Hi all!
We are starting the first experiments on Albert@Home with an OpenCL version of the Fermi Gamma Ray Pulsar (FGRP) search.
As this is our first GPU version of the app for this search, we are expecting some problems (that's the purpose of testing after all).
There are a few known limitations with this early beta test release:
- performance increase will be relatively low. This version is still doing a lot of computation on the CPU
- therefore, the app needs 1 full CPU core plus 1 GPU
- no screensaver / graphics app support yet
- we are releasing the app for ATI/AMD first, later we will try to use the same OpenCL app for NVIDIA as well (no CUDA version is planned for the immediate future if this should work)
- graphics RAM minimum requirements: 512+ MB installed, 300+MB free
- no support for parallel execution of several tasks as per profile settings. We want to test this first with just one task per GPU
- Macs: Lion is minimum OS version (as with BRP4 opencl on Macs)
Cheers
HB
Copyright © 2024 Einstein@Home. All rights reserved.
FGRP application v 1.05 (OPENCL) feedback thread
On the couple of tasks I've seen running this, the speed-up appears to be about 5x (from ~10,000s to about ~2,000s) on a HD6850 - impressive stuff!
Presumably the improvement won't be as great on nvidia (as their OpenCL implementation reputedly isn't the best)?
My understanding is that
My understanding is that their CUDA is that good that openCL is at least sometimes slower on NVIDIA.
Christoph
the wu's finish, they
the wu's finish, they validate against others.
Great job done! It's not 'usual' that the first release is stable.
I'm glad I could return results from 4 different GPU's (HD5000, HD6000, HD7000) and one APU (A8)
I do wait urgently for the
I do wait urgently for the NVIDIA version!
Kind regards and happy crunching
Martin
Hallo! Can become versions
Hallo!
Can become versions for GPU identified by name?
Kind regards
Martin
Oh, sorry. Yes : the 1.05 !!!
Can somebody check my errored
Can somebody check my errored wus and tell me how to proceed?
The other FGRP are on hold.
Two are looking like one BM bug which I reported and which I at least thought was fixed. Continue to run after completion. The CPU time is as short as the other errors. But I suspect BOINC didn't notice it.
Too tired right now to search the logs.
My BM start up info:
16.03.2013 21:26:58 | | Starting BOINC client version 7.0.52 for windows_x86_64
16.03.2013 21:26:58 | | log flags: file_xfer, sched_ops, task, coproc_debug, cpu_sched, cpu_sched_debug
16.03.2013 21:26:58 | | log flags: cpu_sched_status, sched_op_debug, task_debug, work_fetch_debug
16.03.2013 21:26:58 | | Libraries: libcurl/7.25.0 OpenSSL/1.0.1 zlib/1.2.6
16.03.2013 21:26:58 | | Data directory: C:\ProgramData\BOINC
16.03.2013 21:26:58 | | Running under account Administrator
16.03.2013 21:26:58 | | Processor: 8 GenuineIntel Intel(R) Xeon(R) CPU E5540 @ 2.53GHz [Family 6 Model 26 Stepping 5]
16.03.2013 21:26:58 | | Processor: 256.00 KB cache
16.03.2013 21:26:58 | | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss htt tm pni ssse3 cx16 sse4_1 sse4_2 popcnt syscall nx lm vmx tm2 dca pbe
16.03.2013 21:26:58 | | OS: Microsoft Windows Server 2008 "R2": Enterprise x64 Edition, Service Pack 1, (06.01.7601.00)
16.03.2013 21:26:58 | | Memory: 11.99 GB physical, 23.98 GB virtual
16.03.2013 21:26:58 | | Disk: 232.88 GB total, 10.40 GB free
16.03.2013 21:26:58 | | Local time is UTC +1 hours
16.03.2013 21:26:58 | | VirtualBox version: 4.2.8
16.03.2013 21:26:58 | | CAL: ATI GPU 0: ATI Radeon HD 5x00 series (Cedar) (CAL version 1.4.1741, 1024MB, 991MB available, 208 GFLOPS peak)
16.03.2013 21:26:58 | | OpenCL: AMD/ATI GPU 0: ATI Radeon HD 5x00 series (Cedar) (driver version CAL 1.4.1741 (VM), device version OpenCL 1.1 AMD-APP (898.1), 1024MB, 991MB available, 208 GFLOPS peak)
16.03.2013 21:26:58 | | No NVIDIA library found
Christoph
The first thing i found out
The first thing i found out with the FGRP Opencl for Nvidia is that i have 6 cpu cores 2 are normally free for the mainproject (2 BRP4 tasks simoultaneousy)and this FGRP opencl pushed the normally work on side even though that 2 cores are free.i don´t know if you doing this only here on albert or if that´s not correct.
To the runtime the task needed 1700 seconds on my system but the GPU load fluctuation jumped all the time up and down between 35percentage and zero.
At 40% of task completion the gpu did nothing for around 15seconds then it worked
again until to 95% it did again nothing for 87 seconds then it was uploaded and next task begun.I have to wait for validation now.
The FGRP searh works fine
The FGRP searh works fine here, approx same crunch time,as Binary Radio Pulsar Search on my GPU,
But pays only 70 cred ! BRPS gives 500!
Cred payment is to low on this project!!!
Sadly, thats is how it has become. There are a lot of Boinc project,doing their best to get/keep our "donating" time :)
(cred payment is an important part for some participants, when it comes to choose which project to join/crunch )
morgan
RE: The FGRP searh works
Great! That's what we are testing here.
This is a test project that may even be used to test crediting, which might or might not work as you expect.
What else can we do other than to write
on the front page (and AFAIK even in the welcome message in external account managers)?
BM
BM
No I´ve 2 valid results for
No I´ve 2 valid results for FGRP opencl nvidia but one question.
Is the GTX 580 faster with that task(doubled so fast)as my GTX 660 because of higher DP Performance or what´s the difference there?
Yes of course i know different Hardware and OS can affect that too but doubled speed is a little bit hard for that differences i would say.
watch here
http://albertathome.org/workunit/268614