On 14 Oct 2005, at 8:06 am, mike wrote:
Basically, I'd be looking at an Intel 830D (3.0ghz, dual-core, EM64T) processor-based system, with 1 gig of ram and SATA HD vs. a single-core Opteron 1.8ghz (or a dual-processor NOT dual-core Opteron 1.8ghz) system, same HD and same RAM. Has anyone had the opportunity to benchmark this, or have any real experience with changing the underlying platform?
Our experience with IBM HS20 blade servers shows the 2.4 GHz Opteron blows the 3.2 GHz EM64T out of the water on our bioinformatics codes, which are predominantly limited by integer performance and memory bandwidth. And uses a lot less power (this is significant when you have about 1000 processors in the cluster!). No more EM64T for us, at least for a while.
One thing to be careful with, if it's a dual CPU opteron machine that you're benchmarking, is that you need a really up to date kernel that gets the NUMA topology right; our initial benchmarks were disappointing, and it turned out that the the numa topology stuff in the kernel was exactly backwards, so every process was always accessing memory local to the other CPU. This has been fixed in recent kernels.
Tim