[ofa-general] Re: socket buffer accounting with UDP/ipoib

[Michael S. Tsirkin]

> Quoting Roland Dreier <rdreier at cisco.com>:
> Subject: Re: socket buffer accounting with UDP/ipoib
> 
> I did a quick hack to enable copybreak for UD packets up to 256 bytes
> (see below).  This is still missing copybreak for CM / RC mode.
> However I just wanted to see how it affected performance.  And the
> answer is that on my system (fast quad-core Xeon, 1-port Mellanox PCIe
> HCA) is that it didn't make any difference in small-message latency or
> throughput, at least none that I could measure with netpipe (NPtcp).

Not any benchmark would show an improvement: what we save with copybreak
is actually memory, which only has performance impact if you start reaching
RCVBUF size. And the savings are only if message size is below
the threshold, so you better set NDELAY to see any effect.

Try running a UDP benchmark with small message size and
NDELAY, and looking at number of UDP errors with netstat.

-- 
MST