[PR #134] [CLOSED] [TransferEngine] Improve local-to-local memcpy performance #1347

New issue

Closed

opened 2026-10-29 21:26:37 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-29 21:26:37 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/kvcache-ai/Mooncake/pull/134
Author: @alogfans
Created: 3/11/2025
Status: ❌ Closed

Base: main ← Head: mooncake-store-perf-tune

📝 Commits (10+)

943c3b6 Squashed commit of the following:
1c12cf6 add docs for guide
7760ff2 Squashed commit of the following:
ab45c7b Merge branch 'kvcache-ai:main' into add-mooncake-store
3a0b3bd doc: add mooncake store python API
88583b0 change ip addresses for test
96fbce5 add seperate local memcpy transport
ba1fb3b Merge branch 'main' into mooncake-store-perf-tune
7a3339e Merge branch 'main' into mooncake-store-perf-tune
37bac27 [FIX] make shared memory use the Status-based APIs

📊 Changes

13 files changed (+342 additions, -27 deletions)

View changed files

📝 README.md (+3 -0)
📝 mooncake-p2p-store/build.sh (+2 -1)
📝 mooncake-store/src/client.cpp (+4 -0)
📝 mooncake-store/tests/distributed_object_store_provider.py (+6 -6)
📝 mooncake-store/tests/stress_cluster_benchmark.py (+16 -7)
📝 mooncake-store/tests/stress_workload_test.cpp (+9 -7)
📝 mooncake-transfer-engine/include/error.h (+1 -0)
➕ mooncake-transfer-engine/include/transport/shm_transport/shm_transport.h (+112 -0)
📝 mooncake-transfer-engine/src/CMakeLists.txt (+4 -4)
📝 mooncake-transfer-engine/src/multi_transport.cpp (+5 -2)
📝 mooncake-transfer-engine/src/transport/CMakeLists.txt (+3 -0)
➕ mooncake-transfer-engine/src/transport/shm_transport/CMakeLists.txt (+3 -0)
➕ mooncake-transfer-engine/src/transport/shm_transport/shm_transport.cpp (+174 -0)

📄 Description

In the previous version, local-to-local memcpy is implemented in RDMA transport, and this is done in sync. Therefore this leads bad performance of Mooncake Store put/get. I implemented a ShmTransport to handle this case. Memory copy is done in async now. In the future it can be extended to support shared memory.

Also modified some bench code of Mooncake Store.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/kvcache-ai/Mooncake/pull/134 **Author:** [@alogfans](https://github.com/alogfans) **Created:** 3/11/2025 **Status:** ❌ Closed **Base:** `main` ← **Head:** `mooncake-store-perf-tune` --- ### 📝 Commits (10+) - [`943c3b6`](https://github.com/kvcache-ai/Mooncake/commit/943c3b6bedb3ef0309560cc16dcf31506b1b479c) Squashed commit of the following: - [`1c12cf6`](https://github.com/kvcache-ai/Mooncake/commit/1c12cf6c0694bcc9adb06de0952e8b47b311f069) add docs for guide - [`7760ff2`](https://github.com/kvcache-ai/Mooncake/commit/7760ff26ae2fa6bf250a4d3d5ae82e61874c8efa) Squashed commit of the following: - [`ab45c7b`](https://github.com/kvcache-ai/Mooncake/commit/ab45c7bb5d7771e7d175e2579ac6a1d41c1325bb) Merge branch 'kvcache-ai:main' into add-mooncake-store - [`3a0b3bd`](https://github.com/kvcache-ai/Mooncake/commit/3a0b3bd49c66e102e9775813e1ad165de6f46f58) doc: add mooncake store python API - [`88583b0`](https://github.com/kvcache-ai/Mooncake/commit/88583b0e8b39985d83919a9d78236b325803fb1d) change ip addresses for test - [`96fbce5`](https://github.com/kvcache-ai/Mooncake/commit/96fbce5950cf143ecb1ba0d6617ce8d3e02774b9) add seperate local memcpy transport - [`ba1fb3b`](https://github.com/kvcache-ai/Mooncake/commit/ba1fb3b9827141a64c76e547c0bb5a17d2e15948) Merge branch 'main' into mooncake-store-perf-tune - [`7a3339e`](https://github.com/kvcache-ai/Mooncake/commit/7a3339e54de865de79da03b88c20f04977ae332a) Merge branch 'main' into mooncake-store-perf-tune - [`37bac27`](https://github.com/kvcache-ai/Mooncake/commit/37bac274cd28dc1bf412b903276784b50233c8d2) [FIX] make shared memory use the Status-based APIs ### 📊 Changes **13 files changed** (+342 additions, -27 deletions) <details> <summary>View changed files</summary> 📝 `README.md` (+3 -0) 📝 `mooncake-p2p-store/build.sh` (+2 -1) 📝 `mooncake-store/src/client.cpp` (+4 -0) 📝 `mooncake-store/tests/distributed_object_store_provider.py` (+6 -6) 📝 `mooncake-store/tests/stress_cluster_benchmark.py` (+16 -7) 📝 `mooncake-store/tests/stress_workload_test.cpp` (+9 -7) 📝 `mooncake-transfer-engine/include/error.h` (+1 -0) ➕ `mooncake-transfer-engine/include/transport/shm_transport/shm_transport.h` (+112 -0) 📝 `mooncake-transfer-engine/src/CMakeLists.txt` (+4 -4) 📝 `mooncake-transfer-engine/src/multi_transport.cpp` (+5 -2) 📝 `mooncake-transfer-engine/src/transport/CMakeLists.txt` (+3 -0) ➕ `mooncake-transfer-engine/src/transport/shm_transport/CMakeLists.txt` (+3 -0) ➕ `mooncake-transfer-engine/src/transport/shm_transport/shm_transport.cpp` (+174 -0) </details> ### 📄 Description In the previous version, local-to-local memcpy is implemented in RDMA transport, and this is done in sync. Therefore this leads bad performance of Mooncake Store put/get. I implemented a ShmTransport to handle this case. Memory copy is done in async now. In the future it can be extended to support shared memory. Also modified some bench code of Mooncake Store. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-29 21:26:37 +00:00