Linux 2.6.24 Through Linux 2.6.33 Benchmarks

[xianthax]

Out of interest, where is the recommended server configuration instructions. You imply that every LAMP server administrator is going to have more or less the same config. I don't believe there is a secret brotherhood that ensures through verbal tradition every administrator has the same information. Hence, there has to be a few online references as to how to do it right.

I am sure a default vs best-of-class server config vs best-of-class desktop config would be interesting.

Are you willing to stand forward and declare yourself as the configurator of the best-in-class server config? Is there any reader interested in the best-of-class desktop config?

Nothing is sacred, even define and contribute more correct server oriented tests if you have the interested..

Matthew

PS, I have had the same discussion with multi national corporations about more or less the same thing. I don't think they made a document in te end either.

No i'm certainly not an expert on kernel configuration nor do i claim to be. There also is no "best in class" configuration, the ideal configuration settings will be dependent on usage pattern. In fact this is an area where the test suite could likely provide some serious benefit, running a spectrum of server/desktop tests on a kernel and varying various kernel configuration options could be very valuable.

That all being said there certainly is a documented starting place, the kernel configuration settings for common server distros. The test suite could prove very valuable to improve their default configuration options.

The point of my post is that this article provides no actionable information. It isn't going to make anyone choose a particular kernel to gain a performance increase as most of the performance difference come down to kernel defaults rather than increased or decreased efficiency in code paths.

[rekiah]

I used to be an avid reader of Phoronix benchmarks... I found very interesting. My current opinion is that they're not very serious/reliable to say the least.

Anyway, regarding those "regressions" in .33 with PostgreSQL see:

http://www.mail-archive.com/pgsql-pe.../msg34841.html

"The pgbench TPS figure Phoronix has been reporting has always been a fictitious one resulting from unsafe write caching. With 2.6.32 released with ext4 defaulting to proper behavior on fsync"

[mtippett]

I used to be an avid reader of Phoronix benchmarks... I found very interesting. My current opinion is that they're not very serious/reliable to say the least.

Anyway, regarding those "regressions" in .33 with PostgreSQL see:

http://www.mail-archive.com/pgsql-pe.../msg34841.html

"The pgbench TPS figure Phoronix has been reporting has always been a fictitious one resulting from unsafe write caching. With 2.6.32 released with ext4 defaulting to proper behavior on fsync"

Again, the test is for measuring a system. pgbench is just something that exercises some paths. The regression is simply a change. It has become generally known that the cost of fsync implementation is critical for a lot of the database tests.

The TPS figure is the TPS figure for databases running on those systems. As Greg mentions in the first part of his thread. Some interesting benchmark news today suggests a version of ext4 that might actually work for databases is showing up in early packaged distributions: <phoronix reference removed>

The PTS benchmark are being tracked by them and it appears that although it isn't widely communicated, most filesystems that have high performance numbers for these tests are misconfigured. That is the critical part of the value that the PTS tests are adding. Sxooter's post in this thread re-enforces this idea too.

Perhaps statements should be made asserting that anything higher than 100 to 400 in either the SQLITE or pgbench testing indicates that the system is not suitable for database use. That would be interesting. I assume that the first reference to that would get half a dozen highly colorful threads started by those that don't understand that fast isn't really fast.

I don't think that Michael has ever made assertions about *right* or *wrong* implementation in the same way that Greg has. Michael is merely reporting the delta is present. Each reader will need to dig into the relevance of the change.

[mtippett]

No i'm certainly not an expert on kernel configuration nor do i claim to be. There also is no "best in class" configuration, the ideal configuration settings will be dependent on usage pattern.

...

The underlying issue is that *no-one* stands forward willing to state what is and what is not tuned. If *no-one* stands forward, and *no-one* documents it, then we get left with the desktop-configured defaults. The only groups that implied in coming forwards and contributing to the configuration are the desktop crowd. (Although I would expect the default kernel config is more server oriented, but don't have a best-practice to measure that against).

If people want to come forward and be involved in a kernel configuration bake-off, please come forward.

[karl]

... Michael is merely reporting the delta is present. Each reader will need to dig into the relevance of the change.

+1
I agree completely with your post. Deltas are important since they sometimes reveal bugs (and not always with the new code but actually with the old code).

....(Although I would expect the default kernel config is more server oriented, but don't have a best-practice to measure that against)....

I remember that I've read some time ago that in fact the kernel comes by default optimized for _desktop_ usage. (well actually _optimized_ is a strong word). And if you think about it, it makes sense: server guys will want to tune their system base on their _load_ patterns since I don't think there is a configuration that produces _best_ results for all load patterns.

[maccam94]

Out of interest, where is the recommended server configuration instructions.
...
Are you willing to stand forward and declare yourself as the configurator of the best-in-class server config?

I would like to point out that there are two preconfigured kernels in the Ubuntu repositories, linux-image-2.6.xx-generic (default for the Desktop metapackage) and linux-image-2.6.xx-server (the default for the Server metapackage). For running server usage benchmarks, it would probably be a good idea to use the server kernel. It might also be useful to see across the board comparisons between the server and generic kernels.

[xianthax]

The underlying issue is that *no-one* stands forward willing to state what is and what is not tuned. If *no-one* stands forward, and *no-one* documents it, then we get left with the desktop-configured defaults. The only groups that implied in coming forwards and contributing to the configuration are the desktop crowd. (Although I would expect the default kernel config is more server oriented, but don't have a best-practice to measure that against).

If people want to come forward and be involved in a kernel configuration bake-off, please come forward.

What i'm saying is that phoronix is in the best position to become the authority you seek. The framework you've created puts ya'll in the best position to test tons of different configurations that no one currently has the time or resources to manually test and even identify configuration performance dependencies that no one would initially think would impact each other and thus have never thought to try.

Actually one of the only conclusions i drew from this article is that the default kernel configuration is becoming more focused on the desktop as time passed.