Kenneth Chen announced the Linux Kernel Performance Project, "we decided to run a large set of benchmarks covering core components of the Linux kernel (virtual memory management, I/O subsystem, process scheduler, file system, network, device driver, etc) on a regular basis." Results from 13 benchmark utilities are displayed in both table and graph format, run on several different servers. The 2.6.9 kernel [story] is used as the baseline, compared against various more recent kernels up to 2.6.13-rc1. Ken explains:

"Our goal is to work with the Linux community to further enhance the performance of the Linux kernel. The data available on the site allows community members to closely track performance gains and losses with every version of the kernel. Ultimately, we hope that this data will result in performance increases in Linux kernel development."

From: "Chen, Kenneth W" [email blocked]
To: [email blocked]
Subject: [announce] linux kernel performance project launch at sourceforge.net
Date:	Thu, 14 Jul 2005 13:21:25 -0700

I'm pleased to announce that we have established a linux kernel
performance project, hosted at sourceforge.net:

http://kernel-perf.sourceforge.net

As much as discussed various time in the past on LKML that Linux
kernel needs a systematic and disciplined way to measure and track
kernel performance on a regular basis.  To do that, we decided to
run a large set of benchmarks covering core components of the Linux
kernel (virtual memory management, I/O subsystem, process scheduler,
file system, network, device driver, etc) on a regular basis.
Benchmarks are run on a variety of platforms (4P Intel Xeon processor,
2P Xeon, several ia64 server boxes etc) every week, measuring the
latest snapshot of Linus' git development tree. Comprehensive performance
data from our tests will be published for easy access.

Our goal is to work with the Linux community to further enhance the
performance of the Linux kernel. The data available on the site allows
community members to closely track performance gains and losses with
every version of the kernel. Ultimately, we hope that this data will
result in performance increases in Linux kernel development.

The benchmark result pages are populated with a few benchmarks at the
moment.  In the coming weeks, we will be populating more benchmark data.
Happy surfing and hacking!!


		Ken Chen [email blocked]
		Intel Open Source Technology Center



From: Andi Kleen [email blocked]
Subject: Re: [announce] linux kernel performance project launch at sourceforge.net
Date:	15 Jul 2005 00:18:03 +0200

"Chen, Kenneth W" [email blocked] writes:

> I'm pleased to announce that we have established a linux kernel
> performance project, hosted at sourceforge.net:
> 
> http://kernel-perf.sourceforge.net

That's very cool. Thanks a lot.

Would it be possible to add 2.4.30 numbers and perhaps one or two 
distro kernels (let's say RHEL3/4, SLES8/9) to the graphs 
as data points for comparison? These are all very tuned
kernels and would show where mainline is worse than them.

Also how did you run netperf? Locally or to some other machine? 
Perhaps that should be documented.

Some oprofile listings from a few of the test runs would be also nice.

-Andi



From: "Chen, Kenneth W" [email blocked]
Subject: RE: [announce] linux kernel performance project launch at sourceforge.net
Date:	Thu, 14 Jul 2005 15:24:45 -0700

[email blocked] wrote on Thursday, July 14, 2005 3:18 PM
> "Chen, Kenneth W" writes:
> > I'm pleased to announce that we have established a linux kernel
> > performance project, hosted at sourceforge.net:
> > 
> > http://kernel-perf.sourceforge.net
> 
> That's very cool. Thanks a lot.

Thank you.


> Would it be possible to add 2.4.30 numbers and perhaps one or two 
> distro kernels (let's say RHEL3/4, SLES8/9) to the graphs 
> as data points for comparison? These are all very tuned
> kernels and would show where mainline is worse than them.

We did have a distro kernel in the graph originally and later decided
to go with pure mainline kernels for consistency.  We will see what we
can do to add them in the future.


> Also how did you run netperf? Locally or to some other machine? 
> Perhaps that should be documented.

Yes, that was netperf, running locally.  Thanks for the suggestion, I
will document that appropriately.


> Some oprofile listings from a few of the test runs would be also nice.

That is in the works.  We will upload profile data.  I'm having problem
with oprofile on some versions of kernel and that is being investigated
right now.

- Ken



From: Andi Kleen [email blocked]
Subject: Re: [announce] linux kernel performance project launch at sourceforge.net
Date:	Fri, 15 Jul 2005 00:27:48 +0200

> > Some oprofile listings from a few of the test runs would be also nice.
> 
> That is in the works.  We will upload profile data.  I'm having problem
> with oprofile on some versions of kernel and that is being investigated
> right now.

If you run statically compiled kernels you could as well use the
old style readprofile.  It just doesn't work with modules.

-Andi

• Archive of above thread