Paul Durrant writes:
> Just doing some network TX perf. measurement on a Dell 1850 dual Xeon
> box and I see that DMA mapping my buffers seems to be incredibly
> costly: each mapping taking >8us, some >16us.
I don't see anything this bad on an AMD based systems when running a
64-bit kernel. I realize that it is unlikely, but is there any chance
that this could be something which affects only Intel CPUs?
Here is some output from your script when run on a 2.0GHz Athlon64x2
running a 64-bit snv_41 kernel when using my 10GbE driver (without
TSO, with a 1500 byte MTU). I didn't bother to copy your test, and
just ran with normal traffic generated by netperf:
getpfnum
value ------------- Distribution ------------- count
512 | 0
1024 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ 1080289
2048 | 1008
4096 | 86
8192 | 103
16384 | 126
32768 | 23
65536 | 1
131072 | 0
Drew
_______________________________________________
perf-discuss mailing list
[email protected]
