DL385 G2 and SLES 10.1 - Performance

[Rick Carragher]

Hi,

We've installed SLES10.1 (and .2) on a DL385 G2 server (Dual Core Opteron 2.40 GHz). We're running Java and DB2 on it. We're having some fairly severe performance issues that we can't seem to diagnose. The exact same tests on an IBM x3650 server (Dual Core Xeon 3.0 GHz) yield 300% better performance.

We're measuring performance by tracking the CPU time that it takes to perform a use case. For example, Use Case 1 takes 30 CPU seconds on the IBM machine, and 95 on the HP.

My question - is there something in the install of Linux that does not optimize a setting? I'm downloading the latest firmware - I'll apply that. But this seems QUITE odd.

Thanks,
Rick

[rick jones]

Are you binding threads/processes to processors? The DL385 G2, being based on Opteron processors is a "NUMA" (Non-Uniform Memory Access) system whereas your x3650, being based on Xeon is a "UMA" (Uniform Memory Access) system.

On your SLES10 installs there should be a command called "numactl" which can be used to show some of the numa characteristics of the system.

Other pseudo random things - memory is installed "evenly" across the two sockets in the DL385 yes?

Have you tweaked anything in the BIOS like "node interleave?" (IIRC it defaults to disabled.