About :: TODO :: Blog :: RSS :: Old blog :: Projects :: GIT :: Gallery :: Notes

Thu, 26 Oct 2006

Gigabit send/recv benchmark for netchannels and sockets using small packets. Complete rout.




Netchannel with userspace network stack behaves much faster.
Receiving is about 8 MB/sec faster. Receiving CPU usage is 3 times less (90% socket code vs. 30% netchannels+unetstack).

Sending is 10 MB/sec faster. Sending CPU usage is 5 times less (upto 50% vs. upto 10%).

Number of syscalls is about 10 times less for netchannels.

Update: I want to say that with increased reading/writing size netchannels with userspace network stack behave exactly like with 128 byte chunks, and socket test's CPU usage decreases significantly. They become equal with about 1024 byte chunk size.
Userspace network stack does not use hardware checksumm offloading and use C-coded one instead, it also has additional memory copy, which should be eliminated.

/devel/networking :: Link / Comments ()