Testing multi gpu WU on gtx670m
14 Sep 2016 18:21:51 UTC
Topic 201121
(moderation:
Hello, since starting E@H almost a year ago, I still have not changed the work load for my GTX670m. My laptop is my best rig right now https://einsteinathome.org/host/12392381 and I plan on building a desktop rig soon. I used to game, but I'm way more committed to boinc now.
But, for the time being, I'm trying to optimize a fairly inexpensive (now) and quite power efficient laptop to pound out some data. So I finally changed my GPU settings. I'm going to try running 3 concurrent tasks right away, since I have seen that gtx580's have enough memory to do that, and I have equivalent.
I will post successes and probably failures here as well. Is there anything a guy new to gpu computing preferences should know?
Jordan
Make sure you leave CPU cores
)
Make sure you leave CPU cores free so you don't starve your GPU
I have had issues with mixed GPU apps running concurrently. Running BRP4 tasks at the same time as BRP6 tasks would result in a net slowdown. I ended up just running one of the apps (BRP6 in my case).
It's a laptop so cooling is an issue, but you already know that.
The ASUS g75vx has a pretty
)
The ASUS g75vx has a pretty good cooling system for a laptop, so far I havent had any issues with over heating. we'll see if that persists after upping the GPU workload
So, as an update, It still is
)
So, as an update, It still is not running more than 1 task. I currently have a few dozen WU with maybe 1/4 of them as BRP6 tasks, they all are labelled as (.2 cpu + 1 Nvidia GPU)
Not sure if I missed some setting...
Jordan
Okay so I just checked my
)
Okay so I just checked my manager and my compute setting were set to use local prefs, even though I set it to online prefs. I reset it to use online, and once I opened it again, it was set to local. Not sure what the deal is. is it because im using a manager?
Jordan
Some information that might
)
Some information that might help you understand what's going on ...
- This has nothing to do with local or global settings. GPU utilization is no "official" BOINC setting at all, being able to change it is a feature of the Einstein server that your client will never know about. There is a way to make local changes but I wouldn't recommend it in your case.
- The utilization factor is transmitted to your client with new tasks, not on a simple client update.
- It immediately applies to all tasks present, the old tasks will continue to show the old numbers though. You'll have to check how many tasks are actually running.
Okay, so I used to have
)
Okay, so I used to have spotty internet which is now much better, and as a result I switched from storing 5 days of work to 1 (extra) day. When I switched my workload setting is when I changed the compute preferences on the E@H server. If I'm understanding you correctly, I won't get new tasks for around 4 days and after that the compute changes take effect?
Currently I tried setting my cpu factor to 50% thinking i needed more cores available, but if i'm right, I want to put it at 99% to blow through the stored work.
This is fun to learn about.
Jordan
Right, you'll have to wait
)
Right, you'll have to wait some days or make BOINC fetch work earlier. But be careful, run times will be underestimated now and you might get more work than you expect.
When you're running 3 concurrent tasks you can easily try 2 too by suspending all but two tasks. Watch the GPU load, if it's still close to 100%, two concurrent tasks are enough. And I don't think you'll need to reduce CPU load to 50%, try different settings. CUDA doesn't need much CPU, but there's differences between BRP4G and BRP6.
it has started running 3
)
it has started running 3 tasks this morning, need to find a linux gpu utility to monitor load
Jordan
using the watch command in
)
using the watch command in terminal, I now know my gpu is running at 76c. It says it's running in P0, but that it's only using 668mb of the 3gb of memory.
Don't mind the memory, you
)
Don't mind the memory, you can't make use of it.