))
|
2024-03-22 10:29:06 +08:00
|
|
|
# gotestsum
|
|
|
|
GOTESTSUM_VERSION := 1.11.0
|
|
|
|
GOTESTSUM_OUTPUT := $(shell $(INSTALL_PATH)/gotestsum --version 2>/dev/null)
|
|
|
|
INSTALL_GOTESTSUM := $(findstring $(GOTESTSUM_VERSION),$(GOTESTSUM_OUTPUT))
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2023-09-22 09:59:26 +08:00
|
|
|
index_engine = knowhere
|
|
|
|
|
2022-10-16 20:49:27 +08:00
|
|
|
export GIT_BRANCH=master
|
2022-10-12 14:33:23 +08:00
|
|
|
|
2023-09-20 13:59:23 +08:00
|
|
|
ifeq (${ENABLE_AZURE}, false)
|
|
|
|
AZURE_OPTION := -Z
|
|
|
|
endif
|
|
|
|
|
2022-10-17 19:19:25 +08:00
|
|
|
milvus: build-cpp print-build-info
|
2022-10-12 14:33:23 +08:00
|
|
|
@echo "Building Milvus ..."
|
|
|
|
@source $(PWD)/scripts/setenv.sh && \
|
|
|
|
mkdir -p $(INSTALL_PATH) && go env -w CGO_ENABLED="1" && \
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
GO111MODULE=on $(GO) build -pgo=$(PGO_PATH)/default.pgo -ldflags="-r $${RPATH} -X '$(OBJPREFIX).BuildTags=$(BUILD_TAGS)' -X '$(OBJPREFIX).BuildTime=$(BUILD_TIME)' -X '$(OBJPREFIX).GitCommit=$(GIT_COMMIT)' -X '$(OBJPREFIX).GoVersion=$(GO_VERSION)'" \
|
2023-07-13 15:34:33 +08:00
|
|
|
-tags dynamic -o $(INSTALL_PATH)/milvus $(PWD)/cmd/main.go 1>/dev/null
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2023-03-21 17:55:57 +08:00
|
|
|
milvus-gpu: build-cpp-gpu print-gpu-build-info
|
2023-03-14 23:21:56 +08:00
|
|
|
@echo "Building Milvus-gpu ..."
|
|
|
|
@source $(PWD)/scripts/setenv.sh && \
|
|
|
|
mkdir -p $(INSTALL_PATH) && go env -w CGO_ENABLED="1" && \
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
GO111MODULE=on $(GO) build -pgo=$(PGO_PATH)/default.pgo -ldflags="-r $${RPATH} -X '$(OBJPREFIX).BuildTags=$(BUILD_TAGS_GPU)' -X '$(OBJPREFIX).BuildTime=$(BUILD_TIME)' -X '$(OBJPREFIX).GitCommit=$(GIT_COMMIT)' -X '$(OBJPREFIX).GoVersion=$(GO_VERSION)'" \
|
2023-07-13 15:34:33 +08:00
|
|
|
-tags dynamic -o $(INSTALL_PATH)/milvus $(PWD)/cmd/main.go 1>/dev/null
|
2023-03-14 23:21:56 +08:00
|
|
|
|
2020-10-30 16:25:06 +08:00
|
|
|
get-build-deps:
|
|
|
|
@(env bash $(PWD)/scripts/install_deps.sh)
|
|
|
|
|
2021-12-07 17:59:02 +08:00
|
|
|
# attention: upgrade golangci-lint should also change Dockerfiles in build/docker/builder/cpu/<os>
|
2020-12-09 20:07:27 +08:00
|
|
|
getdeps:
|
2023-01-29 10:59:49 +08:00
|
|
|
@mkdir -p $(INSTALL_PATH)
|
2023-09-05 17:39:48 +08:00
|
|
|
@if [ -z "$(INSTALL_GOLANGCI_LINT)" ]; then \
|
2023-11-21 10:04:21 +08:00
|
|
|
echo "Installing golangci-lint into ./bin/" && curl -sSfL https://raw.githubusercontent.com/golangci/golangci-lint/master/install.sh | sh -s -- -b $(INSTALL_PATH) v${GOLANGCI_LINT_VERSION} ; \
|
2023-09-05 17:39:48 +08:00
|
|
|
else \
|
|
|
|
echo "golangci-lint v@$(GOLANGCI_LINT_VERSION) already installed"; \
|
|
|
|
fi
|
|
|
|
@if [ -z "$(INSTALL_MOCKERY)" ]; then \
|
|
|
|
echo "Installing mockery v$(MOCKERY_VERSION) to ./bin/" && GOBIN=$(INSTALL_PATH) go install github.com/vektra/mockery/v2@v$(MOCKERY_VERSION); \
|
|
|
|
else \
|
|
|
|
echo "Mockery v$(MOCKERY_VERSION) already installed"; \
|
2023-09-11 20:43:17 +08:00
|
|
|
fi
|
2024-03-22 10:29:06 +08:00
|
|
|
@if [ -z "$(INSTALL_GOTESTSUM)" ]; then \
|
2024-04-12 19:17:18 +08:00
|
|
|
echo "Install gotestsum v$(GOTESTSUM_VERSION) to ./bin/" && GOBIN=$(INSTALL_PATH) go install gotest.tools/gotestsum@v$(GOTESTSUM_VERSION); \
|
2024-03-22 10:29:06 +08:00
|
|
|
else \
|
|
|
|
echo "gotestsum v$(GOTESTSUM_VERSION) already installed";\
|
|
|
|
fi
|
2020-12-09 20:07:27 +08:00
|
|
|
|
2020-12-31 17:25:45 +08:00
|
|
|
tools/bin/revive: tools/check/go.mod
|
|
|
|
cd tools/check; \
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
$(GO) build -pgo=$(PGO_PATH)/default.pgo -o ../bin/revive github.com/mgechev/revive
|
2020-12-31 17:25:45 +08:00
|
|
|
|
2020-11-23 19:57:05 +08:00
|
|
|
cppcheck:
|
2024-02-22 19:40:53 +08:00
|
|
|
@#(env bash ${PWD}/scripts/core_build.sh -l)
|
|
|
|
@(env bash ${PWD}/scripts/check_cpp_fmt.sh)
|
2020-11-18 10:07:05 +08:00
|
|
|
|
|
|
|
|
2020-10-30 16:25:06 +08:00
|
|
|
fmt:
|
2020-12-29 17:19:44 +08:00
|
|
|
ifdef GO_DIFF_FILES
|
|
|
|
@echo "Running $@ check"
|
|
|
|
@GO111MODULE=on env bash $(PWD)/scripts/gofmt.sh $(GO_DIFF_FILES)
|
|
|
|
else
|
2020-10-30 16:25:06 +08:00
|
|
|
@echo "Running $@ check"
|
2020-11-18 10:07:05 +08:00
|
|
|
@GO111MODULE=on env bash $(PWD)/scripts/gofmt.sh cmd/
|
|
|
|
@GO111MODULE=on env bash $(PWD)/scripts/gofmt.sh internal/
|
2023-02-07 19:24:30 +08:00
|
|
|
@GO111MODULE=on env bash $(PWD)/scripts/gofmt.sh tests/integration/
|
2020-12-03 19:00:11 +08:00
|
|
|
@GO111MODULE=on env bash $(PWD)/scripts/gofmt.sh tests/go/
|
2023-09-08 09:51:17 +08:00
|
|
|
@GO111MODULE=on env bash $(PWD)/scripts/gofmt.sh pkg/
|
2020-12-29 17:19:44 +08:00
|
|
|
endif
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2023-09-21 09:45:27 +08:00
|
|
|
lint-fix: getdeps
|
|
|
|
@mkdir -p $(INSTALL_PATH)
|
|
|
|
@if [ -z "$(INSTALL_GCI)" ]; then \
|
|
|
|
echo "Installing gci v$(GCI_VERSION) to ./bin/" && GOBIN=$(INSTALL_PATH) go install github.com/daixiang0/gci@v$(GCI_VERSION); \
|
|
|
|
else \
|
|
|
|
echo "gci v$(GCI_VERSION) already installed"; \
|
|
|
|
fi
|
|
|
|
@if [ -z "$(INSTALL_GOFUMPT)" ]; then \
|
|
|
|
echo "Installing gofumpt v$(GOFUMPT_VERSION) to ./bin/" && GOBIN=$(INSTALL_PATH) go install mvdan.cc/gofumpt@v$(GOFUMPT_VERSION); \
|
|
|
|
else \
|
|
|
|
echo "gofumpt v$(GOFUMPT_VERSION) already installed"; \
|
|
|
|
fi
|
|
|
|
@echo "Running gofumpt fix"
|
2023-09-26 10:05:25 +08:00
|
|
|
@$(INSTALL_PATH)/gofumpt -l -w internal/
|
|
|
|
@$(INSTALL_PATH)/gofumpt -l -w cmd/
|
|
|
|
@$(INSTALL_PATH)/gofumpt -l -w pkg/
|
2024-05-20 20:47:38 +08:00
|
|
|
@$(INSTALL_PATH)/gofumpt -l -w client/
|
2024-06-21 10:32:02 +08:00
|
|
|
@$(INSTALL_PATH)/gofumpt -l -w tests/go_client/
|
2023-09-26 10:05:25 +08:00
|
|
|
@$(INSTALL_PATH)/gofumpt -l -w tests/integration/
|
2023-09-21 09:45:27 +08:00
|
|
|
@echo "Running gci fix"
|
|
|
|
@$(INSTALL_PATH)/gci write cmd/ --skip-generated -s standard -s default -s "prefix(github.com/milvus-io)" --custom-order
|
|
|
|
@$(INSTALL_PATH)/gci write internal/ --skip-generated -s standard -s default -s "prefix(github.com/milvus-io)" --custom-order
|
|
|
|
@$(INSTALL_PATH)/gci write pkg/ --skip-generated -s standard -s default -s "prefix(github.com/milvus-io)" --custom-order
|
2024-05-20 20:47:38 +08:00
|
|
|
@$(INSTALL_PATH)/gci write client/ --skip-generated -s standard -s default -s "prefix(github.com/milvus-io)" --custom-order
|
2023-09-21 09:45:27 +08:00
|
|
|
@$(INSTALL_PATH)/gci write tests/ --skip-generated -s standard -s default -s "prefix(github.com/milvus-io)" --custom-order
|
|
|
|
@echo "Running golangci-lint auto-fix"
|
2024-05-20 20:47:38 +08:00
|
|
|
@source $(PWD)/scripts/setenv.sh && GO111MODULE=on $(INSTALL_PATH)/golangci-lint run --fix --timeout=30m --config $(PWD)/.golangci.yml;
|
|
|
|
@source $(PWD)/scripts/setenv.sh && cd pkg && GO111MODULE=on $(INSTALL_PATH)/golangci-lint run --fix --timeout=30m --config $(PWD)/.golangci.yml
|
|
|
|
@source $(PWD)/scripts/setenv.sh && cd client && GO111MODULE=on $(INSTALL_PATH)/golangci-lint run --fix --timeout=30m --config $(PWD)/client/.golangci.yml
|
2020-12-31 17:25:45 +08:00
|
|
|
|
2020-10-30 16:25:06 +08:00
|
|
|
#TODO: Check code specifications by golangci-lint
|
2023-01-29 10:59:49 +08:00
|
|
|
static-check: getdeps
|
2020-10-30 16:25:06 +08:00
|
|
|
@echo "Running $@ check"
|
2024-06-21 10:32:02 +08:00
|
|
|
@echo "Start check core packages"
|
2024-06-09 22:55:53 +08:00
|
|
|
@source $(PWD)/scripts/setenv.sh && GO111MODULE=on $(INSTALL_PATH)/golangci-lint run --build-tags dynamic,test --timeout=30m --config $(PWD)/.golangci.yml
|
2024-06-21 10:32:02 +08:00
|
|
|
@echo "Start check pkg package"
|
2024-06-09 22:55:53 +08:00
|
|
|
@source $(PWD)/scripts/setenv.sh && cd pkg && GO111MODULE=on $(INSTALL_PATH)/golangci-lint run --build-tags dynamic,test --timeout=30m --config $(PWD)/.golangci.yml
|
2024-06-21 10:32:02 +08:00
|
|
|
@echo "Start check client package"
|
2024-05-20 20:47:38 +08:00
|
|
|
@source $(PWD)/scripts/setenv.sh && cd client && GO111MODULE=on $(INSTALL_PATH)/golangci-lint run --timeout=30m --config $(PWD)/client/.golangci.yml
|
2024-06-21 10:32:02 +08:00
|
|
|
@echo "Start check go_client e2e package"
|
|
|
|
@source $(PWD)/scripts/setenv.sh && cd tests/go_client && GO111MODULE=on $(INSTALL_PATH)/golangci-lint run --timeout=30m --config $(PWD)/client/.golangci.yml
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2022-06-10 14:58:07 +08:00
|
|
|
verifiers: build-cpp getdeps cppcheck fmt static-check
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2021-09-29 14:41:58 +08:00
|
|
|
# Build various components locally.
|
2021-04-09 14:07:03 +08:00
|
|
|
binlog:
|
|
|
|
@echo "Building binlog ..."
|
2022-07-21 19:48:27 +08:00
|
|
|
@source $(PWD)/scripts/setenv.sh && \
|
2022-10-13 20:57:23 +08:00
|
|
|
mkdir -p $(INSTALL_PATH) && go env -w CGO_ENABLED="1" && \
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
GO111MODULE=on $(GO) build -pgo=$(PGO_PATH)/default.pgo -ldflags="-r $${RPATH}" -o $(INSTALL_PATH)/binlog $(PWD)/cmd/tools/binlog/main.go 1>/dev/null
|
2021-02-20 10:15:37 +08:00
|
|
|
|
2022-10-12 11:37:23 +08:00
|
|
|
MIGRATION_PATH = $(PWD)/cmd/tools/migration
|
2022-10-17 15:07:25 +08:00
|
|
|
meta-migration:
|
2022-10-12 11:37:23 +08:00
|
|
|
@echo "Building migration tool ..."
|
2023-01-11 19:31:39 +08:00
|
|
|
@source $(PWD)/scripts/setenv.sh && \
|
|
|
|
mkdir -p $(INSTALL_PATH) && go env -w CGO_ENABLED="1" && \
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
GO111MODULE=on $(GO) build -pgo=$(PGO_PATH)/default.pgo -ldflags="-r $${RPATH} -X '$(OBJPREFIX).BuildTags=$(BUILD_TAGS)' -X '$(OBJPREFIX).BuildTime=$(BUILD_TIME)' -X '$(OBJPREFIX).GitCommit=$(GIT_COMMIT)' -X '$(OBJPREFIX).GoVersion=$(GO_VERSION)'" \
|
2023-07-13 15:34:33 +08:00
|
|
|
-tags dynamic -o $(INSTALL_PATH)/meta-migration $(MIGRATION_PATH)/main.go 1>/dev/null
|
2022-10-12 11:37:23 +08:00
|
|
|
|
2023-02-07 19:24:30 +08:00
|
|
|
INTERATION_PATH = $(PWD)/tests/integration
|
2024-03-22 10:29:06 +08:00
|
|
|
integration-test: getdeps
|
2023-02-07 19:24:30 +08:00
|
|
|
@echo "Building integration tests ..."
|
2024-03-22 10:29:06 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_intergration_test.sh "$(INSTALL_PATH)/gotestsum --")
|
2023-02-07 19:24:30 +08:00
|
|
|
|
2021-06-22 11:34:04 +08:00
|
|
|
BUILD_TAGS = $(shell git describe --tags --always --dirty="-dev")
|
2023-03-20 19:03:57 +08:00
|
|
|
BUILD_TAGS_GPU = ${BUILD_TAGS}-gpu
|
2022-02-09 14:27:46 +08:00
|
|
|
BUILD_TIME = $(shell date -u)
|
2021-06-22 11:34:04 +08:00
|
|
|
GIT_COMMIT = $(shell git rev-parse --short HEAD)
|
|
|
|
GO_VERSION = $(shell go version)
|
2023-02-10 14:24:32 +08:00
|
|
|
|
2021-12-15 10:21:10 +08:00
|
|
|
print-build-info:
|
2022-09-08 17:14:36 +08:00
|
|
|
$(shell git config --global --add safe.directory '*')
|
2021-12-15 10:21:10 +08:00
|
|
|
@echo "Build Tag: $(BUILD_TAGS)"
|
|
|
|
@echo "Build Time: $(BUILD_TIME)"
|
|
|
|
@echo "Git Commit: $(GIT_COMMIT)"
|
|
|
|
@echo "Go Version: $(GO_VERSION)"
|
|
|
|
|
2023-03-21 17:55:57 +08:00
|
|
|
print-gpu-build-info:
|
|
|
|
$(shell git config --global --add safe.directory '*')
|
|
|
|
@echo "Build Tag: $(BUILD_TAGS_GPU)"
|
|
|
|
@echo "Build Time: $(BUILD_TIME)"
|
|
|
|
@echo "Git Commit: $(GIT_COMMIT)"
|
|
|
|
@echo "Go Version: $(GO_VERSION)"
|
2021-04-15 09:28:46 +08:00
|
|
|
|
2022-10-19 20:45:32 +08:00
|
|
|
update-milvus-api: download-milvus-proto
|
|
|
|
@echo "Update milvus/api version ..."
|
2023-06-09 01:28:37 +08:00
|
|
|
@(env bash $(PWD)/scripts/update-api-version.sh $(PROTO_API_VERSION))
|
2022-10-17 19:19:25 +08:00
|
|
|
|
2022-10-16 20:49:27 +08:00
|
|
|
download-milvus-proto:
|
|
|
|
@echo "Download milvus-proto repo ..."
|
|
|
|
@(env bash $(PWD)/scripts/download_milvus_proto.sh)
|
2022-10-12 14:33:23 +08:00
|
|
|
|
2023-04-21 14:16:33 +08:00
|
|
|
build-3rdparty:
|
|
|
|
@echo "Build 3rdparty ..."
|
2024-02-20 19:16:52 +08:00
|
|
|
@(env bash $(PWD)/scripts/3rdparty_build.sh -o ${use_opendal})
|
2023-04-21 14:16:33 +08:00
|
|
|
|
2023-12-22 12:00:43 +08:00
|
|
|
generated-proto-without-cpp: download-milvus-proto
|
|
|
|
@echo "Generate proto ..."
|
|
|
|
@mkdir -p ${GOPATH}/bin
|
|
|
|
@which protoc-gen-go 1>/dev/null || (echo "Installing protoc-gen-go" && cd /tmp && go install github.com/golang/protobuf/protoc-gen-go@v1.3.2)
|
|
|
|
@(env bash $(PWD)/scripts/generate_proto.sh)
|
|
|
|
|
2023-04-21 14:16:33 +08:00
|
|
|
generated-proto: download-milvus-proto build-3rdparty
|
|
|
|
@echo "Generate proto ..."
|
|
|
|
@mkdir -p ${GOPATH}/bin
|
|
|
|
@which protoc-gen-go 1>/dev/null || (echo "Installing protoc-gen-go" && cd /tmp && go install github.com/golang/protobuf/protoc-gen-go@v1.3.2)
|
|
|
|
@(env bash $(PWD)/scripts/generate_proto.sh)
|
|
|
|
|
|
|
|
build-cpp: generated-proto
|
2023-09-26 10:31:27 +08:00
|
|
|
@echo "Building Milvus cpp library ..."
|
2024-02-20 19:16:52 +08:00
|
|
|
@(env bash $(PWD)/scripts/core_build.sh -t ${mode} -n ${use_disk_index} -y ${use_dynamic_simd} ${AZURE_OPTION} -x ${index_engine} -o ${use_opendal})
|
2022-03-16 16:51:22 +08:00
|
|
|
|
2023-04-21 14:16:33 +08:00
|
|
|
build-cpp-gpu: generated-proto
|
2023-09-20 13:59:23 +08:00
|
|
|
@echo "Building Milvus cpp gpu library ... "
|
2024-02-20 19:16:52 +08:00
|
|
|
@(env bash $(PWD)/scripts/core_build.sh -t ${mode} -g -n ${use_disk_index} -y ${use_dynamic_simd} ${AZURE_OPTION} -x ${index_engine} -o ${use_opendal})
|
2023-03-14 23:21:56 +08:00
|
|
|
|
2023-04-21 14:16:33 +08:00
|
|
|
build-cpp-with-unittest: generated-proto
|
2023-09-20 13:59:23 +08:00
|
|
|
@echo "Building Milvus cpp library with unittest ... "
|
2024-02-20 19:16:52 +08:00
|
|
|
@(env bash $(PWD)/scripts/core_build.sh -t ${mode} -u -n ${use_disk_index} -y ${use_dynamic_simd} ${AZURE_OPTION} -x ${index_engine} -o ${use_opendal})
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2023-04-21 14:16:33 +08:00
|
|
|
build-cpp-with-coverage: generated-proto
|
2022-04-11 09:49:34 +08:00
|
|
|
@echo "Building Milvus cpp library with coverage and unittest ..."
|
2024-02-20 19:16:52 +08:00
|
|
|
@(env bash $(PWD)/scripts/core_build.sh -t ${mode} -a ${use_asan} -u -c -n ${use_disk_index} -y ${use_dynamic_simd} ${AZURE_OPTION} -x ${index_engine} -o ${use_opendal})
|
2022-04-11 09:49:34 +08:00
|
|
|
|
2023-04-21 14:16:33 +08:00
|
|
|
check-proto-product: generated-proto
|
|
|
|
@(env bash $(PWD)/scripts/check_proto_product.sh)
|
|
|
|
|
2022-04-11 09:49:34 +08:00
|
|
|
|
2021-09-29 14:41:58 +08:00
|
|
|
# Run the tests.
|
2020-10-31 16:30:20 +08:00
|
|
|
unittest: test-cpp test-go
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2022-09-08 17:36:35 +08:00
|
|
|
test-util:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t util)
|
|
|
|
|
|
|
|
test-storage:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t storage)
|
|
|
|
|
|
|
|
test-allocator:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t allocator)
|
|
|
|
|
|
|
|
test-config:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t config)
|
|
|
|
|
|
|
|
test-tso:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t tso)
|
|
|
|
|
|
|
|
test-kv:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t kv)
|
|
|
|
|
|
|
|
test-mq:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t mq)
|
|
|
|
|
|
|
|
test-rootcoord:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t rootcoord)
|
|
|
|
|
2022-01-24 17:18:46 +08:00
|
|
|
test-indexnode:
|
|
|
|
@echo "Running go unittests..."
|
2022-09-08 17:36:35 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t indexnode)
|
2022-01-24 17:18:46 +08:00
|
|
|
|
2022-10-08 15:38:58 +08:00
|
|
|
test-indexcoord:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t indexcoord)
|
|
|
|
|
2021-10-24 15:16:00 +08:00
|
|
|
test-proxy:
|
|
|
|
@echo "Running go unittests..."
|
2022-09-08 17:36:35 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t proxy)
|
2021-10-24 15:16:00 +08:00
|
|
|
|
2022-03-15 21:51:21 +08:00
|
|
|
test-datacoord:
|
|
|
|
@echo "Running go unittests..."
|
2022-09-08 17:36:35 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t datacoord)
|
2022-03-15 21:51:21 +08:00
|
|
|
|
2022-01-04 23:33:24 +08:00
|
|
|
test-datanode:
|
|
|
|
@echo "Running go unittests..."
|
2022-09-08 17:36:35 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t datanode)
|
2021-10-24 15:16:00 +08:00
|
|
|
|
2022-03-08 17:39:58 +08:00
|
|
|
test-querynode:
|
|
|
|
@echo "Running go unittests..."
|
2022-09-08 17:36:35 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t querynode)
|
2022-03-08 17:39:58 +08:00
|
|
|
|
2021-10-27 23:00:24 +08:00
|
|
|
test-querycoord:
|
|
|
|
@echo "Running go unittests..."
|
2022-09-08 17:36:35 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t querycoord)
|
2021-10-27 23:00:24 +08:00
|
|
|
|
2022-09-08 17:36:35 +08:00
|
|
|
test-metastore:
|
|
|
|
@echo "Running go unittests..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh -t metastore)
|
2021-10-27 23:00:24 +08:00
|
|
|
|
2021-07-16 14:52:21 +08:00
|
|
|
test-go: build-cpp-with-unittest
|
2020-10-30 16:25:06 +08:00
|
|
|
@echo "Running go unittests..."
|
2021-12-08 10:55:04 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_unittest.sh)
|
2020-10-30 16:25:06 +08:00
|
|
|
|
|
|
|
test-cpp: build-cpp-with-unittest
|
2020-11-03 11:25:47 +08:00
|
|
|
@echo "Running cpp unittests..."
|
2021-12-08 10:55:04 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_cpp_unittest.sh)
|
2020-10-30 16:25:06 +08:00
|
|
|
|
2021-09-29 14:41:58 +08:00
|
|
|
# Run code coverage.
|
2021-09-22 16:01:53 +08:00
|
|
|
codecov: codecov-go codecov-cpp
|
2021-09-18 14:19:51 +08:00
|
|
|
|
2021-09-22 16:01:53 +08:00
|
|
|
# Run codecov-go
|
2022-04-11 09:49:34 +08:00
|
|
|
codecov-go: build-cpp-with-coverage
|
2021-09-18 14:19:51 +08:00
|
|
|
@echo "Running go coverage..."
|
2021-04-24 15:10:24 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_go_codecov.sh)
|
|
|
|
|
2024-03-27 15:29:10 +08:00
|
|
|
# Run codecov-go without build core again, used in github action
|
|
|
|
codecov-go-without-build: getdeps
|
|
|
|
@echo "Running go coverage..."
|
|
|
|
@(env bash $(PWD)/scripts/run_go_codecov.sh "$(INSTALL_PATH)/gotestsum --")
|
|
|
|
|
2021-09-22 16:01:53 +08:00
|
|
|
# Run codecov-cpp
|
2022-04-11 09:49:34 +08:00
|
|
|
codecov-cpp: build-cpp-with-coverage
|
2021-09-18 14:19:51 +08:00
|
|
|
@echo "Running cpp coverage..."
|
2021-09-17 14:21:50 +08:00
|
|
|
@(env bash $(PWD)/scripts/run_cpp_codecov.sh)
|
|
|
|
|
2021-09-29 14:41:58 +08:00
|
|
|
# Build each component and install binary to $GOPATH/bin.
|
2022-10-12 14:33:23 +08:00
|
|
|
install: milvus
|
2020-10-30 16:25:06 +08:00
|
|
|
@echo "Installing binary to './bin'"
|
2021-04-15 09:28:46 +08:00
|
|
|
@mkdir -p $(GOPATH)/bin && cp -f $(PWD)/bin/milvus $(GOPATH)/bin/milvus
|
2022-11-02 15:39:35 +08:00
|
|
|
@mkdir -p $(LIBRARY_PATH)
|
2022-11-23 10:39:11 +08:00
|
|
|
-cp -r -P $(PWD)/internal/core/output/lib/*.dylib* $(LIBRARY_PATH) 2>/dev/null
|
|
|
|
-cp -r -P $(PWD)/internal/core/output/lib/*.so* $(LIBRARY_PATH) 2>/dev/null
|
|
|
|
-cp -r -P $(PWD)/internal/core/output/lib64/*.so* $(LIBRARY_PATH) 2>/dev/null
|
2020-10-30 16:25:06 +08:00
|
|
|
@echo "Installation successful."
|
|
|
|
|
2023-03-20 19:03:57 +08:00
|
|
|
gpu-install: milvus-gpu
|
|
|
|
@echo "Installing binary to './bin'"
|
|
|
|
@mkdir -p $(GOPATH)/bin && cp -f $(PWD)/bin/milvus $(GOPATH)/bin/milvus
|
|
|
|
@mkdir -p $(LIBRARY_PATH)
|
|
|
|
-cp -r -P $(PWD)/internal/core/output/lib/*.dylib* $(LIBRARY_PATH) 2>/dev/null
|
|
|
|
-cp -r -P $(PWD)/internal/core/output/lib/*.so* $(LIBRARY_PATH) 2>/dev/null
|
|
|
|
-cp -r -P $(PWD)/internal/core/output/lib64/*.so* $(LIBRARY_PATH) 2>/dev/null
|
|
|
|
@echo "Installation successful."
|
|
|
|
|
2020-10-30 16:25:06 +08:00
|
|
|
clean:
|
|
|
|
@echo "Cleaning up all the generated files"
|
2021-01-11 18:35:54 +08:00
|
|
|
@rm -rf bin/
|
|
|
|
@rm -rf lib/
|
2021-09-08 16:32:00 +08:00
|
|
|
@rm -rf $(GOPATH)/bin/milvus
|
2022-02-10 17:03:48 +08:00
|
|
|
@rm -rf cmake_build
|
2022-05-26 22:58:03 +08:00
|
|
|
@rm -rf internal/core/output
|
2021-08-24 15:51:51 +08:00
|
|
|
|
2021-12-15 10:21:10 +08:00
|
|
|
milvus-tools: print-build-info
|
2021-08-24 15:51:51 +08:00
|
|
|
@echo "Building tools ..."
|
|
|
|
@mkdir -p $(INSTALL_PATH)/tools && go env -w CGO_ENABLED="1" && GO111MODULE=on $(GO) build \
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
-pgo=$(PGO_PATH)/default.pgo -ldflags="-X 'main.BuildTags=$(BUILD_TAGS)' -X 'main.BuildTime=$(BUILD_TIME)' -X 'main.GitCommit=$(GIT_COMMIT)' -X 'main.GoVersion=$(GO_VERSION)'" \
|
2021-08-24 15:51:51 +08:00
|
|
|
-o $(INSTALL_PATH)/tools $(PWD)/cmd/tools/* 1>/dev/null
|
2021-12-28 09:52:10 +08:00
|
|
|
|
2022-06-24 21:12:15 +08:00
|
|
|
rpm-setup:
|
2021-12-28 09:52:10 +08:00
|
|
|
@echo "Setuping rpm env ...;"
|
2021-12-30 16:59:20 +08:00
|
|
|
@build/rpm/setup-env.sh
|
2021-12-28 09:52:10 +08:00
|
|
|
|
|
|
|
rpm: install
|
|
|
|
@echo "Note: run 'make rpm-setup' to setup build env for rpm builder"
|
|
|
|
@echo "Building rpm ...;"
|
|
|
|
@yum -y install rpm-build rpmdevtools wget
|
|
|
|
@rm -rf ~/rpmbuild/BUILD/*
|
|
|
|
@rpmdev-setuptree
|
|
|
|
@wget https://github.com/etcd-io/etcd/releases/download/v3.5.0/etcd-v3.5.0-linux-amd64.tar.gz && tar -xf etcd-v3.5.0-linux-amd64.tar.gz
|
|
|
|
@cp etcd-v3.5.0-linux-amd64/etcd bin/etcd
|
|
|
|
@wget https://dl.min.io/server/minio/release/linux-amd64/archive/minio.RELEASE.2021-02-14T04-01-33Z -O bin/minio
|
|
|
|
@cp -r bin ~/rpmbuild/BUILD/
|
|
|
|
@cp -r lib ~/rpmbuild/BUILD/
|
|
|
|
@cp -r configs ~/rpmbuild/BUILD/
|
|
|
|
@cp -r build/rpm/services ~/rpmbuild/BUILD/
|
2021-12-30 16:59:20 +08:00
|
|
|
@QA_RPATHS="$$[ 0x001|0x0002|0x0020 ]" rpmbuild -ba ./build/rpm/milvus.spec
|
2022-09-23 10:50:51 +08:00
|
|
|
|
2023-09-04 21:19:48 +08:00
|
|
|
generate-mockery-types: getdeps
|
|
|
|
# RootCoord
|
|
|
|
$(INSTALL_PATH)/mockery --name=RootCoordComponent --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_rootcoord.go --with-expecter --structname=RootCoord
|
|
|
|
# Proxy
|
|
|
|
$(INSTALL_PATH)/mockery --name=ProxyComponent --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_proxy.go --with-expecter --structname=MockProxy
|
|
|
|
# QueryCoord
|
2023-09-11 20:43:17 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=QueryCoordComponent --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_querycoord.go --with-expecter --structname=MockQueryCoord
|
2023-09-04 21:19:48 +08:00
|
|
|
# QueryNode
|
2023-09-11 20:43:17 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=QueryNodeComponent --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_querynode.go --with-expecter --structname=MockQueryNode
|
2023-09-04 21:19:48 +08:00
|
|
|
# DataCoord
|
|
|
|
$(INSTALL_PATH)/mockery --name=DataCoordComponent --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_datacoord.go --with-expecter --structname=MockDataCoord
|
|
|
|
# DataNode
|
|
|
|
$(INSTALL_PATH)/mockery --name=DataNodeComponent --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_datanode.go --with-expecter --structname=MockDataNode
|
|
|
|
# IndexNode
|
|
|
|
$(INSTALL_PATH)/mockery --name=IndexNodeComponent --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_indexnode.go --with-expecter --structname=MockIndexNode
|
|
|
|
|
2023-09-26 09:57:25 +08:00
|
|
|
# Clients
|
|
|
|
$(INSTALL_PATH)/mockery --name=RootCoordClient --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_rootcoord_client.go --with-expecter --structname=MockRootCoordClient
|
|
|
|
$(INSTALL_PATH)/mockery --name=QueryCoordClient --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_querycoord_client.go --with-expecter --structname=MockQueryCoordClient
|
|
|
|
$(INSTALL_PATH)/mockery --name=DataCoordClient --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_datacoord_client.go --with-expecter --structname=MockDataCoordClient
|
|
|
|
$(INSTALL_PATH)/mockery --name=QueryNodeClient --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_querynode_client.go --with-expecter --structname=MockQueryNodeClient
|
|
|
|
$(INSTALL_PATH)/mockery --name=DataNodeClient --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_datanode_client.go --with-expecter --structname=MockDataNodeClient
|
|
|
|
$(INSTALL_PATH)/mockery --name=IndexNodeClient --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_indexnode_client.go --with-expecter --structname=MockIndexNodeClient
|
|
|
|
$(INSTALL_PATH)/mockery --name=ProxyClient --dir=$(PWD)/internal/types --output=$(PWD)/internal/mocks --filename=mock_proxy_client.go --with-expecter --structname=MockProxyClient
|
|
|
|
|
2023-09-04 21:19:48 +08:00
|
|
|
generate-mockery-rootcoord: getdeps
|
|
|
|
$(INSTALL_PATH)/mockery --name=IMetaTable --dir=$(PWD)/internal/rootcoord --output=$(PWD)/internal/rootcoord/mocks --filename=meta_table.go --with-expecter --outpkg=mockrootcoord
|
|
|
|
$(INSTALL_PATH)/mockery --name=GarbageCollector --dir=$(PWD)/internal/rootcoord --output=$(PWD)/internal/rootcoord/mocks --filename=garbage_collector.go --with-expecter --outpkg=mockrootcoord
|
|
|
|
|
|
|
|
generate-mockery-proxy: getdeps
|
|
|
|
$(INSTALL_PATH)/mockery --name=Cache --dir=$(PWD)/internal/proxy --output=$(PWD)/internal/proxy --filename=mock_cache.go --structname=MockCache --with-expecter --outpkg=proxy --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=timestampAllocatorInterface --dir=$(PWD)/internal/proxy --output=$(PWD)/internal/proxy --filename=mock_tso_test.go --structname=mockTimestampAllocator --with-expecter --outpkg=proxy --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=LBPolicy --dir=$(PWD)/internal/proxy --output=$(PWD)/internal/proxy --filename=mock_lb_policy.go --structname=MockLBPolicy --with-expecter --outpkg=proxy --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=LBBalancer --dir=$(PWD)/internal/proxy --output=$(PWD)/internal/proxy --filename=mock_lb_balancer.go --structname=MockLBBalancer --with-expecter --outpkg=proxy --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=shardClientMgr --dir=$(PWD)/internal/proxy --output=$(PWD)/internal/proxy --filename=mock_shardclient_manager.go --structname=MockShardClientManager --with-expecter --outpkg=proxy --inpackage
|
2023-09-12 10:19:17 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=channelsMgr --dir=$(PWD)/internal/proxy --output=$(PWD)/internal/proxy --filename=mock_channels_manager.go --structname=MockChannelsMgr --with-expecter --outpkg=proxy --inpackage
|
2023-09-04 21:19:48 +08:00
|
|
|
|
|
|
|
generate-mockery-querycoord: getdeps
|
|
|
|
$(INSTALL_PATH)/mockery --name=QueryNodeServer --dir=$(PWD)/internal/proto/querypb/ --output=$(PWD)/internal/querycoordv2/mocks --filename=mock_querynode.go --with-expecter --structname=MockQueryNodeServer
|
|
|
|
$(INSTALL_PATH)/mockery --name=Broker --dir=$(PWD)/internal/querycoordv2/meta --output=$(PWD)/internal/querycoordv2/meta --filename=mock_broker.go --with-expecter --structname=MockBroker --outpkg=meta
|
2024-05-17 10:11:37 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=TargetManagerInterface --dir=$(PWD)/internal/querycoordv2/meta --output=$(PWD)/internal/querycoordv2/meta --filename=mock_target_manager.go --with-expecter --structname=MockTargetManager --inpackage
|
2023-09-04 21:19:48 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Scheduler --dir=$(PWD)/internal/querycoordv2/task --output=$(PWD)/internal/querycoordv2/task --filename=mock_scheduler.go --with-expecter --structname=MockScheduler --outpkg=task --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=Cluster --dir=$(PWD)/internal/querycoordv2/session --output=$(PWD)/internal/querycoordv2/session --filename=mock_cluster.go --with-expecter --structname=MockCluster --outpkg=session --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=Balance --dir=$(PWD)/internal/querycoordv2/balance --output=$(PWD)/internal/querycoordv2/balance --filename=mock_balancer.go --with-expecter --structname=MockBalancer --outpkg=balance --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=Controller --dir=$(PWD)/internal/querycoordv2/dist --output=$(PWD)/internal/querycoordv2/dist --filename=mock_controller.go --with-expecter --structname=MockController --outpkg=dist --inpackage
|
|
|
|
|
2023-09-05 14:13:49 +08:00
|
|
|
generate-mockery-querynode: getdeps build-cpp
|
|
|
|
@source $(PWD)/scripts/setenv.sh # setup PKG_CONFIG_PATH
|
2023-10-11 20:45:35 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=QueryHook --dir=$(PWD)/internal/querynodev2/optimizers --output=$(PWD)/internal/querynodev2/optimizers --filename=mock_query_hook.go --with-expecter --outpkg=optimizers --structname=MockQueryHook --inpackage
|
2023-09-04 21:19:48 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Manager --dir=$(PWD)/internal/querynodev2/cluster --output=$(PWD)/internal/querynodev2/cluster --filename=mock_manager.go --with-expecter --outpkg=cluster --structname=MockManager --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=SegmentManager --dir=$(PWD)/internal/querynodev2/segments --output=$(PWD)/internal/querynodev2/segments --filename=mock_segment_manager.go --with-expecter --outpkg=segments --structname=MockSegmentManager --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=CollectionManager --dir=$(PWD)/internal/querynodev2/segments --output=$(PWD)/internal/querynodev2/segments --filename=mock_collection_manager.go --with-expecter --outpkg=segments --structname=MockCollectionManager --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=Loader --dir=$(PWD)/internal/querynodev2/segments --output=$(PWD)/internal/querynodev2/segments --filename=mock_loader.go --with-expecter --outpkg=segments --structname=MockLoader --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=Segment --dir=$(PWD)/internal/querynodev2/segments --output=$(PWD)/internal/querynodev2/segments --filename=mock_segment.go --with-expecter --outpkg=segments --structname=MockSegment --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=Worker --dir=$(PWD)/internal/querynodev2/cluster --output=$(PWD)/internal/querynodev2/cluster --filename=mock_worker.go --with-expecter --outpkg=worker --structname=MockWorker --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=ShardDelegator --dir=$(PWD)/internal/querynodev2/delegator/ --output=$(PWD)/internal/querynodev2/delegator/ --filename=mock_delegator.go --with-expecter --outpkg=delegator --structname=MockShardDelegator --inpackage
|
|
|
|
|
|
|
|
generate-mockery-datacoord: getdeps
|
|
|
|
$(INSTALL_PATH)/mockery --name=compactionPlanContext --dir=internal/datacoord --filename=mock_compaction_plan_context.go --output=internal/datacoord --structname=MockCompactionPlanContext --with-expecter --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=Handler --dir=internal/datacoord --filename=mock_handler.go --output=internal/datacoord --structname=NMockHandler --with-expecter --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=allocator --dir=internal/datacoord --filename=mock_allocator_test.go --output=internal/datacoord --structname=NMockAllocator --with-expecter --inpackage
|
2023-11-07 14:02:20 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=RWChannelStore --dir=internal/datacoord --filename=mock_channel_store.go --output=internal/datacoord --structname=MockRWChannelStore --with-expecter --inpackage
|
2023-11-21 09:26:22 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=IndexEngineVersionManager --dir=internal/datacoord --filename=mock_index_engine_version_manager.go --output=internal/datacoord --structname=MockVersionManager --with-expecter --inpackage
|
2023-11-23 17:30:25 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=TriggerManager --dir=internal/datacoord --filename=mock_trigger_manager.go --output=internal/datacoord --structname=MockTriggerManager --with-expecter --inpackage
|
2023-12-11 17:52:37 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Cluster --dir=internal/datacoord --filename=mock_cluster.go --output=internal/datacoord --structname=MockCluster --with-expecter --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=SessionManager --dir=internal/datacoord --filename=mock_session_manager.go --output=internal/datacoord --structname=MockSessionManager --with-expecter --inpackage
|
2023-11-23 17:30:25 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=compactionPlanContext --dir=internal/datacoord --filename=mock_compaction_plan_context.go --output=internal/datacoord --structname=MockCompactionPlanContext --with-expecter --inpackage
|
2023-12-05 18:44:37 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=CompactionMeta --dir=internal/datacoord --filename=mock_compaction_meta.go --output=internal/datacoord --structname=MockCompactionMeta --with-expecter --inpackage
|
2023-12-25 19:24:46 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=ChannelManager --dir=internal/datacoord --filename=mock_channelmanager.go --output=internal/datacoord --structname=MockChannelManager --with-expecter --inpackage
|
2024-05-07 15:49:30 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=SubCluster --dir=internal/datacoord --filename=mock_subcluster.go --output=internal/datacoord --structname=MockSubCluster --with-expecter --inpackage
|
2024-04-12 16:01:19 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Broker --dir=internal/datacoord/broker --filename=mock_coordinator_broker.go --output=internal/datacoord/broker --structname=MockBroker --with-expecter --inpackage
|
2024-06-06 17:37:51 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=WorkerManager --dir=internal/datacoord --filename=mock_worker_manager.go --output=internal/datacoord --structname=MockWorkerManager --with-expecter --inpackage
|
2023-09-04 21:19:48 +08:00
|
|
|
|
|
|
|
generate-mockery-datanode: getdeps
|
|
|
|
$(INSTALL_PATH)/mockery --name=Allocator --dir=$(PWD)/internal/datanode/allocator --output=$(PWD)/internal/datanode/allocator --filename=mock_allocator.go --with-expecter --structname=MockAllocator --outpkg=allocator --inpackage
|
2023-10-11 17:03:34 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Broker --dir=$(PWD)/internal/datanode/broker --output=$(PWD)/internal/datanode/broker/ --filename=mock_broker.go --with-expecter --structname=MockBroker --outpkg=broker --inpackage
|
2023-10-31 02:30:16 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=MetaCache --dir=$(PWD)/internal/datanode/metacache --output=$(PWD)/internal/datanode/metacache --filename=mock_meta_cache.go --with-expecter --structname=MockMetaCache --outpkg=metacache --inpackage
|
2023-11-04 12:10:17 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=SyncManager --dir=$(PWD)/internal/datanode/syncmgr --output=$(PWD)/internal/datanode/syncmgr --filename=mock_sync_manager.go --with-expecter --structname=MockSyncManager --outpkg=syncmgr --inpackage
|
2023-12-26 10:40:47 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=MetaWriter --dir=$(PWD)/internal/datanode/syncmgr --output=$(PWD)/internal/datanode/syncmgr --filename=mock_meta_writer.go --with-expecter --structname=MockMetaWriter --outpkg=syncmgr --inpackage
|
2024-01-05 21:42:47 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Serializer --dir=$(PWD)/internal/datanode/syncmgr --output=$(PWD)/internal/datanode/syncmgr --filename=mock_serializer.go --with-expecter --structname=MockSerializer --outpkg=syncmgr --inpackage
|
2024-02-05 11:33:43 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Task --dir=$(PWD)/internal/datanode/syncmgr --output=$(PWD)/internal/datanode/syncmgr --filename=mock_task.go --with-expecter --structname=MockTask --outpkg=syncmgr --inpackage
|
2023-11-15 15:24:18 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=WriteBuffer --dir=$(PWD)/internal/datanode/writebuffer --output=$(PWD)/internal/datanode/writebuffer --filename=mock_write_buffer.go --with-expecter --structname=MockWriteBuffer --outpkg=writebuffer --inpackage
|
2023-11-16 00:22:20 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=BufferManager --dir=$(PWD)/internal/datanode/writebuffer --output=$(PWD)/internal/datanode/writebuffer --filename=mock_mananger.go --with-expecter --structname=MockBufferManager --outpkg=writebuffer --inpackage
|
2023-11-30 14:30:28 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=BinlogIO --dir=$(PWD)/internal/datanode/io --output=$(PWD)/internal/datanode/io --filename=mock_binlogio.go --with-expecter --structname=MockBinlogIO --outpkg=io --inpackage
|
2023-11-30 18:42:32 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=FlowgraphManager --dir=$(PWD)/internal/datanode --output=$(PWD)/internal/datanode --filename=mock_fgmanager.go --with-expecter --structname=MockFlowgraphManager --outpkg=datanode --inpackage
|
2023-12-27 16:00:48 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=ChannelManager --dir=$(PWD)/internal/datanode --output=$(PWD)/internal/datanode --filename=mock_channelmanager.go --with-expecter --structname=MockChannelManager --outpkg=datanode --inpackage
|
2024-05-23 09:53:40 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Compactor --dir=$(PWD)/internal/datanode/compaction --output=$(PWD)/internal/datanode/compaction --filename=mock_compactor.go --with-expecter --structname=MockCompactor --outpkg=compaction --inpackage
|
2023-09-04 21:19:48 +08:00
|
|
|
|
|
|
|
generate-mockery-metastore: getdeps
|
2023-09-11 20:43:17 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=RootCoordCatalog --dir=$(PWD)/internal/metastore --output=$(PWD)/internal/metastore/mocks --filename=mock_rootcoord_catalog.go --with-expecter --structname=RootCoordCatalog --outpkg=mocks
|
|
|
|
$(INSTALL_PATH)/mockery --name=DataCoordCatalog --dir=$(PWD)/internal/metastore --output=$(PWD)/internal/metastore/mocks --filename=mock_datacoord_catalog.go --with-expecter --structname=DataCoordCatalog --outpkg=mocks
|
|
|
|
$(INSTALL_PATH)/mockery --name=QueryCoordCatalog --dir=$(PWD)/internal/metastore --output=$(PWD)/internal/metastore/mocks --filename=mock_querycoord_catalog.go --with-expecter --structname=QueryCoordCatalog --outpkg=mocks
|
2023-09-04 21:19:48 +08:00
|
|
|
|
|
|
|
generate-mockery-utils: getdeps
|
|
|
|
# dependency.Factory
|
|
|
|
$(INSTALL_PATH)/mockery --name=Factory --dir=internal/util/dependency --output=internal/util/dependency --filename=mock_factory.go --with-expecter --structname=MockFactory --inpackage
|
|
|
|
# tso.Allocator
|
|
|
|
$(INSTALL_PATH)/mockery --name=Allocator --dir=internal/tso --output=internal/tso/mocks --filename=allocator.go --with-expecter --structname=Allocator --outpkg=mocktso
|
2023-11-03 10:14:16 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=SessionInterface --dir=$(PWD)/internal/util/sessionutil --output=$(PWD)/internal/util/sessionutil --filename=mock_session.go --with-expecter --structname=MockSession --inpackage
|
2023-12-13 10:46:44 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=GrpcClient --dir=$(PWD)/internal/util/grpcclient --output=$(PWD)/internal/mocks --filename=mock_grpc_client.go --with-expecter --structname=MockGrpcClient
|
2023-12-20 19:22:42 +08:00
|
|
|
# proxy_client_manager.go
|
|
|
|
$(INSTALL_PATH)/mockery --name=ProxyClientManagerInterface --dir=$(PWD)/internal/util/proxyutil --output=$(PWD)/internal/util/proxyutil --filename=mock_proxy_client_manager.go --with-expecter --structname=MockProxyClientManager --inpackage
|
|
|
|
$(INSTALL_PATH)/mockery --name=ProxyWatcherInterface --dir=$(PWD)/internal/util/proxyutil --output=$(PWD)/internal/util/proxyutil --filename=mock_proxy_watcher.go --with-expecter --structname=MockProxyWatcher --inpackage
|
|
|
|
|
2023-09-04 21:19:48 +08:00
|
|
|
generate-mockery-kv: getdeps
|
|
|
|
$(INSTALL_PATH)/mockery --name=TxnKV --dir=$(PWD)/internal/kv --output=$(PWD)/internal/kv/mocks --filename=txn_kv.go --with-expecter
|
|
|
|
$(INSTALL_PATH)/mockery --name=MetaKv --dir=$(PWD)/internal/kv --output=$(PWD)/internal/kv/mocks --filename=meta_kv.go --with-expecter
|
|
|
|
$(INSTALL_PATH)/mockery --name=WatchKV --dir=$(PWD)/internal/kv --output=$(PWD)/internal/kv/mocks --filename=watch_kv.go --with-expecter
|
|
|
|
$(INSTALL_PATH)/mockery --name=SnapShotKV --dir=$(PWD)/internal/kv --output=$(PWD)/internal/kv/mocks --filename=snapshot_kv.go --with-expecter
|
2023-09-27 10:21:26 +08:00
|
|
|
$(INSTALL_PATH)/mockery --name=Predicate --dir=$(PWD)/internal/kv/predicates --output=$(PWD)/internal/kv/predicates --filename=mock_predicate.go --with-expecter --inpackage
|
2023-09-04 21:19:48 +08:00
|
|
|
|
2023-11-29 20:52:27 +08:00
|
|
|
generate-mockery-chunk-manager: getdeps
|
|
|
|
$(INSTALL_PATH)/mockery --name=ChunkManager --dir=$(PWD)/internal/storage --output=$(PWD)/internal/mocks --filename=mock_chunk_manager.go --with-expecter
|
|
|
|
|
2023-09-04 21:19:48 +08:00
|
|
|
generate-mockery-pkg:
|
|
|
|
$(MAKE) -C pkg generate-mockery
|
|
|
|
|
2024-06-11 10:38:01 +08:00
|
|
|
generate-mockery-log:
|
|
|
|
$(INSTALL_PATH)/mockery --config $(PWD)/internal/logservice/.mockery.yaml
|
|
|
|
|
|
|
|
generate-mockery: generate-mockery-types generate-mockery-kv generate-mockery-rootcoord generate-mockery-proxy generate-mockery-querycoord generate-mockery-querynode generate-mockery-datacoord generate-mockery-pkg generate-mockery-log
|
2023-09-11 20:43:17 +08:00
|
|
|
|
2024-05-10 14:31:37 +08:00
|
|
|
generate-yaml: milvus-tools
|
|
|
|
@echo "Updating milvus config yaml"
|
|
|
|
@$(PWD)/bin/tools/config gen-yaml && mv milvus.yaml configs/milvus.yaml
|
2024-03-25 15:51:09 +08:00
|
|
|
|
|
|
|
MMAP_MIGRATION_PATH = $(PWD)/cmd/tools/migration/mmap/tool
|
|
|
|
mmap-migration:
|
|
|
|
@echo "Building migration tool ..."
|
|
|
|
@source $(PWD)/scripts/setenv.sh && \
|
|
|
|
mkdir -p $(INSTALL_PATH) && go env -w CGO_ENABLED="1" && \
|
Upgrade go from 1.20 to 1.21 (#33047)
Signed-off-by: shaoting-huang [shaoting-huang@zilliz.com]
issue: https://github.com/milvus-io/milvus/issues/32982
# Background
Go 1.21 introduces several improvements and changes over Go 1.20, which
is quite stable now. According to
[Go 1.21 Release Notes](https://tip.golang.org/doc/go1.21), the big
difference of Go 1.21 is enabling Profile-Guided Optimization by
default, which can improve performance by around 2-14%. Here are the
summary steps of PGO:
1. Build Initial Binary (Without PGO)
2. Deploying the Production Environment
3. Run the program and collect Performance Analysis Data (CPU pprof)
4. Analyze the Collected Data and Select a Performance Profile for PGO
5. Place the Performance Analysis File in the Main Package Directory and
Name It default.pgo
6. go build Detects the default.pgo File and Enables PGO
7. Build and Release the Updated Binary (With PGO)
8. Iterate and Repeat the Above Steps
<img width="657" alt="Screenshot 2024-05-14 at 15 57 01"
src="https://github.com/milvus-io/milvus/assets/167743503/b08d4300-0be1-44dc-801f-ce681dabc581">
# What does this PR do
There are three experiments, search benchmark by Zilliz test platform,
search benchmark by open-source
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file),
and search benchmark with PGO. We do both search benchmarks by Zilliz
test platform and by VectorDBBench to reduce reliance on a single
experimental result. Besides, we validate the performance enhancement
with PGO.
## Search Benchmark Report by Zilliz Test Platform
An upgrade to Go 1.21 was conducted on a Milvus Standalone server,
equipped with 16 CPUs and 64GB of memory. The search performance was
evaluated using a 1 million entry local dataset with an L2 metric type
in a 768-dimensional space. The system was tested for concurrent
searches with 50 concurrent tasks for 1 hour, each with a 20-second
interval. The reason for using one server rather than two servers to
compare is to guarantee the same data source and same segment state
after compaction.
Test Sequence:
1. Go 1.20 Initial Run: Insert data, build index, load index, and
search.
2. Go 1.20 Rebuild: Rebuild the index with the same dataset, load index,
and search.
3. Go 1.21 Load: Upload to Go 1.21 within the server. Then load the
index from the second run, and search.
4. Go 1.21 Rebuild: Rebuild the index with the same dataset, load index,
and search.
Search Metrics:
| Metric | Go 1.20 | Go 1.20 Rebuild Index | Go 1.21 | Go 1.21 Rebuild
Index |
|----------------------------|------------------|-----------------|------------------|-----------------|
| `search requests` | 10,942,683 | 16,131,726 | 16,200,887 | 16,331,052
|
| `search fails` | 0 | 0 | 0 | 0 |
| `search RT_avg` (ms) | 16.44 | 11.15 | 11.11 | 11.02 |
| `search RT_min` (ms) | 1.30 | 1.28 | 1.31 | 1.26 |
| `search RT_max` (ms) | 446.61 | 233.22 | 235.90 | 147.93 |
| `search TP50` (ms) | 11.74 | 10.46 | 10.43 | 10.35 |
| `search TP99` (ms) | 92.30 | 25.76 | 25.36 | 25.23 |
| `search RPS` | 3,039 | 4,481 | 4,500 | 4,536 |
### Key Findings
The benchmark tests reveal that the index build time with Go 1.20 at
340.39 ms and Go 1.21 at 337.60 ms demonstrated negligible performance
variance in index construction. However, Go 1.21 offers slightly better
performance in search operations compared to Go 1.20, with improvements
in handling concurrent tasks and reducing response times.
## Search Benchmark Report By VectorDb Bench
Follow
[VectorDBBench](https://github.com/zilliztech/VectorDBBench?tab=readme-ov-file)
to create a VectorDb Bench test for Go 1.20 and Go 1.21. We test the
search performance with Go 1.20 and Go 1.21 (without PGO) on the Milvus
Standalone system. The tests were conducted using the Cohere dataset
with 1 million entries in a 768-dimensional space, utilizing the COSINE
metric type.
Search Metrics:
Metric | Go 1.20 | Go 1.21 without PGO
-- | -- | --
Load Duration (seconds) | 1195.95 | 976.37
Queries Per Second (QPS) | 841.62 | 875.89
99th Percentile Serial Latency (seconds) | 0.0047 | 0.0076
Recall | 0.9487 | 0.9489
### Key Findings
Go 1.21 indicates faster index loading times and larger search QPS
handling.
## PGO Performance Test
Milvus has already added
[net/http/pprof](https://pkg.go.dev/net/http/pprof) in the metrics. So
we can curl the CPU profile directly by running
`curl -o default.pgo
"http://${MILVUS_SERVER_IP}:${MILVUS_SERVER_PORT}/debug/pprof/profile?seconds=${TIME_SECOND}"`
to collect the profile as the default.pgo during the first search. Then
I build Milvus with PGO and use the same index to run the search again.
The result is as below:
Search Metrics
| Metric | Go 1.21 Without PGO | Go 1.21 With PGO | Change (%) |
|---------------------------------------------|------------------|-----------------|------------|
| `search Requests` | 2,644,583 | 2,837,726 | +7.30% |
| `search Fails` | 0 | 0 | N/A |
| `search RT_avg` (ms) | 11.34 | 10.57 | -6.78% |
| `search RT_min` (ms) | 1.39 | 1.32 | -5.18% |
| `search RT_max` (ms) | 349.72 | 143.72 | -58.91% |
| `search TP50` (ms) | 10.57 | 9.93 | -6.05% |
| `search TP99` (ms) | 26.14 | 24.16 | -7.56% |
| `search RPS` | 4,407 | 4,729 | +7.30% |
### Key Findings
PGO led to a notable enhancement in search performance, particularly in
reducing the maximum response time by 58% and increasing the search QPS
by 7.3%.
### Further Analysis
Generate a diff flame graphs between two CPU profiles by running `go
tool pprof -http=:8000 -diff_base nopgo.pgo pgo.pgo -normalize`
<img width="1894" alt="goprofiling"
src="https://github.com/milvus-io/milvus/assets/167743503/ab9e91eb-95c7-4963-acd9-d1c3c73ee010">
Further insight of HnswIndexNode and Milvus Search Handler
<img width="1906" alt="hnsw"
src="https://github.com/milvus-io/milvus/assets/167743503/a04cf4a0-7c97-4451-b3cf-98afc20a0b05">
<img width="1873" alt="search_handler"
src="https://github.com/milvus-io/milvus/assets/167743503/5f4d3982-18dd-4115-8e76-460f7f534c7f">
After applying PGO to the Milvus server, the CPU utilization of the
faiss::fvec_L2 function has decreased. This optimization significantly
enhances the performance of the
[HnswIndexNode::Search::searchKnn](https://github.com/zilliztech/knowhere/blob/e0c9c41aa22d8f6e6761a0a54460e4573de15bfe/src/index/hnsw/hnsw.cc#L203)
method, which is frequently invoked by Knowhere during high-concurrency
searches. As the explanation from Go release notes, the function might
be more aggressively inlined by Go compiler during the second build with
the CPU profiling collected from the first run. As a result, the search
handler efficiency within Milvus DataNode has improved, allowing the
server to process a higher number of search queries per second (QPS).
# Conclusion
The combination of Go 1.21 and PGO has led to substantial enhancements
in search performance for Milvus server, particularly in terms of search
QPS and response times, making it more efficient for handling
high-concurrency search operations.
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
2024-05-22 13:21:39 +08:00
|
|
|
GO111MODULE=on $(GO) build -pgo=$(PGO_PATH)/default.pgo -ldflags="-r $${RPATH} -X '$(OBJPREFIX).BuildTags=$(BUILD_TAGS)' -X '$(OBJPREFIX).BuildTime=$(BUILD_TIME)' -X '$(OBJPREFIX).GitCommit=$(GIT_COMMIT)' -X '$(OBJPREFIX).GoVersion=$(GO_VERSION)'" \
|
2024-05-23 09:53:40 +08:00
|
|
|
-tags dynamic -o $(INSTALL_PATH)/mmap-migration $(MMAP_MIGRATION_PATH)/main.go 1>/dev/null
|