0.思路
最终确定为:不同规模的输入长度,在已经注册好的大池子内,按照 8KB 去随机拿不重叠 block。
1. 结果
flagcx
connector=flagcx role=server device=gpu block=8.0 KB pool=8.0 GB
------------------------------------------------------------------------
size= 16.0 MB | WRs= 2048 (block=8.0 KB) | lat= 2.698 ms | BW= 5.79 GB/s ( 49.75 Gbps)
size= 64.0 MB | WRs= 8192 (block=8.0 KB) | lat= 10.396 ms | BW= 6.01 GB/s ( 51.64 Gbps)
size= 256.0 MB | WRs= 32768 (block=8.0 KB) | lat= 39.821 ms | BW= 6.28 GB/s ( 53.93 Gbps)
size= 1.0 GB | WRs= 131072 (block=8.0 KB) | lat= 164.435 ms | BW= 6.08 GB/s ( 52.24 Gbps)
------------------------------------------------------------------------
Done.
mooncake
connector=mooncake role=server device=gpu block=8.0 KB pool=8.0 GB
------------------------------------------------------------------------
WARNING: Logging before InitGoogleLogging() is written to STDERR
E0615 02:22:08.092213 871382 transfer_metadata.cpp:863] Local segment descriptor not found
size= 16.0 MB | WRs= 2048 (block=8.0 KB) | lat= 1.902 ms | BW= 8.22 GB/s ( 70.57 Gbps)
size= 64.0 MB | WRs= 8192 (block=8.0 KB) | lat= 7.911 ms | BW= 7.90 GB/s ( 67.86 Gbps)
size= 256.0 MB | WRs= 32768 (block=8.0 KB) | lat= 31.864 ms | BW= 7.85 GB/s ( 67.39 Gbps)
size= 1.0 GB | WRs= 131072 (block=8.0 KB) | lat= 196.079 ms | BW= 5.10 GB/s ( 43.81 Gbps)
------------------------------------------------------------------------
Done.
nixl
------------------------------------------------------------------------
2026-06-15 02:22:54 NIXL INFO _api.py:361 Backend UCX was instantiated
2026-06-15 02:22:54 NIXL INFO _api.py:251 Initialized NIXL agent: server
size= 16.0 MB | WRs= 2048 (block=8.0 KB) | lat= 5.176 ms | BW= 3.02 GB/s ( 25.93 Gbps)
size= 64.0 MB | WRs= 8192 (block=8.0 KB) | lat= 19.783 ms | BW= 3.16 GB/s ( 27.14 Gbps)
size= 256.0 MB | WRs= 32768 (block=8.0 KB) | lat= 78.476 ms | BW= 3.19 GB/s ( 27.36 Gbps)
size= 1.0 GB | WRs= 131072 (block=8.0 KB) | lat= 317.682 ms | BW= 3.15 GB/s ( 27.04 Gbps)
------------------------------------------------------------------------
Done.
