MPI-and-threads post-SVN900

[fivemack]

(3.2GHz, DDR1600 memory, no overclocking)

886: 2463 / 4579 / 6288 / 7833 / 8876 / 10249
1.00 / 1.85 / 2.53 / 3.15 / 3.58 / 4.12

923: 3809 / 7048 / 9901 / 12362 / 14144 / 16176
1.00 / 1.85 / 2.60 / 3.23 / 3.70 / 4.23

Speed ratios 1.55 / 1.54 / 1.57 / 1.58 / 1.59 / 1.58

Notice the better scaling in v.886 on i7/3930 vs Haswell, presumably because that version goes to main memory more quickly and there's more memory bandwidth for it to work with.

Note also that the new version is a bit faster on Haswell than on the i7/3930 at equivalent thread counts, and by more than the ratio in clock speeds, though the advantage goes down 1.23 / 1.20 / 1.17 / 1.14 ; I suspect this is because Haswell has better low-level caches. This reduces substantially my desire to get an i7/4930 when they are launched.

[fivemack]

With SVN923, it's worth using six threads on a six-core submodule. 2x4 and 4x2 grids are better than 1x8 or 8x1. It looks about twice as fast as the Haswell machine.

grid / threads / dimensions-per-second (more is better)
1,8 2 105.2
2,4 2 120.6
4,2 2 121.6
8,1 2 103.4

1,8 3 129.2
2,4 3 149.6
4,2 3 150.3
8,1 3 125.5

1,8 4 146.9
2,4 4 175.2
4,2 4 172.8
8,1 4 142.5

1,8 5 160.2
2,4 5 192.5
4,2 5 187.1
8,1 5 152.4

1,8 6 165.9
2,4 6 201.5
4,2 6 199.3
8,1 6 162.7

I will be running L2910 with 2,4 -t6 tonight to check that the speed remains this good on larger matrices.