An interesting observation on BRP times
9 Feb 2014 1:20:48 UTC
Topic 197375
(moderation:
Since we have run out of Arecibo BRP data I now only run Perseus data on my GPU. The BRP4s ran consistently in about 1.25 hours and the BRP5s ran between 4.25 and 5 hours. I run 3 tasks at a time and would normally have some of each going at once. Now that it is pure BRP5 the times are a steady 4.25 hours. That's good for the RAC.
I feel the 2 different apps must have been competing for some resource with the BRP4s winning. Since everything validated it is no a big deal but if anyone knows what was going on I would love to find out.
An interesting observation on BRP times
)
Yes I see the same results. When we first started with BRP5's I and archae86
commented on this. When running combo BRP4 and BRP5 my times for BRP5 were varing from 19,000 to 22,000 now its consistent at 16,000
see http://einsteinathome.org/node/196873&nowrap=true#124659
Seems to affect Nvidia cards more than ATI
Interesting. I guess this
)
Interesting.
I guess this is related to BRP5 (Perseus Arm) work units using shorter FFTs than BRP4 (Arecibo) work units, if I remember correctly. So BRP5 is less computing intensive (in the sense of operations done per byte of data processed). I would not be surprised to see that the effect you see is more or less pronounced for different cards, depending on whether the memory bandwidth or the computing throughput is the bottleneck.
Cheers
HB