Disable memcpy by default and improve stress workload test

[xiaguan]

Changes

1. Disable memcpy by default for better RDMA performance

• Set MC_STORE_MEMCPY default to false in transfer_task.cpp
• RDMA performance is more stable than memcpy operations

2. Improve stress workload test

• Rewrite stress_workload_test.cpp as standalone benchmark
• Add separate throughput calculations for PUT and GET operations
• Include comprehensive performance metrics:
  • Latency percentiles (P50, P90, P95, P99)
  • Operation-specific throughput (ops/sec)
  • Data throughput (MB/s) for PUT and GET separately
• Add configurable parameters for threads, operations, buffer sizes

[Copilot]

Pull Request Overview

This PR disables memcpy by default to stabilize RDMA performance and transforms the stress workload test into a standalone, configurable benchmark with detailed metrics.

• Disable MC_STORE_MEMCPY default to false in transfer logic.
• Rewrite stress_workload_test.cpp as a benchmark with separate PUT/GET throughput, latency percentiles, and configurable parameters.
• Update test CMake configuration to remove GTest and link GFlags.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 2 comments.

File	Description
tests/stress_workload_test.cpp	Converted from gtest to standalone benchmark; added flags and metrics
tests/CMakeLists.txt	Removed gtest deps and added gflags linkage
src/transfer_task.cpp	Changed default MC_STORE_MEMCPY setting to disabled

Comments suppressed due to low confidence (1)

mooncake-store/tests/stress_workload_test.cpp:24

• [nitpick] The flag name test_operation_nums is a bit ambiguous; consider renaming it to num_operations_per_thread or operations_per_thread for clearer intent.

DEFINE_int32(test_operation_nums, 100, "Number of operations per thread");

[ykwd]

RDMA performance is more stable than memcpy operations

This is rather counter-intuitive. I wonder why RDMA has better performance than local memcpy?

[xiaguan]

RDMA performance is more stable than memcpy operations

This is rather counter-intuitive. I wonder why RDMA has better performance than local memcpy?

#467 (comment)