I think it's already been mentioned that there will be a new application for future GW searches and a GPU version is also planned. I'd rather wait for that before making hardware decisions.
I ran some tests on a few machines. Using an app_config to limit it to half the cores. The Hi work units take around 9 hours, and that was while running Seti optimised app on the other CPU cores. I also had a machine without any special config and it could run 7 at a time and the 8th work unit sat there with an "Awaiting memory" status. Unfortunately BOINC didn't schedule other work to the core so it was idle.
As Christian said the 2008Mb work units happen in the 1400Mhz (and above) frequency range, I had a few machines pick these up.
The Lo work units on the other hand are fairly quick and don't use much memory so no fiddling is needed for them.
I think it's already been mentioned that there will be a new application for future GW searches and a GPU version is also planned.
I am glad you mentioned that. I was certain that a GPU version had been stated, but couldn't find it. I recall that it was to be a CUDA version at that. But the last mention I found was that the GW searches could only be done on a CPU, so I figured they had given up on it. Maybe someone can state more definitively where they stand on that?
I was certain that a GPU version had been stated, but couldn't find it.
@Bernd Machenschalk's comments here are probably what you were searching for. The usual life cycle for a GPU app is to get a CPU version running and then once there is sufficient data, a GPU app appears in due course.
In any regards it's good to start crunching LIGO data again.
... I recall that it was to be a CUDA version at that. But the last mention I found was that the GW searches could only be done on a CPU, so I figured they had given up on it. Maybe someone can state more definitively where they stand on that?
This is a message that Bernd posted well over a year ago. There have possibly been more recent statements as well. Personally, I'm very confident there will be a GPU version of the GW app. They certainly wont have 'given up'. These things always take time and a lot of effort to get right.
I think you are wrong if you are suggesting it will be CUDA only. I don't recall any such statement. If you want the best information, just search through posts made by Bernd or Christian. My impression is that it could be like what has happened with FGRPB1G. It will probably be an OpenCL app capable of running on both AMD and NVIDIA GPUs. If it doesn't perform well on NVIDIA, a CUDA version could follow at a later date.
As you might have noticed the Windows app was deployed yesterday and is working fine (for a Windows app). The deadline for those tasks is still 5 days as we want to have the results as fast as possible to see if the scientific setup we have chosen is as efficient as we want it to be. We don't necessarily need all of the work and might cut this run short after we gathered enough to make a decision.
It has just been a day now, but I'm seeing invalids between my Windows machine and Linux / MacOs. My one valid work unit is with another Windows machine and the third machine, a MacOS, was marked as invalid. My pending queue also has several inconclusive results.
I've checked my temps, run a couple of quick benchmarks and rebooted last night looking for a hardware fault but everything looks okay. Run times seem consistent between this run and the last gravity wave run.
It has just been a day now, but I'm seeing invalids between my Windows machine and Linux / MacOs. My one valid work unit is with another Windows machine and the third machine, a MacOS, was marked as invalid. My pending queue also has several inconclusive results.
I've checked my temps, run a couple of quick benchmarks and rebooted last night looking for a hardware fault but everything looks okay. Run times seem consistent between this run and the last gravity wave run.
This is the reason for the many inconclusive tasks and also for the backlog of tasks waiting for validation. I deployed an updated Linux version which should hopefully fix that. We'll see how that goes tomorrow.
I think it's already been
)
I think it's already been mentioned that there will be a new application for future GW searches and a GPU version is also planned. I'd rather wait for that before making hardware decisions.
I ran some tests on a few
)
I ran some tests on a few machines. Using an app_config to limit it to half the cores. The Hi work units take around 9 hours, and that was while running Seti optimised app on the other CPU cores. I also had a machine without any special config and it could run 7 at a time and the 8th work unit sat there with an "Awaiting memory" status. Unfortunately BOINC didn't schedule other work to the core so it was idle.
As Christian said the 2008Mb work units happen in the 1400Mhz (and above) frequency range, I had a few machines pick these up.
The Lo work units on the other hand are fairly quick and don't use much memory so no fiddling is needed for them.
BOINC blog
floyd_7 wrote:I think it's
)
I am glad you mentioned that. I was certain that a GPU version had been stated, but couldn't find it. I recall that it was to be a CUDA version at that. But the last mention I found was that the GW searches could only be done on a CPU, so I figured they had given up on it. Maybe someone can state more definitively where they stand on that?
Jim1348 wrote: I was certain
)
@Bernd Machenschalk's comments here are probably what you were searching for. The usual life cycle for a GPU app is to get a CPU version running and then once there is sufficient data, a GPU app appears in due course.
In any regards it's good to start crunching LIGO data again.
Jim1348 wrote:... I recall
)
This is a message that Bernd posted well over a year ago. There have possibly been more recent statements as well. Personally, I'm very confident there will be a GPU version of the GW app. They certainly wont have 'given up'. These things always take time and a lot of effort to get right.
I think you are wrong if you are suggesting it will be CUDA only. I don't recall any such statement. If you want the best information, just search through posts made by Bernd or Christian. My impression is that it could be like what has happened with FGRPB1G. It will probably be an OpenCL app capable of running on both AMD and NVIDIA GPUs. If it doesn't perform well on NVIDIA, a CUDA version could follow at a later date.
Cheers,
Gary.
Thanks for both comments. I
)
Thanks for both comments. I don't mind if it is OpenCl or not, as long as it is efficient.
As you might have noticed the
)
As you might have noticed the Windows app was deployed yesterday and is working fine (for a Windows app). The deadline for those tasks is still 5 days as we want to have the results as fast as possible to see if the scientific setup we have chosen is as efficient as we want it to be. We don't necessarily need all of the work and might cut this run short after we gathered enough to make a decision.
It has just been a day now,
)
It has just been a day now, but I'm seeing invalids between my Windows machine and Linux / MacOs. My one valid work unit is with another Windows machine and the third machine, a MacOS, was marked as invalid. My pending queue also has several inconclusive results.
I've checked my temps, run a couple of quick benchmarks and rebooted last night looking for a hardware fault but everything looks okay. Run times seem consistent between this run and the last gravity wave run.
I have two of them running on
)
I have two of them running on my 2 core Opteron 1210 plus a GPU task running on the GTX 750 Ti. OS is SuSE Linux Leap 42.2.
Tullio
Robert_56 wrote:It has just
)
This is the reason for the many inconclusive tasks and also for the backlog of tasks waiting for validation. I deployed an updated Linux version which should hopefully fix that. We'll see how that goes tomorrow.