Hey,
I just came back to crunching seeing that CUDA is available to my graphics cards. Started all that stuff, but it seems I only get tasks for CPU.
Boinc sends requests for GPU work but never gets any tasks, whereas my CPU tasks flood the computer...
Didn't find anything on GPU issues, so what's that about? Why don't I get any work?

---

Hey,
I just came back to crunching seeing that CUDA is available to my graphics cards. Started all that stuff, but it seems I only get tasks for CPU.
Boinc sends requests for GPU work but never gets any tasks, whereas my CPU tasks flood the computer...
Didn't find anything on GPU issues, so what's that about? Why don't I get any work?

Because Rosetta does not do gpu processing right now. Only a few Boinc projects do...Seti and Milky Way come to mind right away. Milky Way is very iffy as to when and if you can get work, some have no trouble, some can't get any. If you want to crunch outside of Boinc, at Folding@Home you can use both Nvidia and ATI cards.

---

Just confirming mikey's comments.

Rosetta@home is coded to run on CPU, not GPU. So, your machine is getting all of the work that the project is able to use it for. There have been some issues with the BOINC core client requesting too much work in the newer versions that support CUDA. So this may be why you feel like your machine is flooded with CPU work. Both because that is all you can get from R@h and because the client requested more then it really needed.

The BOINC folks are working on revisions to the work fetch coding in the core client to avoid those problems in the future.

---

Rosetta Moderator: Mod.Sense

Hey,
I just came back to crunching seeing that CUDA is available to my graphics cards. Started all that stuff, but it seems I only get tasks for CPU.
Boinc sends requests for GPU work but never gets any tasks, whereas my CPU tasks flood the computer...
Didn't find anything on GPU issues, so what's that about? Why don't I get any work?

Because Rosetta does not do gpu processing right now. Only a few Boinc projects do...Seti and Milky Way come to mind right away. Milky Way is very iffy as to when and if you can get work, some have no trouble, some can't get any. If you want to crunch outside of Boinc, at Folding@Home you can use both Nvidia and ATI cards.

Seti@home is also having work problems at the moment because Astropulse which is a subproject is not working, causing huge strain on the other work units. I am able to get some but not enough to fill my cache. GPUGrid seems to be the most reliable project for giving GPU work, however, unless you have a reasonably fast nVidia card (64 cores or above) you will not meet the WU deadlines.

Has there been any talk of even bringing GPU work to Rosetta? I would think bringing the power of GPU folding would benefit this project greatly if done right.

Has there been any talk of even bringing GPU work to Rosetta? I would think bringing the power of GPU folding would benefit this project greatly if done right.

So much talk that it is a dead horse that no one wants resurrected.
Is this Boinc client requesting GPU work a strain on the Rosetta servers? If so I will go back to version 6.4.5 and get rid of 6.6.63

---

Has there been any talk of even bringing GPU work to Rosetta? I would think bringing the power of GPU folding would benefit this project greatly if done right.

So much talk that it is a dead horse that no one wants resurrected.
Is this Boinc client requesting GPU work a strain on the Rosetta servers? If so I will go back to version 6.4.5 and get rid of 6.6.63

I just checked through all my messages on one of my CUDA ready machines which isn't actually CUDA crunching anything currently. I see no GPU requests from this machine (using 6.6.36).

Will 6.6.36 BOINC manager enable a CUDA enabled device to work on Rosetta @Home.
I hope so. Because the head of BOINC came out in a statement a while back (6 months ago) that to reach an exoflop would require GPU type devices. I dont think Rosseta will ever get enough computing power untill an exoflop is reached.

---

Will 6.6.36 BOINC manager enable a CUDA enabled device to work on Rosetta @Home.
I hope so. Because the head of BOINC came out in a statement a while back (6 months ago) that to reach an exoflop would require GPU type devices. I dont think Rosseta will ever get enough computing power untill an exoflop is reached.

No Rosetta@home does not have a GPU application, nor is it going to have one in the near future. There is a huge thread on the topic on this forum, have a search for it. It will give you all the details for the discussion.

Will 6.6.36 BOINC manager enable a CUDA enabled device to work on Rosetta @Home.
I hope so. Because the head of BOINC came out in a statement a while back (6 months ago) that to reach an exoflop would require GPU type devices. I dont think Rosseta will ever get enough computing power untill an exoflop is reached.

No Rosetta@home does not have a GPU application, nor is it going to have one in the near future. There is a huge thread on the topic on this forum, have a search for it. It will give you all the details for the discussion.

Turning an application into a GPU application takes a lot of programming work, especially for memory-hungry applications such as minirosetta. Few BOINC projects have invested enough programming time to do this.

For now, adding GPUGRID seems the best option for making use of Nvidia cards at all:

http://www.gpugrid.net/

Expect it to require a version of BOINC more recent than most other BOINC projects recommend; version 6.6.28 the last I looked. At least some versions of BOINC more recent than 6.6.28 are known to have problems using GPUs correctly; read their forums if you need details. Also, it helps to give GPUGRID a share of the CPU time based on counting the GPU as another CPU core, and assuming that it will make full use of the GPU but little use of the other CPU cores. Also, expect a generous return of BOINC credits if it is able to use your Nvidia card - it can't use all models.

However, note the option for turning off BOINC access to the GPU even if you have one:

http://boinc.berkeley.edu/wiki/Client_configuration

The most recent versions of BOINC default to having it turned off, so you're likely to need to change this setting.

There are plans to eventually allow BOINC to access ATI graphics cards, but it appears that the standard method (and therefore the ability to share the ATI card between BOINC projects) is months away from being ready to use (probably when the 6.8.* versions of BOINC are ready to use). I expect these cards to also require a heavy reprogramming effort.

My client still requests GPU work often, even though I have the setting for it unchecked in the preference, and I am using the online preferences anyway.

---

My client still requests GPU work often, even though I have the setting for it unchecked in the preference, and I am using the online preferences anyway.

This is a work fetch issue where the design decision was to allow the client to ask all projects for all classes of work. So, BOINC will ask all projects for CPU work and GPU work with out discrimination. Over time the exponential back-off is supposed to reduce the calls for work of the "wrong" classes, but, that takes some time.

More sadly, there are still bugs in this area of the system and so you actually still get more calls for GPU work on CPU only projects than should be expected. There are now about 6 projects that are working on GPU tasks and applications so this should start to be less of an issue, though most projects have problems of one sort or another with the GPU side (GPU Grid being a notable exception for the most part their tasks just run now ... YMMV)...

Oh, and if you are attached to GPU Grid, there BOINC asks for CPU work and never gets any because they don't have any ... :)

... There are now about 6 projects that are working on GPU tasks and ...

Which are those six? I can only come up with three and a half: Seti, gpugrid, AQUA (they also do multicore), milkyway (this is the half. ATI works, but CUDA isn't up an running yet.)

---

I know that Primegrid where trying to create an ATI application however ran into problems and dropped the idea for now. Anyone else got clarification of all projects which have or are developing GPU apps??

On a side note: Folding@Home which is non BOINC based project supports both ATI and Nvidia GPUs.

... There are now about 6 projects that are working on GPU tasks and ...

Which are those six? I can only come up with three and a half: Seti, gpugrid, AQUA (they also do multicore), milkyway (this is the half. ATI works, but CUDA isn't up an running yet.)

There are several projects in various stages of GPU development, but none what you would call 'close'. All say they are 'working on it' but will give no commitment. A couple of projects have looked at it and for one reason or another aren't going to GPU processing. One project says it can't produce enough work to keep the GPU's busy, most say the GPU's don't supply enough precision for what they currently do. They say they will 'look again in the future if things change' but don't get your hopes up! Going to GPU processing has turned Milky Way into a project that either has work or doesn't have work. The GPU's blast thru the work faster than the CPU's and the project can't make work fast enough, so they often run out!! Of course all that crunching then needs to be analyzed, so more work for the project at a faster rate! And since alot of the projects use a ton of volunteers, they are kinda between a rock and a hard place.

---

Going to GPU processing has turned Milky Way into a project that either has work or doesn't have work. The GPU's blast thru the work faster than the CPU's and the project can't make work fast enough, so they often run out!! Of course all that crunching then needs to be analyzed, so more work for the project at a faster rate! And since alot of the projects use a ton of volunteers, they are kinda between a rock and a hard place.

Means there's more CPUs going spare for the other projects :)

---

Going to GPU processing has turned Milky Way into a project that either has work or doesn't have work. The GPU's blast thru the work faster than the CPU's and the project can't make work fast enough, so they often run out!! Of course all that crunching then needs to be analyzed, so more work for the project at a faster rate! And since alot of the projects use a ton of volunteers, they are kinda between a rock and a hard place.

Means there's more CPUs going spare for the other projects :)

Yes it does but that is not what the Users that like the idea behind Milky Way want to see. It is a balancing thing, blast thru the units or try to even them out. Personally I would like to see more projects do a dual kind of workunit thing, one kind of unit for cpu's and a different kind of gpu's. That way both sets of people are happy and there is fewer problems for the project, they just have 2 kinds of work to make. You can tailor the workuntis to what gets it, the cpu or the gpu. Although I guess that can be a problem in itself!

---

... There are now about 6 projects that are working on GPU tasks and ...

Which are those six? I can only come up with three and a half: Seti, gpugrid, AQUA (they also do multicore), milkyway (this is the half. ATI works, but CUDA isn't up an running yet.)

... A couple of projects have looked at it and for one reason or another aren't going to GPU processing. One project says it can't produce enough work to keep the GPU's busy, most say the GPU's don't supply enough precision for what they currently do. They say they will 'look again in the future if things change' but don't get your hopes up! Going to GPU processing has turned Milky Way into a project that either has work or doesn't have work. The GPU's blast thru the work faster than the CPU's and the project can't make work fast enough, so they often run out!! Of course all that crunching then needs to be analyzed, so more work for the project at a faster rate! And since alot of the projects use a ton of volunteers, they are kinda between a rock and a hard place.

science doesn't translate well to a massively parallel algorithm is a perfectly fine argument, but "too fast" isn't.

Everyone in the BOINC community needs to be a good citizen. And one way to be a good citizen is to responsibly use the donated resources by supplying the most efficient apps for your project. This means GPU and SSEx when applicable. If you run out of work units either you could expand your project or free up the compute resources for another project. The donated BOINC compute resource is "free" to a project just like donations to the Red Cross are "free". They still need to be used responsibly because if the Red Cross doesn't need them right now there are plenty of other causes which do need those resources.

Milkyway has been having WU issues for a couple of months, but this is a mark of progress. They first massively optimized their program and used SSEx optimizations to gain something around an order of magnitude of improvement, then because of their openness a user developed an ATI GPU program which gave something like 2 orders of magnitude improvement, now they are working on a CUDA version. Their ability to do science has dramatically improved and when they don't have a WU then I crunch on another project. This is much better than inefficiently slogging through 16 hr work units which we now know can be processed in minutes or seconds, that time is now free to work on someone else's project or do much more work for MW.

This is not an indictment against any project, just a general request to the BOINC admin community to use the donated resource wisely.

---

This is not an indictment against any project, just a general request to the BOINC admin community to use the donated resource wisely.

I think eventually this will happen, but unfortunately this is not a sprint but a marathon. What I would like to see is Boinc auto installed on every computer sold in the WHOLE WORLD, kinda like all the junk that comes on all the new ones now. Then as the User sets up their pc they pick a project, or 2 or 3, and start helping! Distributed Computing has gone from the early days of one project with its own specialized software to Boinc, written by the same people, to over a dozen projects now available for us to choose from. More even than on the list of projects on the list built into Boinc!! One day Distributed Computing will be run off the tiniest of computers taking only minutes instead of hours to crunch a unit and the contributions to Science will be ASTOUNDING! BUT we are not there yet, first we crawl then we walk, then we run, Boinc is still crawling, it doesn't even handle multicore processors and time scheduling of its own projects very well yet!! And gpu's can bring cpu processing to a crawl, and vice versa, with Boinc's sorry use of processor scheduling!

---