Re: [Qemu-devel] [PATCH] implement moderation registers for e1000

[Luigi Rizzo]

On Fri, Feb 08, 2013 at 11:59:12AM +0100, Stefan Hajnoczi wrote:
...
> >>     http://info.iet.unipi.it/~luigi/papers/20130206-qemu.pdf is 404.
> >>
> >>
> >> sorry, fixed now.
> >> And, will resubmit a fixed patch with uninit and fixed braces in the
> >> if() statement.
> >>
> >> I am happy to make this default to off. But it would be good if you could
> >> actually give it a try. Note that linux and FreeBSD (and presumably 
> >> windows)
> >> do use moderation by default so enabling the emulation of the
> >> registers makes the guest OS behave differently (and closer to bare metal).
> >>
> >> To test that the patch itself does not cause regression in the emulator
> >> one should also turn off moderation in the guest (one of the tests i
> >> have run).
> >
> > I think making the default on is fine, but you need to add compatibility
> > options to leave it off in older machine types (pc-1.4 and earlier).
> 
> Latency regression.  Would need to see real results to understand how bad it 
> is.

most of the numbers in the paper are for FreeBSD,
not sure if they qualify as "real" here :)

For Linux, in guest-host configuration, we have these numbers for TCP_RR

        CONFIGURATION           TCP_RR (KTps)

        1 VCPU TX itr=0         17.7
        1 VCPU TX itr=100        7.3
        2 VCPU TX itr=0         13.9
        2 VCPU TX itr=100        7.4

These are computed forcing the value in the itr register.
However, by default,
linux seems to dynamically adjust the itr values depending
on the load so it is difficult to make repeatable experiments,
This is why I'd encourage you to run some tests with what you
think is an appropriate configuration.

cheers
luigi