Msieve Lanczos scalability

[Jeff Gilchrist]

I'm not sure if people have posted this or not but with a recent C129 GNFS I was playing around with the -t option for the multi-threaded Lanczos and here are the estimated completion times from msieve on an 8 CPU Xeon 2.0 GHz system.

./msieve.core2 -v -nc2 -t 8
linear algebra completed 13150 of 940512 dimensions (1.4%, ETA 1h 9m)

./msieve.core2 -v -nc2 -t 4
linear algebra completed 13148 of 940512 dimensions (1.4%, ETA 1h16m)

./msieve.core2 -v -nc2 -t 2
linear algebra completed 13670 of 940512 dimensions (1.5%, ETA 1h54m)

./msieve.core2 -v -nc2
linear algebra completed 13156 of 940512 dimensions (1.4%, ETA 3h29m)

The jump from 1 to 2 is very good, and even a significant increase from 2 to 4 CPUs. You get a little faster from 4 to 8 but not really enough to tie up half the CPUs. That is probably largely due to the crappy memory bus of the Xeon and things are very saturated at that point. I wonder how well the improved bus the Core i7 would do with 8 CPUs.

Jeff.

[fivemack]

The 'benchmarks on i7' thread in the Hardware forum has timing results for msieve at various threadings on core2 / phenom / i7 which you might find interesting.