Linux Benchmarking

This page presents some information on Linux benchmarking; and in particular, UnixBench Version 5, a version of the venerable Byte benchmarks updated to handle multi-CPU systems.

Results

Here are some benchmarking results from UnixBench. Click on a system name to see the full results:

Some observations:

• OpenSUSE 10.3 is a bit slower than 10.2; it runs at about 75% - 85% as fast. Shame, since 10.2 was about 50% faster than 10.1.
• Linux gets a good speedup from utilising 2 CPUs: 160% (Core Duo) to 180% (Core 2 Duo). Considering that the system benchmark covers a pretty broad range, including filesystem tests that are actually hurt by 2 CPUs (due to greater contention; see the "USAGE" file), this seems pretty good.
• The 64-bit Linux install gains about 10%-15% over a 32-bit install on the same hardware.
• UnixBench 5.1 discards the worst 1/3 of the results for each test, in an attempt to improve consistency. This causes a gain in the test scores of about 1%-2% relative to 5.0.
• The graphics benchmarks, specially the 3D ones, are still very much in their infancy. Still, the Compaq — the only machine here with 3D hardware — clearly wins there.

UnixBench Version 5

UnixBench is a very ancient (1983!!!) benchmarking suite which originated at Monash University, and was then taken up and expanded by Byte magazine. It is a general-purpose benchmark designed to provide a basic evaluation of the performance of a Unix-like system (as compared to other Unix-like systems); it runs a set of tests to evaluate various aspects of system performance, and then generates a set of scores. (Other benchmarks, such as lmbench, do a more low-level analysis of a system's performance.) Given that UnixBench was not modified, as far as I can tell, in ten years, maybe there's something newer that everyone is using — but if so, then I haven't found it (please tell me!).

By agreement with David Niemi, the previous maintainer, I have taken on UnixBench. So after a 10-year gap, here are the latest updates to UnixBench:

• The main change I made in version 5.0 was to make the tests multi-CPU aware, so that they can generate useful results on systems with more than 1 CPU. The way it does this is pretty simplistic, but hopefully it may be useful.
• Version 5.1 adds some rudimentary 2D and 3D graphics benchmarks. They're crude — the 3D side is just "glxgears" — but hopefully useful to get a rough measure of a system's performance. These tests are optional and segregated into separate index groups.
• Version 5.1.1 sets the "LANG" environment variable. This fixes a major inconsistency between different system configurations caused by the differences in handling different character sets.
• Version 5.1.2 fixes a couple of bugs, most notably that if the installation pathname contained a space, it previously bombed out — now it works.

Running the benchmarks is pretty easy; download and unpack the tarball, enter the directory, and do "make; Run". There is more information in the included documentation files, which you can get from the links below.

Here is the code, along with the documentation (such as it is):

Previous UnixBench Versions

Versions prior to 5.0 can be downloaded from David Niemi's site. Unixbench 5.0 was derived from version 4.1.0.

Contact Information

You can email me at:

    johantheghost at yahoo period com

Ian Smith, December 2007