Work Unit odditys

Ryan
Ryan
Joined: 25 Nov 14
Posts: 36
Credit: 54,343,600
RAC: 2,149
Topic 197847

Can the same type of WU come in different sizes?
I have this type of WU running on 2 boxes :

Binary pulsar radio search (Perseus arm survey) 1.39 (BRP5-opencl-ati)

Running on my 290x its taking aprox 4:30 mins, on my 270 its taking aprox 1:30?

Is this right? it seems to be taking a long time on my 290x for a GPU unit? how long should these units take?

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,491
Credit: 63,169,175,770
RAC: 54,269,657

Work Unit odditys

Hi Ryan, welcome to the project!!

Quote:
Can the same type of WU come in different sizes?


Not in this case.

Quote:

I have this type of WU running on 2 boxes :

Binary pulsar radio search (Perseus arm survey) 1.39 (BRP5-opencl-ati)

Running on my 290x its taking aprox 4:30 mins, on my 270 its taking aprox 1:30?

Is this right? it seems to be taking a long time on my 290x for a GPU unit? how long should these units take?


Your 290x is possibly being starved for CPU support. You could try running 4 concurrent GPU tasks (which would automatically reserve 2 CPU cores for GPU support and you should get better performance. You could also experiment with further freeing up of CPU cores through computing preferences.

Are you sure your card is in a 16x slot? PCIe bandwidth is also very important for good performance and that could possibly be the problem.

Your R9 270 (basically a HD7850) will also perform better if you run 4 concurrent GPU tasks. You should get 4 completed tasks in a little over 4 hours - ie just over an hour each, rather than 1:30 when running them singly. I get that on 7850s even when driving them with a low end dual core / 4 thread i3. In your E@H preferences, set the GPU utilization factor to 0.25 for BRP searches. The warning is just for CYA purposes. I've never had a problem nor seen anyone report one if sensible cooling solutions are being used.

If you change the GPU utilization factor on the website, it will only kick in on your host after a new task is downloaded. It will then apply to all tasks in your work cache. For some reason you have a whole bunch of tasks on your 5930K and only a single 'in progress' task on your 3770K. Are you using different cache settings on the two machines? What values do you actually have for the two settings?

Cheers,
Gary.

Ryan
Ryan
Joined: 25 Nov 14
Posts: 36
Credit: 54,343,600
RAC: 2,149

Ok just checked and on the

Ok just checked and on the 270 tjhe GPU is being maxed out, hitting 100% usage, on the 290 its barly touching 20%, confirmed with power draw and fan speed, any ideas on getting the 290 back up to speed? a setting I may have wrong?

With regards to the 3770k, I have it set to only use 2 CPU cores and the GPU, the CPU atm is running through a load of Seti units I still have.

Ryan
Ryan
Joined: 25 Nov 14
Posts: 36
Credit: 54,343,600
RAC: 2,149

Hi, In answer to your

Hi,

In answer to your other question, yep its sat in a 16x slot

Ryan
Ryan
Joined: 25 Nov 14
Posts: 36
Credit: 54,343,600
RAC: 2,149

Also it does seem to have

Also it does seem to have downloaded a lot of CPU units, I have this and Milky Way at home running, both set to 100 in resource usage and Einstien is all it seems to crunch even though a small amount of Milky way has been downloaded.

CPU wise things also might be going slow, Gamma ray pulsar search (FGRP4-SSE2) units are taking about 11 hours.

Just set the 3770k to do cpu units as a test, its downloaded a Gravitational wave search unit which it thinks will take 31 hours, I have one queued on the 5930k and that estimates 77 hours to completion.

Strange things, I can only assume something with the boinc client, I may try a reinstall

Edit, just downloaded Milky way unit on the 3770k and its taking slightly longer than the 5930k so its working as should be, just seems to be issues on this machine with Einstein, might do a reinstall of the client and a full reset of the options, start again and see what happens.

Ryan
Ryan
Joined: 25 Nov 14
Posts: 36
Credit: 54,343,600
RAC: 2,149

Ok reinstalling Boinc seems

Ok reinstalling Boinc seems to have fixed it, getting full GPU usage now and the WU is estimating about a hour 50.

Does this look fine now for the WU's including the CPU ones?, still looks a little slow for me as the 270 is doing them in a hour and a half?

https://i.imgur.com/Z25uuMC.png

Also its not downloaded a crap ton of units like it did before.

Ryan
Ryan
Joined: 25 Nov 14
Posts: 36
Credit: 54,343,600
RAC: 2,149

Ok dropped the usable CPU

Ok dropped the usable CPU cores down to two on this machine and the 290x is now estimating just over a hour per unit and the GPU is and I can hear it really ramping up.

So looks like I was starving the GPU even though it still says running 0.5 CPU and one GPU, whats up with that?

So really how many free cores does the GPU units need? for Seti one seems to be fine.

Sigh edit, the 270 is showing 53 min estimate for the Aricebo units so this is still running slower :s

What should a 290x do that unit in?

Gary Roberts
Gary Roberts
Moderator
Joined: 9 Feb 05
Posts: 5,491
Credit: 63,169,175,770
RAC: 54,269,657

Five separate posts in less

Five separate posts in less than an hour??

You need to slow down a bit!! :-). - or at least keep adding your extra thoughts/concerns to the original message so it's a bit easier for an old geezer to keep up :-).

Quote:
With regards to the 3770k, I have it set to only use 2 CPU cores and the GPU ...


Which is why the GPU is able to go full steam.

Quote:
Also it does seem to have downloaded a lot of CPU units, I have this and Milky Way at home running, both set to 100 in resource usage and Einstien is all it seems to crunch even though a small amount of Milky way has been downloaded.


And do you also have Seti as part of the mix? If you do, and Seti can't supply, it might explain why you got a bunch of Einstein tasks. BOINC tends to do things like that. Until you get properly established, keep your cache settings low - say 0.5 to 1.0 days until things settle.

Quote:
CPU wise things also might be going slow, Gamma ray pulsar search (FGRP4-SSE2) units are taking about 11 hours.


From what I can see, there is no problem with CPU tasks.

With your actions in re-installing BOINC on the 5930K, you have managed to create a second host ID for that machine. BOINC is designed to be installed 'over the top' of an existing installation and if you do that, the original ID will be re-used and a bunch of unnecessary downloading will be avoided. Also, you wont consign all the tasks on the original installation to 'limbo land' until they eventually time out and get re-issued to someone else.

If you'd like to be kind to a whole bunch of other people, you should 'merge' your original host ID 11694600 by clicking the 'merge' link at the bottom of the 'host details' page. Be warned that this will get back over 300 tasks that were in progress on the old ID. You probably have enough computing power to process them all well within deadline, but if you really don't want them just abort them all and they can be re-issued immediately to other hosts without having to wait two weeks to expire.

Here's a suggestion for you. You have two really nice crunching boxes. Einstein has 4 really nice searches just itching to be crunched :-). For a whole bunch of complicated reasons, if you do all 4 searches on each machine, you might be troubled (until you get used to things) by wonky estimates of crunching times that never seem to settle down properly. So just put two of the searches on one box and the other two on the second. You will find that a lot easier to understand and tinker with.

Your 5930K will get back a bunch of FGRP4 tasks and BRP5 (Perseus Arm Survey) tasks if you do the above merge. So why not put both your new and old IDs for this machine into say, the 'home' venue and set the project preferences to allow only the BRP5 and FGRP4 runs. Your 3770K could be put into the 'school' venue where you would select just the BRP4G (Arecibo GPU) and the Gravity Wave searches. Once you have done that, each host will clear out any tasks for the 'wrong' searches and then only get new tasks for the 'right' searches in future.

You should treat estimates as potentially being 'wrong' for a while until things settle down. All 4 searches have reasonably repeatable crunch times. However, this project still uses DCF (Duration Correction Factor) to refine estimates and the BOINC limitation of a single, 'per project' DCF rather than 'per search' DCF, causes each individual search to potentially adversely affect every other search. That's why it will be easier to have just two searches per host, at least initially :-).

Quote:
Just set the 3770k to do cpu units as a test, its downloaded a Gravitational wave search unit which it thinks will take 31 hours, I have one queued on the 5930k and that estimates 77 hours to completion.


I would expect that either machine will be able to do a GW task in something like 10-15 hours. The estimates will get better as completed work adjusts the DCF. I would also expect that FGRP4 tasks would take around 8-12 hours. The completed work on your 'old' 5930K shows around 8.5 hours.

Quote:
Ok dropped the usable CPU cores down to two on this machine and the 290x is now estimating just over a hour per unit and the GPU is and I can hear it really ramping up.


That machine has 6 cores / 12 threads. One free virtual core should be more than enough to support a single GPU task.

Quote:
So looks like I was starving the GPU even though it still says running 0.5 CPU and one GPU, whats up with that?


Each task comes with a 'built-in' estimate of the CPU and GPU resource requirments for that task. For AMD GPUs it's exactly as you list. If you were crunching 2 GPU tasks concurrently, the numbers would change to 0.5 CPU + 0.5 GPU per task so that for both to run BOINC would ensure that a full free CPU core was available, automatically. There are ways of refining these resource requirements and you may choose to use these features as you tweak your setups.

Quote:
So really how many free cores does the GPU units need? for Seti one seems to be fine.


For AMD GPUs (which do perform really well on this project) the automatic 'built-ins' work quite well. You will (eventually) want to run multiple GPU tasks concurrently if you are at all interested in performance. If you do get to that stage, I would suggest the following scheme:-

1 or 2 concurrent GPU tasks - 1 free core
3 or 4 concurrent GPU tasks - 2 free cores
5 or 6 concurrent GPU tasks - 3 free cores

At the end of the day you can't make hard and fast rules - there are too many variables. You need to experiment and find what works best - that's part of the fun :-).

Quote:

Sigh edit, the 270 is showing 53 min estimate for the Aricebo units so this is still running slower :s

What should a 290x do that unit in?


Estimates are irrelevant. Just wait for completed tasks and see what they take. On your 'old' 5930K, BRP5 seemed to be taking around 2.5 hours. With a free core or two they should take a lot less, probably just over an hour. Please note that BRP5 (Perseus Arm Survey) and BRP4G (Arecibo) are completely different searches. BRP5 should take 2.5 to 3 times longer to crunch compared to BRP4G. They 'pay' 3.3 times more so that's a small 'bonus' for doing them :-).

Cheers,
Gary.

Claggy
Claggy
Joined: 29 Dec 06
Posts: 560
Credit: 2,557,091
RAC: 0

RE: And do you also have

Quote:
And do you also have Seti as part of the mix? If you do, and Seti can't supply, it might explain why you got a bunch of Einstein tasks. BOINC tends to do things like that. Until you get properly established, keep your cache settings low - say 0.5 to 1.0 days until things settle.


Of Ryan's four hosts, two have the minimum buffer set to 0.1 days, one to 10 days, and the last to 0 days.

Claggy

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.