Hey Rasputin42,

Are those bc5 tasks? I have noticed that those run more than 3 times slower, so it appears they are stalled, but they are still running.

And GPU temps are really low when running them.

Hey Brent,
They are not running slower, they do not run at all (after the initial about 30%).
Elapsed time is progressing, but the percentage stays exactly the same.

The 720M is simply to slow running multiple instances.
LowPerformancePath is active so use sleep is activated as well.

Try only one instance.

---

With each crime and every kindness we birth our future.

With any "new" GPU installation it is worth running with only one task for a few days just to see what the thing will do in the base situation, then step it up to two for a few more days, finally if that is OK, push up to three. As Mike says I very much doubt that a GTX720M is up to running more than one task at a time.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

It sounds as though system has gone into "trash mode" - it is trying to run too many tasks at a time on the GPU and is failing, so BOINC thinks tasks are getting near their deadline, BOINC pushes them up the priority tree and others get stuck in waiting mode. Trying to run too many tasks ata time will cause the GPU's task scheduling will struggle, particularly with the current crop of low angle tasks. You would probably do better to drop back to one task at a time on the GPU, whereas you are used to running two or three higher angle tasks.

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

Well, the are all "fresh" tasks. That card runs 3 cuda50 tasks with no problem, but i was only running 2 at once. It runs one (SOG task) in about 26min.

As far as i know, if there is enough memory, it will run multiple instances, but less efficient, if you run too many.

How many times must this be said - the critical thing with GPUs is not MEMORY, but the number of GPU "cores" and their management.
There is probably enough memory to support half a dozen tasks, but trying to run more than a couple of tasks (particularly SoG) the GPU's internal task manager will be seriously struggling long before you reach that number.

Another thing to consider is that the current data from the servers is dominated by guppi (from the GBT) for which CUDA is not best suited - my GTX960 rig would quite happily run three "normal" Arecibo MB tasks, but try running three guppi at once and it started to sweat, it is much happier running two of them - that is quite a hit!

---

Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?

My main problem, even though I have the sleep command running, is that I have a gpu wu 'running HP' and a gpu wu in 'waiting to run', it's probably nothing, but I thought I'd mention it.

I have 3 cpu and 3 gpu wu's running, plus some SoG has been downloaded.

I'm also getting this, note the days figure:

SETI@home 8.00 setiathome_v8 (cuda42) 25se10ad.23501.12750.7.34.2_1 00:49:30 (00:00:10) 0.35 0.001 3372d,22:39:51 71.0 °C 0.04C + 0.33NV Running Pegasus

This one is going to go Waiting to run, how can I stop this from happening?

Help...

The days figure is growing...

. . Can you monitor the memory usage on your GPU card? If there is insufficient memory it can exit a task leaving it in the "Waiting to run" state.

. . . Hello Richard,

. . . You haven't mentioned it but is there any chance that SSE4.1 support has been added in 0.45 Beta?

No, it hasn't - no developer has supplied me with any updated CPU applications since the v0.44 launch to support SaH v8.

I need to deploy it on my system with the GT730 which is Core2 Duo based. It might be helpful if SSE4.1 is available to help things along.

Stephen

You don't strictly 'need' it. SIMD hardware support is cumulative - extra capabilities are added to newer CPU designs, but the old ones are never removed. There are one or two gaps where Intel and AMD followed different pathways for a while, but during that phase of development, the incremental steps were relatively small.

Sure, SSSE3 and SSE4.1 would be 'nice to have', but your Core2 Duo will get along pretty well with SSE3 until the developers can catch their breath and regroup.

. . It is doing so as we speak. But I was just wondering.

:)

Hey Rasputin42,

Are those bc5 tasks? I have noticed that those run more than 3 times slower, so it appears they are stalled, but they are still running.

And GPU temps are really low when running them.

. . SoG is running nicely on this machine but blc5 certainly take longer than blc6, they fooled me into thinking something was wrong at one point. I am wondering how blc7 will behave as I have a lot of them coming up.

With any "new" GPU installation it is worth running with only one task for a few days just to see what the thing will do in the base situation, then step it up to two for a few more days, finally if that is OK, push up to three. As Mike says I very much doubt that a GTX720M is up to running more than one task at a time.

. . My GT730 is only barely capable of running doubles, and comes to grief when a Guppie swims by.

How many times must this be said - the critical thing with GPUs is not MEMORY, but the number of GPU "cores" and their management.
There is probably enough memory to support half a dozen tasks, but trying to run more than a couple of tasks (particularly SoG) the GPU's internal task manager will be seriously struggling long before you reach that number.

Another thing to consider is that the current data from the servers is dominated by guppi (from the GBT) for which CUDA is not best suited - my GTX960 rig would quite happily run three "normal" Arecibo MB tasks, but try running three guppi at once and it started to sweat, it is much happier running two of them - that is quite a hit!

. . FWIW, I received the suggestion to run -sbs 512 on my GTX950 to persuade Guppies to play nice. But it wanted more than the 2048MB on the card and it was doing that, dropping out of WUs and leaving them in "waiting to run" state.

. . The max it will support is -sbs 384 so it may be a lack of sufficient memory, worth checking it out.

Wait for something that works better, I get the same thing, only time to completion becomes days, instead of minutes, I tried 3, then 6, then 33, same result, I can't figure this out, I just went back to cuda42, I give up.

Zoom, I would suggest posting your entire commandline rather than just snips of it.

It's hard to tell what the computer see when you only post a portion of it.

-use_sleep -instances_per_device N: 2

I think, that should be:

-use_sleep -instances_per_device 2

-use_sleep -instances_per_device N: 2

I think, that should be:

-use_sleep -instances_per_device 2

Rasputin is correct.

Zoom try this

-use_sleep -sbs 512 -total_GPU_instances_num 2 -instance_per_device 2

Edit..

Only 1 GPU in your machine correct?

You can always decrease the -sbs to 256 if you want and see if that works better

I have noticed, that when running 2 instances of sog, it runs both tasks for a while and then one makes no progress any more.It finishes the other and starts a new one which it continues to process. The first one is still making no progress, but the elapsed time keeps going.If i suspend all other tasks, it finally makes progress again and eventually finishes.
I have used -instances_per_per_device 2.
Any suggestions?

Link to such task result?

I have noticed, that when running 2 instances of sog, it runs both tasks for a while and then one makes no progress any more.It finishes the other and starts a new one which it continues to process. The first one is still making no progress, but the elapsed time keeps going.If i suspend all other tasks, it finally makes progress again and eventually finishes.
I have used -instances_per_per_device 2.
Any suggestions?


Link to such task result?

I will have to set it up again.Currently doing cuda50.

I have noticed, that when running 2 instances of sog, it runs both tasks for a while and then one makes no progress any more.It finishes the other and starts a new one which it continues to process. The first one is still making no progress, but the elapsed time keeps going.If i suspend all other tasks, it finally makes progress again and eventually finishes.
I have used -instances_per_per_device 2.
Any suggestions?

Link to such task result?

Stephen's message 1794739 earlier, about VRAM overcommit, may be relevant.

Link to result would be helpful - we can check if it's a BOINC client which supports temporary exit.

I have noticed, that when running 2 instances of sog, it runs both tasks for a while and then one makes no progress any more.It finishes the other and starts a new one which it continues to process. The first one is still making no progress, but the elapsed time keeps going.If i suspend all other tasks, it finally makes progress again and eventually finishes.
I have used -instances_per_per_device 2.
Any suggestions?

Link to such task result?

Stephen's message 1794739 earlier, about VRAM overcommit, may be relevant.

Link to result would be helpful - we can check if it's a BOINC client which supports temporary exit.

No, it's not.
A habit to post at least link to result under question should be developed.
App prints how much memory it uses for particular task in its stderr.

Regarding SoG task progress: if task is VHAR and processing enters in SoG-only phase when it will enqueue all work to be done for particular task to GPU and then await task completion. What BOINC shows - unrelevant (as we know it shows its own guesses). Cause there is no lack of work for GPU to do even from single task there is nothing bad if runtime decides to finish kernel sequence from one task before switching to another (even if it's real effect that I strongly doubt to be so). There is no such thing as pre-emptive context switching in GPU so far AFAIK after all.