[netperf-talk] global question concerning Netperf test and SMP support

[Rick Jones]

> I still can not explain the negative CPU utilzation in remote. What
> considerations could we do with the confidence interval? Is it
> acceptable? Could we ignore it?

CPU utilization should never be negative.

I suppose since the system of interest is the one on which netperf is 
running (?) the remote CPU util could be ignored for the purposes of 
confidence intervals - in that case I would just remove the -C from the 
netperf command line.

But CPU util going negative is a bug and makes me uncomfortable.

> Next, the changes are just concerning the socket size and the message size.
> We clearly observe the differences between the different configuration.
> The best throughput is obtained for 56k socket size. These performance
> and optimal configuration are dependent of the equipment.
>
> Why does we configure 56k and netperf print 56k x 2 as socket size. Is
> it what you called linuxism here?
> http://www.netperf.org/pipermail/netperf-talk/2006-October/000176.html

Yes.  Linux uses the value of SO_SNDBUF/SO_RCVBUF as both a "data byes" 
and "buffer bytes" limit, so up to net.core.[rw]mem_max it wil take 2X 
the user's requested size, which presumes then a "buffer efficiency" no 
worse than 50%.

rick