Saturday, April 21

Geek

Kurumi And Midori

Midori has now been set up.  Benchmarky time!

Kurumi is a 3.2GHz Xeon E3 1230; 4 cores with hyper-threading for 8 logical cores, and turbo mode up to 3.6GHz.

Midori is a 2.1GHz Opteron 6272; two dies on a single package, each with 4 modules with 2 cores each but with some shared resources, for 16 physical cores, and turbo mode up to 3GHz.

The Xeon E3 core is more efficient than the Opteron core, as well as running at higher clock speeds.  But the Opteron gives four times as many cores at a very modest price.

Let's see how they run for single-threaded applications, moderately-threaded ones, and then huge horrible messes of applications.  These scores are aggregate throughput on my little Python benchmark.

Threads Kurumi Midori
1 385 289
2 771 564
4 1216 1085
8 1627 1819
16 1629 2116
32 1626 2178

Hmm.

Things are pretty much as expected up to four threads; the higher clocks and stronger cores on the Xeons outrun the Opteron.  At eight threads, the Xeon delivers a better than expected result from hyperthreading, but the Opteron delivers an 80% scaling factor and outstrips the Xeon.

At 16 threads there's no improvement for the Xeon...  But there's not much from the Opteron either.  At 32 threads there's little change.

So...

The Opteron is 25% slower for single-threaded applications, and 30% faster for multi-threaded ones.  The performance at 16 threads is disappointing; that's probably due to a combination of the lower clock speed when all cores are active and the shared resources between core pairs.  In fact, the performance gained by the extra cores in the Opteron is less than the gain from hyperthreading in the Xeon.  That's a surprise, and one I hope AMD can address in future versions of the chip.

But it supports four times as much memory, so it's a much better chip for large servers.

A dual Xeon E5 2620 (the other option I was looking at) would be as much as 40% faster, based on these benchmarks, but it would also be about 40% more expensive (for the same amount of memory and storage). And - here's the tricky part - up through 8 threads it would offer no advantage at all.  A dual E5 2620 would offer 12 full cores vs. 8 dual-core modules, but its lower maximum clock speed would offset the stronger architecture, and it would offer no performance advantage until you get beyond eight threads.

A dual Opteron 6272 would be about 40% faster again - at least, for heavily threaded workloads - and at the same price as the dual E5 2620.  But I've decided to go for dual servers once we get to that point.

Posted by: Pixy Misa at 05:47 PM | No Comments | Add Comment | Trackbacks (Suck)
Post contains 456 words, total size 4 kb.

Comments are disabled. Post is locked.
42kb generated in CPU 0.03, elapsed 0.1766 seconds.
50 queries taking 0.1555 seconds, 275 records returned.
Powered by Minx 1.1.6c-pink.