search

fixstars / lightning-kit

Reactive data transfer framework
Other
1 stars 0 forks source link

tcp with dpdk gpudirect performance #13

Open masaruito110 opened 5 months ago

masaruito110 commented 5 months ago

Performance on VMWare VM.

VM has RTX A6000 Ampere arch, ConnectX6, PCIe gen2, CPU is Intel(R) Xeon(R) Silver 4214R CPU @ 2.40GHz

Used program is 722a917b4eebfd00f649dba893845912e7c1126a

process session/process chunk size [MByte] Gbps/session
1 1 1 42.92
1 1 2 52.61
1 1 4 56.45
1 1 8 57.31
1 1 16 58.14
1 1 32 57.86
1 1 40 57.63
1 1 48 57.35
1 1 60 57.95
masaruito110 commented 5 months ago

when chunk size is over 10MByte, sometimes illegal memory access happens. The performance looks saturated over 4MByte chunk size. I think 4MByte chunk size is enough.