[Xen-devel] SEDF regression

["Diwaker Gupta" <diwaker.lists@xxxxxxxxx>]

I know that SEDF is not the default scheduler anymore, but I still
wanted to bring this to attention if some folks are continuing with
SEDF. I've been doing benchmarking with SEDF with very simple web
workloads, and I'm seeing some very strange behavior that was not
present before.

The setup is as follows: I have a single VM (nosmp) apart from dom0
(nosmp). The VM runs a web server which is loaded using httperf from
another machine. I'm measuring web server throughput as a function of
the relative weights of domain-0 and domain-1.

I get a through of around 1200 req/s with the following slice/period
configuration:
dom0: 3ms out of 10ms
dom1: 7ms out of 10ms

But, with only around 600 req/s with the following:
dom0: 30ms out of 100ms
dom1: 70ms out of 100ms

In both cases, the extra time flag is set to 0 and the latency hint is
also set to 0.

Thats half the throughput with a roughly equivalent set of weights. I
haven't isolated the exact changeset that might be the culprit, but
until about a few months back, this didn't happen. If anyone has seen
similar behavior, please let me know -- I'd be really interested in
hearing about it.

Thanks,
Diwaker
--
Web/Blog/Gallery: http://floatingsun.net/blog

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel