[netperf-talk] netperf2.4.4 on linux, socket size issue ?

[Rick Jones]

Andrew Gallatin wrote:
> Rick Jones writes:
>  > beginning.  So, while in both cases above, netperf is reporting the same 
>  > socket buffer sizes, in one case, Linux has gone behind netperf's back 
>  > and ended-up using even larger socket buffers.  This is another in a 
>  > series of things where Linux wanted to be different from everyone else 
>  > and in so doing changed/violated some basic assumptions about stack 
>  > behavour implicit in netperf. (and perhaps other benchmarks too)
> 
> 
> I'm not generally a Linux fan, but I'd say the benefits of autotuning
> far outweigh the drawbacks.  If the application cares, it can set the
> buffer sizes itself.  FWIW, the autotuning has spread to FreeBSD, and
> even Solaris is talking about doing it.

I am probably just being a curmudgeon :)  What irks me most though is 
that by default it seems the autotuning is allowed to take a socket 
buffer much farther than an application is allowed to directly.

Also, at present, if one runs the omni tests over a 1GbE link, you will 
notice that the autotuning takes the socket buffer size to the max, 
which is (generally, depending on sysctl settings) much, Much MUCH 
larger than it needs to be - I see 4MB when all it really needs is 
128KB.  I keep meaning to reproduce that on a kernel.org kernel and 
report it to netdev but other priorities in the day job keep getting in 
the way :(

rick jones