diff --git a/internal/core/src/storage/ThreadPool.h b/internal/core/src/storage/ThreadPool.h
index 9c1659648b..521ddd9c89 100644
--- a/internal/core/src/storage/ThreadPool.h
+++ b/internal/core/src/storage/ThreadPool.h
@@ -41,10 +41,13 @@ class ThreadPool {
         max_threads_size_ = CPU_NUM * thread_core_coefficient;
 
         // only IO pool will set large limit, but the CPU helps nothing to IO operations,
-        // we need to limit the max thread num, each thread will download 16 MiB data,
-        // it should be not greater than 256 (4GiB data) to avoid OOM and send too many requests to object storage
-        if (max_threads_size_ > 256) {
-            max_threads_size_ = 256;
+        // we need to limit the max thread num, each thread will download 16~64 MiB data,
+        // according to our benchmark, 16 threads is enough to saturate the network bandwidth.
+        if (min_threads_size_ > 16) {
+            min_threads_size_ = 16;
+        }
+        if (max_threads_size_ > 16) {
+            max_threads_size_ = 16;
         }
         LOG_INFO("Init thread pool:{}", name_)
             << " with min worker num:" << min_threads_size_
diff --git a/internal/querynodev2/segments/pool.go b/internal/querynodev2/segments/pool.go
index 4119cd2e99..bbf21b91a7 100644
--- a/internal/querynodev2/segments/pool.go
+++ b/internal/querynodev2/segments/pool.go
@@ -79,8 +79,12 @@ func initDynamicPool() {
 func initLoadPool() {
 	loadOnce.Do(func() {
 		pt := paramtable.Get()
+		poolSize := hardware.GetCPUNum() * pt.CommonCfg.MiddlePriorityThreadCoreCoefficient.GetAsInt()
+		if poolSize > 16 {
+			poolSize = 16
+		}
 		pool := conc.NewPool[any](
-			hardware.GetCPUNum()*pt.CommonCfg.MiddlePriorityThreadCoreCoefficient.GetAsInt(),
+			poolSize,
 			conc.WithPreAlloc(false),
 			conc.WithDisablePurge(false),
 			conc.WithPreHandler(runtime.LockOSThread), // lock os thread for cgo thread disposal