Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
1)
Message boards :
Number crunching :
The Dog Days of Summer Challenge
(Message 39527)
Posted 4301 days ago by Snf*
Most short wu challenges have seen these kind of problems. This one might be slightly worse but it's not that big a difference. The main problem is the short wus. This combined with the low wu limits makes things worse.
If you are going to continue to run challenges with short tasks, I suggest you make the wus longer by combining several tasks in one wu. I know it's some work, but so is babysitting an overloaded server.
When it comes to running out of work, we are all in the same boat. While there may be some randomness, most of us have problems getting work from the overloaded server, so we all run dry some periods of time.
It won't be much of a difference rescheduling the challenge unless you address the problems with the large amount of short wus. Sure, you might get fewer participants in a rerun (and survive that way), but then you have kind of missed the point with a challenge.
/Snf
|
2)
Message boards :
Number crunching :
sli/crossfire
(Message 34285)
Posted 4462 days ago by Snf*
Nowadays (at least for a year), you can choose to have SLI/CF enabled and BOINC will still see multiple separate GPUs. So if you have 2 GPUs, BOINC will run one task on each of them.
I run all of my setups this way, since it eliminates the sometimes needed dummy plug in a card not connected to a monitor.
DNETC can run one job on several GPUs, but in reality it only starts 2 threads in the same job. This could easily be done with sieves as well, but it has no real benefit. You simply run 2 tasks in parallel instead.
Some tasks are very linear (e g LLR) and are not easily threadable. This is because the calculations made, depend on the previous calculation. Usually you need to modify the present algorithm or invent a new one to make the calculation more threadable.
Also if you use multiple threads you might lose performance (i e the added threads might not give 100% speed per thread). In this case it becomes a choice between throughput (total work done) and latency (time for each task to complete).
/Snf
|
3)
Message boards :
Number crunching :
Getting CPU WUs when not requested
(Message 34148)
Posted 4466 days ago by Snf*
All WUs (GPU/CPU) complete successfully if I allow them to. The problem is that I get CPU PPS sieve WUs which I don't really want. I have no CUDA problems whatsoever. This is a task selection issue, not a CUDA/GPU issue.
I do get both types of task (GPU/CPU) and I complete both types of task. The problem is that I get CPU tasks I don't want.
/Snf
Okay but i believe all problems are not independent from each other. They are caused by the same basis.
I think Rytis must change again the predefined project settings if no CPU-app is selected, if this is possible with the current boinc server software...
I have specifically set both GPU and CPU projects and I want to run both, but with different projects. The problem is that sometimes I get the wrong WUs for CPU (i e I get GPU project(s), but for CPU).
It's probably the same problem that gives people who only want GPU WUs, CPU WUs, so in that sense the problems are related.
/Snf
|
4)
Message boards :
Number crunching :
Getting CPU WUs when not requested
(Message 34140)
Posted 4466 days ago by Snf*
14.03.2011 14:45:47 PrimeGrid Message from server: A minimum of 200 MB (preferably 200 MB) of video RAM is needed to process tasks using your computer's NVIDIA GPU
Maybe this answered your CPU-tasks.
Not in my case. I do not own a recent GPU with less than 512 MB.
/Snf
Okay, but the cuda-app want to have at least 200MB free video RAM...
Maybe you were running a video game or something else that used video ram. I believe there are some games which do not "release" the memory properly and thus require a re-boot before you can use GPU for calculating.
All WUs (GPU/CPU) complete successfully if I allow them to. The problem is that I get CPU PPS sieve WUs which I don't really want. I have no CUDA problems whatsoever. This is a task selection issue, not a CUDA/GPU issue.
I do get both types of task (GPU/CPU) and I complete both types of task. The problem is that I get CPU tasks I don't want.
/Snf
|
5)
Message boards :
Number crunching :
Getting CPU WUs when not requested
(Message 34130)
Posted 4466 days ago by Snf*
14.03.2011 14:45:47 PrimeGrid Message from server: A minimum of 200 MB (preferably 200 MB) of video RAM is needed to process tasks using your computer's NVIDIA GPU
Maybe this answered your CPU-tasks.
Not in my case. I do not own a recent GPU with less than 512 MB.
/Snf
|
6)
Message boards :
Number crunching :
PPS sieve CPU WUs not aborting correctly
(Message 34117)
Posted 4467 days ago by Snf*
As related to my other problem I have discovered that PPS sieve CPU units do not always abort correctly (when running). If aborted through BOINC the sieve process does not always die. This leaves a runaway process hogging resources. It will eventually die (probably when it has run to completion).
I have seen this on 64 bit Windows (Vista/7). Cleaning up through the task manager works fine, but is a hassle. It happens quite often for me, so it should be easy to reproduce.
/Snf
|
7)
Message boards :
Number crunching :
Getting CPU WUs when not requested
(Message 34116)
Posted 4467 days ago by Snf*
I have set up all my machines with capable GPUs to run PPS sieve on the GPU and some other type of WU on the CPU.This worked like a charm until a few days ago. Now I start getting PPS sieve CPU units from time to time, which is not what I wanted.
This happens when the client requests both CPU and GPU work at the same time. It either gets GPU work (fine), or it gets CPU work (but from the GPU project list). This applies to CW sieve as well (tested). It never gets work for both GPU and CPU. The type of work received is deterministic in any specific situation (as in if I abort the CPU WU, I will get a new one). I also only get one CPU WU at a time. When requesting only one type of work at a time, it seems to work as intended.
When the state of the client changes, it flips to getting GPU work. When there is enough GPU work, it starts asking for CPU work only and gets out of the loop. Of course when the work is completed, there is a chance of it flipping back again. There will still be some unwanted PPS sieve CPU WUs though. It seems like there is a mix up when requesting both types of work at the same time.
I have tried updating the prefs, reset and re-attaching. Furthermore this affects all my machines with this setup so it's not a client issue, since it worked in the past. This is most likely related to some kind of server update, intentional or not.
Any advice on how to proceed?
/Snf
|
8)
Message boards :
Number crunching :
The Year of the Rabbit Challenge
(Message 32008)
Posted 4506 days ago by Snf*
I am getting gypped. The systems with video cards where I say get jobs from other projects and only have PPS LLR checked is sending me all kinds of stuff instead of just PPS LLR on the CPUs even though there are plenty.
This is skewing my work and will really upset me if continues through the Rabbit challenge.
Just deselect GPUs on the same page where you pick which subprojects you want to participate in.
Also, make sure your computer is set to the correct Default/Home/Work/School for what you want to run.
You both missed the point...only PPS LLR is checked so that CPUs will get those while the "work from other projects..." is selected so that the GPUs do not sit idle. However, all sorts of non-PPS LLR apps are loading to the CPUs as well.
I agree with Pooh Bear...very frustrating.
I have seen this as well. Ended up moving my GPUs off primegrid during the challenge.
This is probably due to the server being unable to maintain enough PPS LLR work (i e the work queue is empty from time to time).
/Snf
|
9)
Message boards :
Number crunching :
2011 Challenge Series suggestions
(Message 30813)
Posted 4528 days ago by Snf*
SoB WUs don't crash because of "some unknown error," they crash because the slightest hardware issue (memory, CPU usage, etc) can cause them to miss a checkpoint, or something like that
That, by definition, is a some unknown error. More informative error messages may help.
I forget exactly what I was told regarding SoB errors, but it seems like the application is prone to errors on older or not-so-optimized hardware. I'd imagine that people who do lots of SoB tasks have overclocked CPUs or memory (or lots of components) which can make things unstable.
I do not follow that line because it does crash with new systems and without overclocking.
Perhaps is it just the time required as the same app is about to handle other subprojects without the same crashes.
Even new hardware may be marginal. The quality of consumer computer hardware is often more marginal than you might be aware of. There are a lot of stress test you can run to verify if you hardware is really stable.
A lot of consumer computer hardware is marginal because of everyone wanting to make (more) money. Marginal components are simply used to save money. In many cases this won't be noticeable until you really stress the hardware and very few programs do that.
Even if you buy components and assemble the computer yourself, you will still have to test it to be sure it's stable. The cheaper components you use, the more likely you are to get an unstable system.
/Snf
|
10)
Message boards :
Number crunching :
Dual ATI GPUs on PPS (seive)
(Message 30772)
Posted 4529 days ago by Snf*
I have had the same problem, but it disappeared when I upgraded to 10.12 (ocl, which includes ocl 2.3).
The second gpu will use a lot of cpu, but will produce valid results in PG.
It could be worth a try...
/Snf
|
Next 10 posts
|
