Benchmarking The Linux 2.6.24 Through 2.6.29 Kernels

[phoronix]

Phoronix: Benchmarking The Linux 2.6.24 Through 2.6.29 Kernels

With the release yesterday of the Linux 2.6.29 kernel, we have set out to explore how the desktop performance has evolved over the past six major kernel releases. On a few occasions in the past we have provided kernel benchmarks (at one point even benchmarking 12 kernels), but this time around we have included nearly two dozen benchmarks using the Phoronix Test Suite. How has the Linux performance evolved since the release of the Linux 2.6.24 kernel back in early 2008? Well, simply put, the Linux 2.6.29 kernel in a few areas does pack some serious performance boosts.

http://www.phoronix.com/vr.php?view=13622

[ssam]

love the little arrows, much easier to skim through now.

but i'd still like to see error bars.

if you do multiple runs with the same kernel there will be some variation. if there is a 1% run to run variation, then a 0.1% kernel to kernel variation is insignificant.

also would be nice to see boot, and program launch times.

[numasan]

Nice article, very interesting. There's just one typo:

For those unfamiliar with GraphicsMagick, this open-source package was originally derived from GraphicsMagick, but since then it has adopted an incredible number of changes. In the most recent release of GraphicsMagick, it now uses OpenMP to provide multi-threaded image processing.

I guess you missed an "ImageMagick" in there
Great to see those kernel improvements with graphical operations. I didn't think the kernel had impact with that..

[smitty3268]

Yes, I wonder what caused that, it's very surprising.

Maybe a big lock got broken up in the kernel, and ImageMagick is able to utilize a 2nd core now that was previously just always waiting on syscalls from the 1st one?

That doesn't really seem right.

[drees]

When you see doubling of performance, it would be nice to see some analysis of why those tests improved so much.

For IO bound tests (like SQLite) was there less IO? Or did IO complete faster?

For CPU bound tests (like OpenSSL & GraphicsMachine) was CPU utilization higher?

Seeing improvements of 100% or more is quite unexpected indeed. There either was a significant improvement in the kernel, or a bug in the benchmark routines.

[Kano]

"On a clean Ubuntu 8.10 (x86_64) system we had installed the Linux 2.6.24, Linux 2.6.25, Linux 2.6.26, Linux 2.6.27, Linux 2.6.28, and Linux 2.6.29 kernels"

I can not verify those results with my system. Of course I do not use Ubuntu 8.10 but something Debian 5.0 based and those results are completely different. I only tested the OpenSSL benchmark, with 2.6.29 and running fglrx 9.2 i got 87.87 and with 2.6.28 i got 87.97, that's a very minor diff. Both tests with live mode from usb storage with E8400 cpu at stock speeds, running in performance mode of speedstep.

[Linuxhippy]

> For CPU bound tests (like OpenSSL & GraphicsMachine) was CPU
> utilization higher?

Ok, I can understand the IO bound tests, maybe SQLLite hit a corner case somewhere triggering very ugly behaviour.

What I don't understand are the CPU test. Are those really the same binaries? Could it be that those binaries have been compiled with OpenMP support, but <=2.6.28 were single-core kernels?

- Clemens

[Michael]

What I don't understand are the CPU test. Are those really the same binaries?

Yep, they are the same exact binaries. No tests were reinstalled or modified between kernels.

[Kano]

But why do i see basially no diff between 2.6.28 + 2.6.29? I always have got high results with OpenSSL, but you have got lots of low results, something weird must be going on there.

[Michael]

But why do i see basially no diff between 2.6.28 + 2.6.29? I always have got high results with OpenSSL, but you have got lots of low results, something weird must be going on there.

I don't know, yet. Still running some other tests to see on what hardware the speed differences can be found on.