Welcome to MilkyWay@home

GPU app teaser


Advanced search

Message boards : Application Code Discussion : GPU app teaser
Message board moderation

To post messages, you must log in.

Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 19 · Next

AuthorMessage
Honza

Send message
Joined: 28 Aug 07
Posts: 31
Credit: 86,152,236
RAC: 0
50 million credit badge10 year member badge
Message 9519 - Posted: 1 Feb 2009, 13:58:57 UTC - in response to Message 9513.  

@ the guys, where it runs, had you similiar problems?

Well, I had some issues but not app related.
Running under Win 2008 with Catalyst 8.12, automatic regulation of fan speed is not working on my HD3870. It was set to ~20% which cause GPU to overheat under full load. Driver stopped responding, W2K8 was able to recover without crashing but some WUs trashed of course.

Running ATI Tray Tool and setting fan speed to 40-50% helped was good solution.
I'm able to monitor GPU load, temps and fan speed with ATT without problems.
In general, app is working just fine (with comments in Message 9432).


BOINC Project specifications and hardware requirements
ID: 9519 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
bobgoblin

Send message
Joined: 8 Dec 07
Posts: 60
Credit: 67,028,931
RAC: 0
50 million credit badge10 year member badge
Message 9521 - Posted: 1 Feb 2009, 15:41:51 UTC - in response to Message 9518.  
Last modified: 1 Feb 2009, 16:38:26 UTC

I installed it this morning on a vista 64 i7 with ati hd 4800, and they've been cranking out fine since. the short ones are running 5 - 9 seconds; the longer apps are running just shy of 10 minutes each.
ID: 9521 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jedirock
Avatar

Send message
Joined: 8 Nov 08
Posts: 178
Credit: 6,140,854
RAC: 0
5 million credit badge10 year member badge
Message 9524 - Posted: 1 Feb 2009, 17:08:15 UTC - in response to Message 9513.  

Got some errors with the workunits. I made the mistake of setting the app in, then putting Milkyway wide open. About 7 tasks quit with an error before I suspended the rest. Looking in one of the reported WUs, the exit code is 0xc0000135. Googling for that returns many results for BOINC, most of which seem to say it's a missing DLL. I'm presuming this to be brook.dll. So maybe the app_info has to be tweaked so it's also copied to the slots directory? I'm not sure how to verify what files are in there to check, as BOINC deletes them too quickly for me.

Hmm, what have you downloaded?
The brook.dll is supplied in the zip file and also correctly set up in the also supplied app_info.xml. Just copy all 3 files to your Milkyway folder (and completely quit BOINC before). But maybe you need to download new WUs, as for the ones you already have there could be some issues with the data in the client_state.xml.

@ the guys, where it runs, had you similiar problems?

The files are there, and the only thing I've tweaked in the app_info is the avg_ncpus to 1. However, could it be caused by an older version of BOINC? I'm running 6.2.19.
ID: 9524 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9526 - Posted: 1 Feb 2009, 17:21:40 UTC - in response to Message 9510.  

* ~20s GPU crunch times on i920@3.5 + 3850/256.

What just comes to my mind is that the app uses about 20 MB of the video RAM per concurrent WU (actually 10 MB plus 10 MB for every stream, so in case of two streams it will be 30 MB). With an i7 (8 virtual CPUs) and the default entries in the app_info.xml it will run 16 WUs at once if you are only connected to MW.
16 * 20MB = 320MB, what is more than what you have on your card. Have you tested what happens in that case? It would have to be exclusively the short WUs, as the long (multi stream) ones are calculated on the CPU until I extend the app. In the moment I don't test how much video RAM ist still free before I allocate it. There is no error checking for the case something goes wrong. I have not thought of the 256MB variants and 512MB are enough for any sensible amount of concurrent WUs, but it would be interesting to see what happens, just in case.
ID: 9526 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9527 - Posted: 1 Feb 2009, 17:28:44 UTC - in response to Message 9524.  
Last modified: 1 Feb 2009, 17:29:00 UTC

The files are there, and the only thing I've tweaked in the app_info is the avg_ncpus to 1. However, could it be caused by an older version of BOINC? I'm running 6.2.19.

You may try to delete the line with <api_version> in the app_info.xml, but actually it should even run with the old 5.x.x clients. Remember that you have to stop BOINC also for editing the app_info.xml.
ID: 9527 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Honza

Send message
Joined: 28 Aug 07
Posts: 31
Credit: 86,152,236
RAC: 0
50 million credit badge10 year member badge
Message 9528 - Posted: 1 Feb 2009, 17:40:46 UTC - in response to Message 9524.  
Last modified: 1 Feb 2009, 17:40:55 UTC

The files are there, and the only thing I've tweaked in the app_info is the avg_ncpus to 1. However, could it be caused by an older version of BOINC? I'm running 6.2.19.

It runs on 5.10.30 as well.
BOINC Project specifications and hardware requirements
ID: 9528 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jedirock
Avatar

Send message
Joined: 8 Nov 08
Posts: 178
Credit: 6,140,854
RAC: 0
5 million credit badge10 year member badge
Message 9529 - Posted: 1 Feb 2009, 17:41:28 UTC - in response to Message 9527.  
Last modified: 1 Feb 2009, 17:44:44 UTC

The files are there, and the only thing I've tweaked in the app_info is the avg_ncpus to 1. However, could it be caused by an older version of BOINC? I'm running 6.2.19.

You may try to delete the line with <api_version> in the app_info.xml, but actually it should even run with the old 5.x.x clients. Remember that you have to stop BOINC also for editing the app_info.xml.

Yes, I know I have to stop BOINC. I'll try deleting the API version line, and try again.

EDIT: Same result. The hosts tasks are here if you want to look over them.
ID: 9529 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9530 - Posted: 1 Feb 2009, 17:46:46 UTC - in response to Message 9529.  
Last modified: 1 Feb 2009, 17:53:25 UTC

The files are there, and the only thing I've tweaked in the app_info is the avg_ncpus to 1. However, could it be caused by an older version of BOINC? I'm running 6.2.19.

You may try to delete the line with <api_version> in the app_info.xml, but actually it should even run with the old 5.x.x clients. Remember that you have to stop BOINC also for editing the app_info.xml.

Yes, I know I have to stop BOINC. I'll try deleting the API version line, and try again.

I googled a bit and it appears you may need the .NET framework installed. Thought I've unchecked that option in the compiler (it's not used either way), but who knows.

Edit: Hmm, but normally one needs the .NET framework 2.0 either way for the Catalyst driver, isn't it?

Edit2: Has anyone tried the new Cat 9.1? It shouldn't make a difference, as the CAL compiler (part of the driver) still generates the same binary code to run on the GPU.
ID: 9530 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jedirock
Avatar

Send message
Joined: 8 Nov 08
Posts: 178
Credit: 6,140,854
RAC: 0
5 million credit badge10 year member badge
Message 9531 - Posted: 1 Feb 2009, 20:09:55 UTC - in response to Message 9530.  

The files are there, and the only thing I've tweaked in the app_info is the avg_ncpus to 1. However, could it be caused by an older version of BOINC? I'm running 6.2.19.

You may try to delete the line with <api_version> in the app_info.xml, but actually it should even run with the old 5.x.x clients. Remember that you have to stop BOINC also for editing the app_info.xml.

Yes, I know I have to stop BOINC. I'll try deleting the API version line, and try again.

I googled a bit and it appears you may need the .NET framework installed. Thought I've unchecked that option in the compiler (it's not used either way), but who knows.

Edit: Hmm, but normally one needs the .NET framework 2.0 either way for the Catalyst driver, isn't it?

Edit2: Has anyone tried the new Cat 9.1? It shouldn't make a difference, as the CAL compiler (part of the driver) still generates the same binary code to run on the GPU.

I don't know about .NET for Catalyst, but I'm running Windows 7 beta x64. It should have .NET already installed, same version as or later than Vista. As for drivers, no I haven't tried 9.1. 8.12 is all that's available for Win7 for now.
ID: 9531 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Honza

Send message
Joined: 28 Aug 07
Posts: 31
Credit: 86,152,236
RAC: 0
50 million credit badge10 year member badge
Message 9532 - Posted: 1 Feb 2009, 20:28:55 UTC - in response to Message 9531.  
Last modified: 1 Feb 2009, 20:31:00 UTC

I don't know about .NET for Catalyst, but I'm running Windows 7 beta x64. It should have .NET already installed, same version as or later than Vista. As for drivers, no I haven't tried 9.1. 8.12 is all that's available for Win7 for now.

Well, what's good for Vista (and server 2008) should be good for Win 7 (and server 2008 R2). Including drivers like Catalyst.
Win 7 should be shipped with .NET 2 (server 2008 and 2008 R2 comes with it). It has both x86 and x64 .NET 2 version in it.

Anyway, if you installed (full) Catalyst, CCC (Catalyst Control Center) needs .NET. If it's working, means you have .NET 2 working.
BOINC Project specifications and hardware requirements
ID: 9532 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
fubared

Send message
Joined: 8 Apr 08
Posts: 2
Credit: 1,035,343
RAC: 0
1 million credit badge10 year member badge
Message 9538 - Posted: 2 Feb 2009, 3:31:46 UTC - in response to Message 9526.  

...I have not thought of the 256MB variants and 512MB are enough for any sensible amount of concurrent WUs, but it would be interesting to see what happens, just in case.



On my system the WUs "stalls" when more than 6 WU are run concurrently, ie the CPU and GPU usage goes to 0%. Restarting Boinc manager solves the problem for a while then it just stalls again.
ID: 9538 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
James Nunley

Send message
Joined: 29 Nov 07
Posts: 39
Credit: 74,300,629
RAC: 0
50 million credit badge10 year member badge
Message 9541 - Posted: 2 Feb 2009, 4:15:59 UTC

As stated 4670 does not seem to be working this is an MSI 4670 512 MB

Seems to get compute error instantly


on this host

http://milkyway.cs.rpi.edu/milkyway/show_host_detail.php?hostid=25888
ID: 9541 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9542 - Posted: 2 Feb 2009, 4:23:50 UTC - in response to Message 9538.  

...I have not thought of the 256MB variants and 512MB are enough for any sensible amount of concurrent WUs, but it would be interesting to see what happens, just in case.


On my system the WUs "stalls" when more than 6 WU are run concurrently, ie the CPU and GPU usage goes to 0%. Restarting Boinc manager solves the problem for a while then it just stalls again.

It appears it needs double the graphics memory I thought, about 40MB per WU. That would explain your findings.
ID: 9542 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9544 - Posted: 2 Feb 2009, 4:31:04 UTC - in response to Message 9541.  
Last modified: 2 Feb 2009, 4:31:56 UTC

As stated 4670 does not seem to be working this is an MSI 4670 512 MB

Seems to get compute error instantly


on this host

http://milkyway.cs.rpi.edu/milkyway/show_host_detail.php?hostid=25888

There are actually two reasons why it is not working. A HD4670 does not support double precision (only HD38x0 and HD48x0) and the app is a 64Bit application. So one needs a WinXP64, Vista64, Server 2003 x64, Server 2008 or a Windows7 installation (maybe I forgot something), but it is not running on WinXP Pro SP3.
ID: 9544 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
ProfileGlenn Rogers
Avatar

Send message
Joined: 4 Jul 08
Posts: 165
Credit: 364,966
RAC: 0
100 thousand credit badge10 year member badge
Message 9551 - Posted: 2 Feb 2009, 8:55:36 UTC - in response to Message 9518.  

Thank you kindly for the info and the link... Lets hope it works out maybe someone wil do an app for w32 and xseries ATI cards..
Glenn.
ID: 9551 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
jedirock
Avatar

Send message
Joined: 8 Nov 08
Posts: 178
Credit: 6,140,854
RAC: 0
5 million credit badge10 year member badge
Message 9649 - Posted: 4 Feb 2009, 0:41:47 UTC

Any more suggestions on what I should do to test the GPU app? I have a Vista x64 installation I can try with to, if that might help.
ID: 9649 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Realkojack

Send message
Joined: 8 Feb 09
Posts: 1
Credit: 668,863
RAC: 7,748
500 thousand credit badge10 year member badge
Message 9682 - Posted: 4 Feb 2009, 22:41:19 UTC

Works perfect here on a Q6600 with HD4870 running on Vista X64. Should get the label "release candidate" imho.
ID: 9682 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9688 - Posted: 4 Feb 2009, 23:27:33 UTC - in response to Message 9682.  

Works perfect here on a Q6600 with HD4870 running on Vista X64. Should get the label "release candidate" imho.

As long the app does not handle WUs with more than one stream (still fighting some bugs of the Stream SDK to get it running) and there is no Linux variant of it, I would call it more a technology demonstrator than a release candidate.
Two further things still missing are the multi GPU Support (a single HD4870 is already that fast, maybe I should restrict it to one GPU ;) and the release of the CPU during the GPU computations.

But I think one could make a public beta, if the worst things are ironed out.
ID: 9688 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9746 - Posted: 7 Feb 2009, 1:00:33 UTC
Last modified: 7 Feb 2009, 1:04:47 UTC

I have the next alpha version ready, still limited to Win64 and running on ATI HD38x0 and HD48x0 cards with Catalyst driver 8.12 or 9.1. But this time it should be able to handle again all WUs flying around, also those with more than one stream. They should take about 50% longer (there is one stream more to calculate). I have squeezed the code a bit, it should need now about 0.1 seconds less for the single stream WUs ;) But there is not much left what one could get out of it performance wise. Don't expect major speed gains for the future!

Let me know how it is running for you!

Link to the latest alpha-version

Next thing to do will be more diagnostic output (GPU detection) as preparation for a major overhaul of the WU scheduling, which will hopefully support multiple GPUs afterwards and maybe also the release of the CPU during the GPU tasks. After that, I will have a look to Linux.
ID: 9746 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
50 million credit badge10 year member badgeextraordinary contributions badge
Message 9827 - Posted: 7 Feb 2009, 4:27:37 UTC - in response to Message 9519.  

Running ATI Tray Tool and setting fan speed to 40-50% helped was good solution.
I'm able to monitor GPU load, temps and fan speed with ATT without problems.
In general, app is working just fine (with comments in Message 9432).

Just a short note to the GPU-Z problem also seen in the first video in this thread. It appears that not only GPU-Z but also the Catalyst control panel (if started) leads to severe lags. In case of GPU-Z it is caused by the monitoring (load, temperature, fan speed) part, the same could also apply to the CCP.

But according to Honzas comment the monitoring works flawlessly with the ATI tray tools, so maybe there is some strange thing going on here.
ID: 9827 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Previous · 1 · 2 · 3 · 4 · 5 · 6 . . . 19 · Next

Message boards : Application Code Discussion : GPU app teaser

©2019 Astroinformatics Group