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

---

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 their CUDA is that good that openCL is at least sometimes slower on NVIDIA.

---

Christoph

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 NVIDIA version!

Kind regards and happy crunching
Martin

---

Hallo!
Can become versions for GPU identified by name?

Kind regards
Martin

Oh, sorry. Yes : the 1.05 !!!

---

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

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 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

---

The FGRP searh works fine here, approx same crunch time,as Binary Radio Pulsar Search on my GPU,

Great! That's what we are testing here.

Cred payment is to low on this project!!!

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

If you care about credit, this project is certainly not the right one for you.

on the front page (and AFAIK even in the welcome message in external account managers)?

BM

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://albert.phys.uwm.edu/workunit.php?wuid=268614

The FGRP searh works fine here, approx same crunch time,as Binary Radio Pulsar Search on my GPU,

Great! That's what we are testing here.

Cred payment is to low on this project!!!

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

If you care about credit, this project is certainly not the right one for you.

on the front page (and AFAIK even in the welcome message in external account managers)?

BM

Hi again

I am aware of the cred info, even : Most of the time Albert@Home will have no work to run at all, and when it has, the applcations are experimental, might be unstable, unreliable and may even damage your computer..-)

I ment to say; when having Fix cred payment! why so low ?

You can set it to 0, but then Testing out the project will take ages..

Cred are a ´bonus´ for me while helping out project with testing !
I´ve been participated in several Alpha and Beta testing project, so this is not something new to me


PS Bernd, like your haircut :)

---

I've got to say I'm quite impressed with the Nvidia OpenCL app! If I compare the times I get here to the times the CPU-app produce over at Einstein the result is:
The FGRP CPU-app over at Einstein takes about 5700-6000s per task.
The Nvidia OpenCL app here takes about 1300s per task.

I also believe this app has the potential to be much better as the current version use a full CPU core and only loads my GTX660Ti to about 30% and that load comes in bursts. Hopefully this can be improved in future releases.

Keep up the good work!

---

Hi gpu1 0-30% gpu2 0% load. Suspend gpu1 still 0-30% load.It seems that gpu2 working on gpu1.

I am very glad to see that FGRP search has been ported to GPUs. Thanks for developing the GPU applications for FGRP search.

Here are some results I have seen on different hardware and OS.

Tasks running per GPU: 1

OS, Driver, Card, CPU Frequency, GPU Usage, Runtime, PCI-E
Linux, 12.11 Beta 11, 7970, 3930K 4.4 GHz, 55%, 955 sec, x16 3.0
Linux, 295.33, GTX 580, 920 4.2 GHz, NA, 1180 sec, x16 2.0
Linux, 295.33, GTX 580, 920 4.2 GHz, NA, 1280 sec, x8 2.0
XP 64, 314.07, GTX 680, 920 4.2 GHz, 0-30%, 1065-1157 sec, x16 2.0

FGRP search CPU only (via Einstein): 3930K 4.4 GHz No HT, 2850 sec

So far there have been no failed tasks due to the new application. I did have one task error out but that was due to something missing in my Linux image. The FGRP tasks take around 35% longer than the BRP4 tasks to complete but complete much quicker than FGRP search on CPUs alone. I did not have GPU usage via Linux NVIDIA as nvidia-smi does not report usage for Geforce cards.

---

Should Nvidia cards with 512Mb RAM receive work with this app? My ATI host has done quite a few units, but so far I haven't seen any on my Nvidia host (8800GT + 9600GT, each with 512Mb RAM, nvidia driver 295.75). The systems are essentially identical except for the video card.

From the server log for your Nvidia host http://albert.phys.uwm.edu/host_sched_logs/5/5805:

2013-03-27 05:29:44.3316 [PID=32301] [version] Checking plan class 'FGRPopencl-nvidia'
2013-03-27 05:29:44.3316 [PID=32301] [version] NVidia device (or driver) doesn't support OpenCL

In principle, the devices support OpenCL, so I suspect the difficulty lies in the driver.

Fab, thanks! (and I'll know where to look in future).

The drivers are claimed to support OpenCL but maybe a more recent version is needed - I'll update the host and see if that fixes it.

I also believe this app has the potential to be much better as the current version use a full CPU core and only loads my GTX660Ti to about 30% and that load comes in bursts. Hopefully this can be improved in future releases.

Keep up the good work!

I doubt there will much improvement for Nvidia until they come up with better OpenCl driver support.

From the server log for your Nvidia host http://albert.phys.uwm.edu/host_sched_logs/5/5805:

2013-03-27 05:29:44.3316 [PID=32301] [version] Checking plan class 'FGRPopencl-nvidia'
2013-03-27 05:29:44.3316 [PID=32301] [version] NVidia device (or driver) doesn't support OpenCL

In principle, the devices support OpenCL, so I suspect the difficulty lies in the driver.

After upgrading to 310.40 (latest), the server log is now showing this:-

2013-03-28 21:21:34.5079 [PID=8389 ]    [version] Checking plan class 'FGRPopencl-nvidia'
2013-03-28 21:21:34.5079 [PID=8389 ]    [version] OpenCL GPU RAM required min: 536870912.000000, supplied: 536674304
2013-03-28 21:21:34.5079 [PID=8389 ]    [version] [AV#625] app_plan() returned false

If I understand that correctly, the app is requesting RAM of at least 512Mb but the card is reporting only 511.8125Mb available. The app only needs "300+MB Free" (post #1 in this thread) so is there any scope for a bit of tolerance here? :)