I'm forwarding this email that I sent to current@ in the hopes that some performance minded people will tell me their results with this new scheduler infrastructure.

Thanks,
Jeff

---------- Forwarded message ----------
Date: Tue, 12 Jun 2007 19:28:39 -0700 (PDT)
From: Jeff Roberson <[EMAIL PROTECTED]>
To: [EMAIL PROTECTED]
Subject: Call for testers, amd64 only, new scheduler.

This is actually a ULE derivative that I have been discussing for some time. It is now stable enough that I would like more people to bang on it and give feedback about performance and tell me about any stability problems they encounter.

You can see in the following graph that the ramp-up times for the sysbench benchmark have improved although the tail has suffered a little:

http://people.freebsd.org/~jeff/sysbench.png

The new scheduler is FreeBSD-7.0-ULE-pcpulocks. On my dual cpu system the results are much less ambiguous. It's a 10-15% win on sysbench across the board. This also improves an artificial thread benchmark included with sysbench by 500% on an 8 way and gives us ebizzy numbers on par with or better than linux.

It also includes a mechanism that should improve bde's nfs buildworld times, although I have not yet verified that. Hopefully he'll try it as well. It should reduce CPU idle time while also negating the effects of thrashing other cpus run queues in the idle thread.

The new patch is at: http://people.freebsd.org/~jeff/schedsmp.diff

You have to change your scheduler to SCHED_SMP. x86 support is forthcoming. Depending on feedback and what re thinks I am going to consider replacing ULE entirely with SCHED_SMP or committing this as a seperate scheduler for 7.0.

Thanks,
Jeff
_______________________________________________
[EMAIL PROTECTED] mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-current
To unsubscribe, send any mail to "[EMAIL PROTECTED]"
_______________________________________________
freebsd-performance@freebsd.org mailing list
http://lists.freebsd.org/mailman/listinfo/freebsd-performance
To unsubscribe, send any mail to "[EMAIL PROTECTED]"

Reply via email to