mfakto: an OpenCL program for Mersenne prefactoring

[airsquirrels]

Originally Posted by kracker

.

Has anyone else been able to verify before/after performance changes from changing the memory allocation type?

[Rodrigo]

Thanks @wombatman and @lalera.

Rodrigo

[lalera]

my first results for v 0.14 and v 0.14p

[kracker]

Originally Posted by lalera

my first results for v 0.14 and v 0.14p

Try benchmarking with an actual workload instead of the selftests, the reason for the selftests are more of a "sanity check" rather than a benchmark or hardware tester.

[lalera]

hi,
the file i posted above (#1345) represents my work in the range i am doing
here is an example of a actual workload on the tf-wave

Factor=N/A,132961627,71,72

v 0.14
got assignment: exp=132961627 bit_min=71 bit_max=72 (3.60 GHz-days)
Starting trial factoring M132961627 from 2^71 to 2^72 (3.60GHz-days)
Using GPU kernel "cl_barrett15_73_gs_2"
Date    Time | class   Pct |   time     ETA | GHz-d/day    Sieve     Wait
Jan 24 17:08 | 4617 100.0% |  0.589   0m00s |    549.62    82485    0.00%
no factor for M132961627 from 2^71 to 2^72 [mfakto 0.14-Win cl_barrett15_73_gs_2]
tf(): total time spent:  9m 27.220s (547.89 GHz-days / day)

v 0.14p
got assignment: exp=132961627 bit_min=71 bit_max=72 (3.60 GHz-days)
Starting trial factoring M132961627 from 2^71 to 2^72 (3.60GHz-days)
Using GPU kernel "cl_barrett15_73_gs_2"
Date    Time | class   Pct |   time     ETA | GHz-d/day    Sieve     Wait
Jan 24 17:21 | 4617 100.0% |  0.578   0m00s |    560.08    82485    0.00%
no factor for M132961627 from 2^71 to 2^72 [mfakto 0.14-MGW cl_barrett15_73_gs_2]
tf(): total time spent:  9m 10.661s (564.37 GHz-days / day)

[lalera]

hi,
thanks to kracker for the program !
i am using it now

[PopGoesTheWease]

Originally Posted by airsquirrels

Has anyone else been able to verify before/after performance changes from changing the memory allocation type?

Just changed over to 0.14p and wow! On my workstation (dual socket Win7x64) there is a ~14% improvement. On my server (quad socket Win2k8R2) there is a ~20% improvement. Same model cards both in x16 slots.

[PopGoesTheWease]

Originally Posted by airsquirrels

Has anyone else been able to verify before/after performance changes from changing the memory allocation type?

Just changed over to 0.14p and wow! On my workstation (dual socket Win7x64) there is a ~14% improvement. On my server (quad socket Win2k8R2) there is a ~20% improvement. Same model cards both in x16 slots.

[airsquirrels]

Originally Posted by PopGoesTheWease

Just changed over to 0.14p and wow! On my workstation (dual socket Win7x64) there is a ~14% improvement. On my server (quad socket Win2k8R2) there is a ~20% improvement. Same model cards both in x16 slots.

Are your CPU cores running prime95 or otherwise loaded? Switching to on-GPU memory would significantly reduce the need for CPU use for memory transfers, which may have been stalling you. Also, if you are using older / slower main RAM that would also explain the high gains despite being in x16 slots.

[derekg]

Using an R9 390X on a 2.0 x8 slot throughput went from 624.15 GHZ-days per day to 696.91.

Derek

[fregames]

I just upgraded to windows 8 and performance has dropped to 7Ghz-Days/Day. I'm usually running at ~275 each. Is there any known issue that would cause this? I'm using 3 R9 270x cards.

--
Turns out I needed to install drivers. my bad