Fermi LAT Gamma-ray pulsar search #3 "FGRP3"
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4307
Credit: 249793180
RAC: 32953
14 Jan 2014 13:11:54 UTC
Topic 197343
(moderation:
We will soon be starting the new "Fermi LAT Gamma-ray pulsar search #3" that supersedes the previous "#2" / "FGRP2" search.
Some features of the new application were already posted elsewhere, but I'll collect these here again.
- improved semicoherent stage with increased sensitivity. In addition, we use larger FFTs, which means longer run-times per sky point, but we'll adjust the number of sky points per workunit such that the overall run-time per task remains roughly the same.
- improved coherent (follow-up) stage: There will only be a single, though deeper, follow-up done at the end of each task. Previously, we had a less sensitive follow-up after every 11 sky points. However, the run-time of the new follow-up depends on the outcome of the previous semicoherent stage and is not predictable - it may vary between 30s and 30m. There is currently no checkpointing during that stage.
- sky positions are calculated on-the-fly, no more (large) skygrid files to handle
- the application was originally meant for processing the new, more sensitive Fermi LAT data that is available since Nov 2013. However, pre-processing this data will take a bit longer as expected. For a start we will feed the new application with the (converted) remaining data sets of FGRP2 (LATeah0005x).
- OpenCL versions for AMD, NVidia and Intel GPUs. As we did with the first BRP GPU apps, so far only the FFT will run on the GPU, the rest of the computation (which is still a lot) is done on the CPU. As the application develops, more and more computation will be shifted to the GPU. Note that in the project preferences setting there is (now) a "GPU utilization factor for FGRP" separate from the BRP one.
We are still busy preparing the input data for the new app, I expect the first tasks to be sent tomorrow.
BM
BM
Fermi LAT Gamma-ray pulsar search #3 "FGRP3"
)
Thanks for the detailed information !
-----
Thanks for the update. (No
)
Thanks for the update. (No sign of them yet...)
RE: - OpenCL versions for
)
For the 'Run only the selected applications' preference, are you planning for separate check boxes for 'Gamma ray pulsar search #3' and 'Gamma ray pulsar search (GPU) #3 - like there is for 'BRPS (Arecibo)' and 'BRPS (Arecibo, GPU)? I hope you will be but I thought I'd better ask ;-). If not, I wonder how it would be possible to prevent FGRP3 GPU tasks for a host where the desired mix is BRP5 on the GPU and FGRP3 on the CPUs?
Also, will the 'Run CPU versions of applications for which GPU versions are available' preference affect both BRP4 and FGRP3? In other words, would you need to use venues if you wanted to have different behaviour for each application?
Cheers,
Gary.
RE: Thanks for the update.
)
Take a look at the server status page. There is a trial batch of 100 WUs (200 tasks) already in progress.
Unfortunately, the initial signs seem to be a bit of a worry. At the time of writing, there are already 18 failed tasks (which have already been replaced by the scheduler - it's now 218 tasks total) and only 10 so far with 'pending' status. I hope it's just due to 18 wayward hosts that somehow managed to stuff up the transition to the new app.
If anyone has an 'in progress' FGRP3 task, can you please advise (apart from the app itself) what files were in the initial download? Is the new app still using JPLEPH.405 or perhaps something different?
Cheers,
Gary.
RE: For the 'Run only the
)
No. Technically BRP4 ("Arecibo") and BRP4G ("Arecibo GPU") are different applications, with different workunit sizes etc. There is no other mechanism for this than the usual app selection. In particular this means that there is no "cross-validation" between Arecibo CPU and GPU tasks.
OTOH at current state the FGRP3 GPU App is still somewhat experimental and we don't really trust its results until at least reasonable percentage is checked against the CPU results (which we trust). With time (and possibly a couple of bug fixes) we should get confidence of the GPU App, and then may split it from the CPU app (and design separate WUs). But not now.
Currently FGRP is not at all affected by the 'Run CPU versions of applications for which GPU versions are available' setting, which is a bug we'll hope to fix in the next app version release.
For now, you may adjust to your needs using different venues, and even this may not cover every possible case. We certainly need to think about this a little more when we find the time.
BM
BM
RE: If anyone has an 'in
)
We are still using the same ("sticky") ephemeris file "JPLEPH.405" that should be present on every machine that ever ran a FGRP task. Apart from that only the application files and the actual data file that is related to the set of WUs (currently only "LATeah0052C.dat") need to be downloaded.
BM
BM
RE: At the time of writing,
)
Not much to worry, actually. Think about it: error tasks finish earlier, thus are also the first to be reported. In the current case 16 of the 18 tasks are from download errors of only a few machines that probably have a flawed internet connection. Not that unusual.
BM
BM
Apparently there are some
)
Apparently there are some occasional issues with the OpenCL versions. We will soon issue another app version, until then we disabled the GPU versions and generated some more FGRP3 work.
BM
BM
Thanks for the update Berndt.
)
Thanks for the update Berndt. To have some clarity. Will we see FGRP 3 in an GPU and a GPU version when it is released? Tried to read from previous posts but i haven't figured that out yet. Sorry if i missed the info if it was posted before.
RE: Thanks for the update
)
FGRP3 is OpenCL-capable, but there are some issues currently - see above.
-----