BTW: as the setup of BRP6 is different from that of the previous runs, it's by no means sure that the setting of running N jobs in parallel that might have worked best for those earlier searches still is best for BRP6.
However, we plan to incorporate some optimizations soon(ish) into the BRP GPU apps, and there will be a mix of tasks from different searches for some more time. So for those who want to squeeze the last bit of performance out of the GPUs, I would recommend to rather wait until after those optimizations have been made before again making experiments to get the optimal number of GPU jobs to run in parallel for you host(s).
I got one and it validated. It took apx 32% more time than the BRP5 tasks take. I was running 3 at a time and unlike when I was mixing BRP4 and BRP5s things seemed to run at normal speed. My GTX660 did not like to run the mix of BRP4 and BRP5. It was labeled Binary Radio Pulsar Search (Parkes PMPS XT) v1.39 (BRP5-cuda32-nv301)
Even though it is a BRP6?
Yep, it's known that running BRP4 and BRP5 tasks concurrently causes a significant performance degradation.
Does anyone have experience whether the same happens with BRP5+BRP6? I think it should not, since both use the same binaries (only the filename differs).
Two PMPS XT jobs
)
Two PMPS XT jobs completed.
on a NVIDIA 650 TI run time was around 53k running 4 concurrent gpu tasks
on a NVIDIA 760 run time was around 20k running 3 concurrent gpu tasks
Thanks, looks ok! BTW: as
)
Thanks, looks ok!
BTW: as the setup of BRP6 is different from that of the previous runs, it's by no means sure that the setting of running N jobs in parallel that might have worked best for those earlier searches still is best for BRP6.
However, we plan to incorporate some optimizations soon(ish) into the BRP GPU apps, and there will be a mix of tasks from different searches for some more time. So for those who want to squeeze the last bit of performance out of the GPUs, I would recommend to rather wait until after those optimizations have been made before again making experiments to get the optimal number of GPU jobs to run in parallel for you host(s).
HBE
Here's another one that's
)
Here's another one that's much slower now, also running BRP5-opencl-ati-lion:
http://einsteinathome.org/host/10654444/tasks
Deadline now 3 days. Will
)
Deadline now 3 days. Will that be extended?
What exactly is the correct
)
What exactly is the correct name for this new "Parkes..." for app_config purposes? So I don't have to redo it. Thanks.
merle
What is freedom of expression? Without the freedom to offend, it ceases to exist.
— Salman Rushdie
RE: What exactly is the
)
It's following the naming of the earlier searches and it's called einsteinbinary_BRP6
To obtain the info for yourself enable in cc_config.xml and you'll get this kind of message in Boinc's event log when a new tasks starts:
16/02/2015 22:28:57 | Einstein@Home | [cpu_sched] Starting task PM0001_00121_286_1 using einsteinbinary_BRP6 version 139 (BRP5-cuda32-nv301) in slot 9
You're after whatever is between the words "using" and "version"
RE: RE: What exactly is
)
Works good for me. Thanks Holmis.
merle
What is freedom of expression? Without the freedom to offend, it ceases to exist.
— Salman Rushdie
RE: Deadline now 3 days.
)
We generated another 10 beams (~2000 workunits, ~4000 tasks) for a larger field test.
The final version will have a 14d deadline as the other searches on E@H.
BM
BM
I got one and it validated.
)
I got one and it validated. It took apx 32% more time than the BRP5 tasks take. I was running 3 at a time and unlike when I was mixing BRP4 and BRP5s things seemed to run at normal speed. My GTX660 did not like to run the mix of BRP4 and BRP5. It was labeled
Binary Radio Pulsar Search (Parkes PMPS XT) v1.39 (BRP5-cuda32-nv301)Even though it is a BRP6?
Yep, it's known that running
)
Yep, it's known that running BRP4 and BRP5 tasks concurrently causes a significant performance degradation.
Does anyone have experience whether the same happens with BRP5+BRP6? I think it should not, since both use the same binaries (only the filename differs).
-----