I happened to see Perseus Arm work downloaded on both of my two GTX660 hosts, so arranged for them to run next.
Each host is running 3 Perseus jobs on one card. Even though this uses the same executable as BRP4 jobs do, on both hosts the CPU consumption of the CPU support job is running quite a bit higher than it does for BRP4 (roughly 50% higher).
This could mean that on some hosts the optimum number of pure CPU jobs might differ for BRP5 vs. BRP4 (the practice often but inaccurately called "reserving cores"). One should not necessarily assume the optimum number of simultaneous GPU jobs is the same between the two either.
I have finished 2 of those in parallel on 650Ti:
Run Time = 24400 s
CPU Time = 3000 s
Don't know yet how much credit is granted, sine it's pending validation...
I had a problem prior to the brp4 1.28 CUDA app in that I was unable to achieve anything above 75% GPU load on my 560ti including running multiple GPU tasks & "reserved CPU cores" for GPU . The release of brp4 1.28 fixed those issues and I was able to achieve 90-95% GPU load running 2 GPU tasks.
It seems these new PA work units have returned me to the days of 70-75% GPU load. Probably related to the extra CPU usage as per archae86's observations earlier.
While my first batch of three BRP5 WUs on each of two GTX660 hosts spent most of their run time together, I've subsequently had substantial run time combining one BRP4 with two BRP5 or two BRP4 with one BRP5. While my use of throttling and my use of a single GW pure CPU job complicate direct comparison with other people, an initial observation is that a single BRP4 jobs runs much faster when sharing the GPU with two BRP5 than when sharing with two more BRP4.
Also, the BRP5 jobs run much slower when sharing the GPU with BRP4 work.
These effects will considerably complicate calculations of "fair" credit and of most productive workload, as I imagine a large fraction of the BRP5 production capacity will be on hosts running multiple jobs per GPU, and with both BRP4 and BRP5 enabled.
Thanks archae86 that answered my question why when I woke this morning
I had two BRP5 tasks at 25% done one using 8 hours of runtime and the other at
2 hours. Event log shows I had been running BRP4 tasks in parallel with BRP5.
cpu time is only 1.6 times longer for BRP5 tasks so might be something else.
GTX660 with i5-3570
Now that I have two BRP5's running together the runtimes are tracking very close
about +.005% each second for both
Just checked my HD7950 its running one BRP5 and BRP4 but the runtimes seem to be tracking normally ie 45 min runtime BRP5 is 11% done its increasing +.004%
per second while BRP4 ia increasing .05% per sec seems normal enough
One last update Now running two BRP5's on the HD7950 and the first one
is still increasing the same +.004% per sec as well as the second BRP5 +.004%
and when I dedicated another core so each BRP5 has its own core the increase is now +.005% for each job per sec
Exactly 3 hours runtime on a
)
Exactly 3 hours runtime on a slightly overclocked 560Ti (single GPU task).
Lets see how long it will take on the 760Ti which might be mine in one or two months :-)
I happened to see Perseus Arm
)
I happened to see Perseus Arm work downloaded on both of my two GTX660 hosts, so arranged for them to run next.
Each host is running 3 Perseus jobs on one card. Even though this uses the same executable as BRP4 jobs do, on both hosts the CPU consumption of the CPU support job is running quite a bit higher than it does for BRP4 (roughly 50% higher).
This could mean that on some hosts the optimum number of pure CPU jobs might differ for BRP5 vs. BRP4 (the practice often but inaccurately called "reserving cores"). One should not necessarily assume the optimum number of simultaneous GPU jobs is the same between the two either.
I have finished 2 of those in
)
I have finished 2 of those in parallel on 650Ti:
Run Time = 24400 s
CPU Time = 3000 s
Don't know yet how much credit is granted, sine it's pending validation...
-----
I have finished two BRP5 unit
)
I have finished two BRP5 unit , not in parallel , on my 560 ti :
1 ) GPU : 11531 s ; CPU : 3806 s
2 ) GPU : 11757 s ; CPU : 3986 s
jeanguy
It is a pity the BRP5
)
It is a pity the BRP5 validator is currently disabled, as my quorum partner and I have both turned in work on this work unit.
BRP5 PAS tasks are slightly
)
BRP5 PAS tasks are slightly over 10x longer than BRP4:
- GTX560 (3 WUs in parallel) - BRP4 (3400 s) - BRP5 (39000 s)
- HD7970 (2 WUs in parallel) - BRP4 (1300 s) - BRP5 (13500 s)
I suppose credit granted for one BRP5 task should be floating around 5000 credits.
I had a problem prior to the
)
I had a problem prior to the brp4 1.28 CUDA app in that I was unable to achieve anything above 75% GPU load on my 560ti including running multiple GPU tasks & "reserved CPU cores" for GPU . The release of brp4 1.28 fixed those issues and I was able to achieve 90-95% GPU load running 2 GPU tasks.
It seems these new PA work units have returned me to the days of 70-75% GPU load. Probably related to the extra CPU usage as per archae86's observations earlier.
Original post here
http://einsteinathome.org/node/196484&nowrap=true#118943
While my first batch of three
)
While my first batch of three BRP5 WUs on each of two GTX660 hosts spent most of their run time together, I've subsequently had substantial run time combining one BRP4 with two BRP5 or two BRP4 with one BRP5. While my use of throttling and my use of a single GW pure CPU job complicate direct comparison with other people, an initial observation is that a single BRP4 jobs runs much faster when sharing the GPU with two BRP5 than when sharing with two more BRP4.
Also, the BRP5 jobs run much slower when sharing the GPU with BRP4 work.
These effects will considerably complicate calculations of "fair" credit and of most productive workload, as I imagine a large fraction of the BRP5 production capacity will be on hosts running multiple jobs per GPU, and with both BRP4 and BRP5 enabled.
Thanks archae86 that answered
)
Thanks archae86 that answered my question why when I woke this morning
I had two BRP5 tasks at 25% done one using 8 hours of runtime and the other at
2 hours. Event log shows I had been running BRP4 tasks in parallel with BRP5.
cpu time is only 1.6 times longer for BRP5 tasks so might be something else.
GTX660 with i5-3570
Now that I have two BRP5's running together the runtimes are tracking very close
about +.005% each second for both
Just checked my HD7950 its running one BRP5 and BRP4 but the runtimes seem to be tracking normally ie 45 min runtime BRP5 is 11% done its increasing +.004%
per second while BRP4 ia increasing .05% per sec seems normal enough
One last update Now running two BRP5's on the HD7950 and the first one
is still increasing the same +.004% per sec as well as the second BRP5 +.004%
and when I dedicated another core so each BRP5 has its own core the increase is now +.005% for each job per sec
How many credits for each WU?
)
How many credits for each WU?