Testing The Different Ubuntu 10.04 Kernels

Next, the Phoronix Test Suite turned from gaming to server performance. In the result above is from the Apache test profile where the Apache server was running on the test machine and it was making many requests on the local host to a static web-page. The Ubuntu -preempt kernel was slower than the -generic and -server kernels for Ubuntu 10.04 by about 12%. The Ubuntu -server and -generic kernels were right on par with the mainline Linux 2.6.32.9 kernel in terms of the number of requests per second that could be sustained. There's a regression introduced in the Linux 2.6.33 kernel and it is still active as of the Linux 2.6.34-rc1 release that severely hampers the Apache performance. The number of requests per second with Apache went from 10,000~12,000 RPS on the 2.6.32-based kernels to just 408 requests with the Linux 2.6.33 kernel. This, of course, like all tests was fully reproducible by the Phoronix Test Suite. This regression does not affect all systems running the post-2.6.32 kernels as our daily kernel tracker found some Apache performance improvements during the Linux 2.6.33 cycle when tested on Intel Atom hardware.

The PostgreSQL database performance was close between the three Ubuntu Lucid kernels with the best lead going to the -preempt kernel while oddly the -server kernel was beat out by -generic, but by a very small margin when averaging in the three test runs carried out by PTS. Looking forward to Ubuntu 10.10 or those that will manually be installing their own 2.6.33+ kernel on Lucid, there are some PostgreSQL boosts to be found with the EXT4 file-system in the newer releases.

When measuring the time it took to compile Apache 2.2.11 with GCC 4.4.3 using eight jobs (the Phoronix Test Suite does physical core count times two), there was not a dramatic difference between the tested kernels. For what it's worth, the -preempt kernel took 42 seconds to build Apache on this AMD Opteron quad-core setup where as the other kernels produced results of 40 seconds.