benchmarks on dual i7-xeon

[fivemack]

This is a 2.26GHz dual-Xeon5520 Mac Pro running OSX.

Test is http://www.chiark.greenend.org.uk/~twomack/i7test.zip (a 130-digit GNFS; I'm afraid it's a 600MB download)

Threads : BLanczosTime
 1  10941
 2   5940
 3   4463
 4   3772
 6   3270
 8   3330
12 3505
16 3818

So scaling from 1 to 4 threads is pretty good, but beyond that gets diminishing returns quite quickly. I don't know how you get malloc to allocate an area on the memory physically attached to a specific CPU.

[jasonp]

Originally Posted by fivemack

So scaling from 1 to 4 threads is pretty good, but beyond that gets diminishing returns quite quickly. I don't know how you get malloc to allocate an area on the memory physically attached to a specific CPU.

The current code makes each thread malloc its own memory, in case the C library is NUMA-aware; hopefully that's enough, and at any rate page migration has been in the 2.6 kernel for a while now (dont know about Darwin)