Linux 2.6.32 Kernel Benchmarks

[kraftman]

I'm starting to think that those big boys should really start testing and profiling their crap. I wonder how 2.6.24 would compare. Good article.

They actually do, but they also do proper benchmarks. I wonder if generic kernels were tested here and still ext4 vs different mount options (or mentioned commit which changed behavior) and maybe even different settings like NO_NEW_FAIR_SLEEPERS which is default in Karmic and it affects benchmarks numbers. Btw, what are you talking about if it was planned change?

And like Michael pointed:

The very significant drop in PostgreSQL's performance in the Linux 2.6.32 kernel with default options can be attributed to this lone Git commit that is for a fix to address cache flushing in ext4_sync_file for the EXT4 file-system. This commit improves data integrity in the event of a power loss or other problem, but carries a high disk performance penalty.

[Luis]

It should be noted than on 2.6.32, both the CPU and the I/O schedulers have been tuned for desktop usage (i.e, low latency), which means a drop in throughput in many cases. But desktop users should "feel" things faster, more responsive.

So nothing to worry about here, on the contrary, it just proves that the kernel devs did a bold movement in favour of desktop users.

[Veerappan]

Well, CK and Dark Shikari. http://x264dev.multimedia.cx/?p=185

@Phoronix team:

Not lowest. Either 'worst', or 'highest'.

Agreed. I looked at the graph, then the comments, and then back at the graph a few times... The average CPU utilization is higher almost across the board, but the PEAK utilization was lower for 2.6.32.

So I guess Michael's technically right, but the text analysis doesn't lead one to look at the average numbers, both of which are important in video playback.

[kraftman]

It should be noted than on 2.6.32, both the CPU and the I/O schedulers have been tuned for desktop usage (i.e, low latency), which means a drop in throughput in many cases. But desktop users should "feel" things faster, more responsive.

Yes, but I/O scheduler will be probably set for throughput for final 2.6.32. Afaik it's set for low latency in -rc and some benchmarks suffer from this. It should be optimized and set default for 2.6.33.

So nothing to worry about here, on the contrary, it just proves that the kernel devs did a bold movement in favour of desktop users.

Yes, if someone wants to have good results in some benchmarks just mount file system using different mode or options.

[Apopas]

I wonder if this boost is just for x264 or if we are gonna see sweet results with lame and oggenc as well

[Kano]

Maybe 2.6.31+BFS and 2.6.32+BFS should be added to compare.

[kraftman]

Yes, but I/O scheduler will be probably set for throughput for final 2.6.32. Afaik it's set for low latency in -rc and some benchmarks suffer from this. It should be optimized and set default for 2.6.33.

I/O scheduler is set for low latency in the final 2.6.32 and this can have impact on some benchmarks. From kernel newbies:

In this release, the CFQ IO scheduler (the one used by default) gets a new feature that greatly helps to reduce the impact that a writer can have on the system interactiveness. The end result is that the desktop experience should be less impacted by background IO activity, but it can cause noticeable performance issues, so people who only cares about throughput (ie, servers) can try to turn it off echoing 0 to /sys/class/block/<device name>/queue/iosched/low_latency. It's worth mentioning that the 'low_latency' setting defaults to on.

http://kernelnewbies.org/LinuxChange...231eaf58a63ed8

[Picander]

Starting by looking at the CPU usage during the playback of a 1080p H.264 video file, the Linux 2.6.32 kernel had the lowest overall CPU usage when using X-Video with MPlayer. However, the CPU usage was only less by 2%.

Looks like there is an error! The graph says
2.6.30 = 32.3
2.6.31 = 37.2
2.6.32 = 39.1

[Apopas]

I installed 2.632 just now and run some tests to compare 2.6.31-r6 vs 2.6.32.
I repeated each one 4 times and these are the average scores:

lame
0m59.607s ---> 2.6.31-r6
0m59.639s ---> 2.6.32


oggenc
0m24.844s ---> 2.6.31-r6
0m24.480s ---> 2.6.32


x264
1m50.563s with 30.80 fps ---> 2.6.31-r6
1m51.355s with 30.58 fps ---> 2.6.32


In other worlds bullshits. Where's the speed improvement in x264?
Should I enable something specific in kernel?

[Kano]

Well BFS increases compile speed by about 5% with the same number of threads. 7zip should be also a good benchmark for this, but single core benchmarks are useless to show multicore speed.