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

same thing for me. HD7870s in crossfire
13.3 BETA 3

I have had 1617 tasks validated and 207 tasks pending in the last week or so via one of my Linux systems with a NVIDIA card. There was one task that failed to validate yesterday but otherwise, the new OpenCL application appears to be running stable overall with Linux/NVIDIA.

---

I'm experiencing a some very strange bug. I've two video cards and they are working very well in Einstein@Home project. However when I try to run FGPR Opencl WUs on them I can see very strange behavior - two WU's are running on the second card simultaneously and nothing is happening on the first card. Nevertheless boinc manage shows that WUs are being processed each one on the separate cards.

I'm experiencing a some very strange bug. I've two video cards and they are working very well in Einstein@Home project. However when I try to run FGPR Opencl WUs on them I can see very strange behavior - two WU's are running on the second card simultaneously and nothing is happening on the first card. Nevertheless boinc manage shows that WUs are being processed each one on the separate cards.

BOINC are having problems with their code (server, client, API - not sure which is most relevant) in that area - identifying which card is which 'device' in OpenCL terminology. Oliver might want to look at that report in connection with the current BOINC developer discussions (mainly being initiated by Charle Fenton).

The application produces with 1 cpu core and 1 gpu a little bit more computation output for FGRP as if i would run it on all of my 6 cores on the fx6100@3600Mhz
and that on a gpu fluctuation on my gtx 660 OC @1150Mhz at around 30%.
It looks like i could run 2 or 3 simultaneously if i would.
Yes i know actually for testing it´s not allowed.

Specifically 2 cpu cores and 0,5 gpu or 3 cpu cores 0,33 gpu.
But with 2GB of gpu memory it´s certainly possible?

I'm experiencing a some very strange bug. I've two video cards and they are working very well in Einstein@Home project. However when I try to run FGPR Opencl WUs on them I can see very strange behavior - two WU's are running on the second card simultaneously and nothing is happening on the first card. Nevertheless boinc manage shows that WUs are being processed each one on the separate cards.

BOINC are having problems with their code (server, client, API - not sure which is most relevant) in that area - identifying which card is which 'device' in OpenCL terminology. Oliver might want to look at that report in connection with the current BOINC developer discussions (mainly being initiated by Charle Fenton).

This is either a problem in the client or a problem communicating the device to run on from the client to the application.

AFAIK we're using the exact same code in the FGRP App than in the OpenCL BRP App, and I know of no such error being reported with the latter. My guess is that it's either a bug in the OpenCL implementation in the driver (in which case using a different driver version might help) or in the client (try a different version here, too).

When two such tasks are running, could you take a look at the command lines? There should be a small integer number following a "--device" option. If this number is identical for the two tasks, then this is definitely a problem in the Client. If it is different, it may be a problem in the App code or the driver.

BM

I think it's a app or driver problem.

Testing with host 5367: two identical GTX 670, BOINC v7.0.64, driver v314.22, Windows 7

Event log says:

19/04/2013 17:19:11 | Albert@Home | [coproc] NVIDIA instance 1: confirming for LATeah0069U_16.0_40_-8.24e-10_1
19/04/2013 17:19:11 | Albert@Home | [coproc] NVIDIA instance 0: confirming for LATeah0069U_16.0_40_-8.31e-10_0

Command line in slot 0 ends --device 1
Command line in slot 1 ends --device 0

But GPU-Z shows



MSI Afterburner shows significant usage on what it calls GPU 1, zero usage on GPU 2.

Thanks, that's helpful.

That's as far as I could track it down, Oliver or HB should be able to pick that up after the weekend.

BM

When two such tasks are running, could you take a look at the command lines? There should be a small integer number following a "--device" option. If this number is identical for the two tasks, then this is definitely a problem in the Client. If it is different, it may be a problem in the App code or the driver.

BM

Sorry, wasn't very informative on my previous post.

The command lines look fine:
--device 0
--device 1
Boinc 7.0.28 and GTX660Ti as fisrt card and GTX560Ti as second.
NVIDIA driver 314.07

Thank you.

This is just to let you know that we are working on it. It definitely is more difficult than I first thought.

It might help if someone of you having this problem could send us (or post here, stripped of any personal info you like) the init_data.xml files from the two slots of the tasks that end up running on the same device.

BM

It might help if someone of you having this problem could send us (or post here, stripped of any personal info you like) the init_data.xml files from the two slots of the tasks that end up running on the same device.

BM

Which of these apps need to be enabled?

Binary Radio Pulsar Search: no
Binary Radio Pulsar Search (single DM): no
Binary Radio Pulsar Search (Perseus Arm Survey): yes
Gravitational Wave S6 LineVeto search (extended): no
Gamma-ray pulsar search #2: yes

---

This is just to let you know that we are working on it. It definitely is more difficult than I first thought.

It might help if someone of you having this problem could send us (or post here, stripped of any personal info you like) the init_data.xml files from the two slots of the tasks that end up running on the same device.

BM

I'll set up a run and send you the files (email OK?) when this GPUGrid task vacates its GPU - ~2.5 hours.

This is just to let you know that we are working on it. It definitely is more difficult than I first thought.

It might help if someone of you having this problem could send us (or post here, stripped of any personal info you like) the init_data.xml files from the two slots of the tasks that end up running on the same device.

BM

I'll set up a run and send you the files (email OK?) when this GPUGrid task vacates its GPU - ~2.5 hours.

Email is ok. But it will probably be enough to post the tags of the two files here ( and whatever I might have missed).

BM

This is just to let you know that we are working on it. It definitely is more difficult than I first thought.

It might help if someone of you having this problem could send us (or post here, stripped of any personal info you like) the init_data.xml files from the two slots of the tasks that end up running on the same device.

BM

I'll set up a run and send you the files (email OK?) when this GPUGrid task vacates its GPU - ~2.5 hours.

Email is ok. But it will probably be enough to post the <gpu...> tags of the two files here (<gpu_type> <gpu_device_num> <gpu_opencl_dev_index> and whatever I might have missed).

BM

Well, those tags look as I would have expected:

<gpu_type>NVIDIA</gpu_type>
<gpu_device_num>0</gpu_device_num>
<gpu_opencl_dev_index>0</gpu_opencl_dev_index>

<gpu_type>NVIDIA</gpu_type>
<gpu_device_num>1</gpu_device_num>
<gpu_opencl_dev_index>1</gpu_opencl_dev_index>

I was a little worried about a couple of 'device_id 0' lower down:

<pci_info>
   <bus_id>7</bus_id>
   <device_id>0</device_id>
   <domain_id>0</domain_id>
</pci_info>
<pci_info>
   <bus_id>8</bus_id>
   <device_id>0</device_id>
   <domain_id>0</domain_id>
</pci_info>

I'll zip up the files and email them so you have everything available in the same place.

Email is ok. But it will probably be enough to post the <gpu...> tags of the two files here (<gpu_type> <gpu_device_num> <gpu_opencl_dev_index> and whatever I might have missed).

BM

Same here:

<gpu_type>NVIDIA</gpu_type>
<gpu_device_num>0</gpu_device_num>
<gpu_opencl_dev_index>0</gpu_opencl_dev_index>

and

<gpu_type>NVIDIA</gpu_type>
<gpu_device_num>1</gpu_device_num>
<gpu_opencl_dev_index>1</gpu_opencl_dev_index>

Email is ok. But it will probably be enough to post the <gpu...> tags of the two files here (<gpu_type> <gpu_device_num> <gpu_opencl_dev_index> and whatever I might have missed).

BM

Same here:

<gpu_type>NVIDIA</gpu_type>
<gpu_device_num>0</gpu_device_num>
<gpu_opencl_dev_index>0</gpu_opencl_dev_index>

and

<gpu_type>NVIDIA</gpu_type>
<gpu_device_num>1</gpu_device_num>


<gpu_opencl_dev_index>1</gpu_opencl_dev_index>

Is it possible to add to the WU output something like:
[INFO ] Using CUDA device #0 "GeForce GTX 660 Ti" (0 CUDA cores / 0.00 GFLOPS)
as it was done for CUDA?

This is not a problem or bug but there is no other thread possibility then news.

The FGRP (gamma ray) via ATI are finished in around 1350 seconds on a old HD5870 card.
On cpu via Einstein@home (also gamma ray, I thus I think the same)it takes my pc around 8200 seconds to finish. So a good improvement.

---

Greetings from,
TJ.

This is not a problem or bug but there is no other thread possibility then news.

The FGRP (gamma ray) via ATI are finished in around 1350 seconds on a old HD5870 card.
On cpu via Einstein@home (also gamma ray, I thus I think the same)it takes my pc around 8200 seconds to finish. So a good improvement.

Thanks for the feedback. Yes, the good old HD 5850 and 5870 are quite fast, if only they would consume less power :-[ .

I'm quite satisfied with the performance for the moment, but there still is some work to do under the hood (fixing a bug of mine in GPU detection, reducing verbosity of log, error handling needs to be more robust ....). But I think I can put out a new fixed version tomorrow latest.

Cheers
HB

---

Will the new version work on cards with 512Mb, do you know? (ref. Message 112402).

Will the new version work on cards with 512Mb, do you know? (ref. Message 112402).

Memory requirements will be unchanged, sorry.
HB

---