Multiple WU on 1 GPU
19 Jan 2018 12:52:01 UTC
Topic 212802
(moderation:
Hey guys,
As I understand it, for a gpu to crunch multiple WUs, I'll need to create an app_config.xml file.
I've done them for a couple other products, however I'm a little confused by how I'm supposed to do this for e@h. The giant list of apps doesn't help lol.
Specs:
i7 4770k
GTX 770
I'd like to allow 2 WUs on this GPU. I also read somewhere that only 1 e@h app uses gpu? which one is it?
Thanks for your patience
<app_config><app><name>hsgamm
)
<app_config><app><name>hsgamma_FGRPB1G</name><gpu_versions><gpu_usage>.5</gpu_usage><cpu_usage>1</cpu_usage></gpu_versions></app></app_config>Roy Loy skrev:As I understand
)
Actually here @Einstein you could also go to your project preferences and scroll to the bottom to find 3 sets of GPU utilization factor settings. Just change the one for FGRP apps (0.5 for x2, 0.33 for x3) and click save, then download new work for the GPU and the setting will apply to all GPU work in your cache. No need for an app_config.xml, although you could use one if you want and it will override the webpage settings. The downside to using the webpage setting is that you need to download new work for that application for the change to take effect, issuing an "Update" command in Boinc will not work with this particular setting. A small increase of the cache setting will often get the job done, just remember to lower it again afterwards.
At this moment that would be "Gamma-ray pulsar binary search #1 on GPUs" or FGRPB1G for short.
awesome thanks! thanks
)
awesome thanks! thanks extremely useful!
my gtx770 actually died lol.. got a r9 280x in lieu of the newer geforce 11xx cards being launched soon.
Holmis написал:Roy Loy
)
Thx it works! I have power load on my GPU's ~66% when it set to 0.5. So I can set 0.33, right?
vitaliy.mikhailenko wrote:Roy
)
It means you can write down the time it's taking to do each wu at the 0.50 setting and then try the 0.33 setting and check the time to completion that way too, at some point it will be slower doing more wu's at the same time.
vitaliy.mikhailenko skrev:Roy
)
As Mikey said, you should run your own tests as every system is different and Nvidia and AMD cards behave differently.
Start by running x1 or 1.0 for the "GPU utilization factor of FGRP apps". Note the average time for a dozen task or so. Then change to x2 or 0.5, again note the average run time for a dozen tasks or so but this time divide the average by 2, your running twice the number of tasks now! Then rinse and repeat for x3 or 0.33 but divide by 3. You might even try x4 or 0.25 but I don't think it's worth it unless your running an AMD card. Find the sweet spot and stick to it even though other metrics might say different.
Remember that Nvidia card do need a CPU thread free or "reserved" to perform at full potential whereas AMD seems to be fine without it. This might also be worthwhile to explore with different settings.
vitaliy.mikhailenko wrote:...
)
You can certainly use that setting but it could easily be quite disappointing. As others have suggested, you really need to test each option with a degree of care in order to find the optimum configuration. All my experience is with AMD GPUs. I find x2 is best for mid-range units. Even for GPUs like RX 580s which are above mid-range, and which run very nicely at x3, the improvement over x2 is relatively small.
As a general rule, my impression is that it's highly likely to see an improvement for running x2 over x1 in situations where there is enough VRAM to sustain two tasks. You don't have a problem there, but others reading this might. As a rule of thumb, calculate on 1GB of VRAM per task. People with 2GB cards should limit themselves to x2. Most of the output gain to be had will come just from changing from x1 to x2.
If you have 3GB or greater, it's worthwhile testing at x3. Don't expect a huge improvement, particularly if it's not a top end, latest architecture card. Another factor to consider is the extra power that will be used compared to the extra output you get. It's not just the cost of power but the extra stress that is put on the machine. These are all things to be aware of in choosing your settings.
As you have an 8-core host and two 1060 GPUs, when running x2 on the GPUs, 4 CPU cores will be prevented by BOINC from running CPU tasks. If you change to x3, that number would increase to 6 cores. Whilst it is possible to override this automatic restriction, the reports show that GPU performance will suffer considerably if you do. Quite a few people running mid to high end nVidia cards choose to disable CPU crunching completely in their preferences. If you were to choose running x3 permanently, that might be a good option to consider.
Cheers,
Gary.
Hi all, i've tested with
)
Hi all, i've tested with different settings - 1 thread for 1 GPU is best for mine 1060(6). Now i overclocked it and have 12 minutes for 1 task.
vitaliy.mikhailenko wrote:Hi
)
It's looking good!!
Yes I agree Mikey, There is
)
Yes I agree Mikey,
There is a theory of physics going on here. We operate as far as I know with Fermion GPU's and not Boson GPU's.
Until we have photonics instead of electronics running our devices, only one electron can occupy one space. If we add more than one task to a GPU we simply are overloading the buss. Bang! goes the GPU.
This is only a suggestion.
Work runs fine on Bosons reacted into Fermions,
Sunny regards,
earthbilly