Two Xeon's go on a date with a RX480...only wants one GPU task at a time
27 Sep 2016 5:40:28 UTC
Topic 201462
(moderation:
Long time listener, first time caller...
There are these identical twins that live in the same place as me. I have on good information that these twins do everything together. Not only that, they can take on a huge amount of work at the same time. I have heard of multitasking, but word has it each of them can multitask EIGHT times each without breaking a sweat. Rumor has it they can even do double that EACH! I have only read about it in the local paper (The Log) that they can each do 16 tasks per twin; so 32 tasks at the same time!
Now I like to workout as much as the next guy but these ladies are a little intimidating. I do have a thing for older women; I love their maturity, they are pretty grounded in reality, have less baggage, know what they want in life and are pretty content with their bodies. I am a bit of a newcomer myself and was sold on this place being low rent but the HVAC does not always work so well as other places I have heard about. (Phase cooling, local pool, hell even a place that is completely under....oil I think) I like it natural and air cooled is the way for me. No natural disasters thank you. I do not need to wake up to a flood or have it rain indoors! Steam baths and showers are for keyboards and mice.
Anyway, I am afraid to take on more than one of them at a time because I think I can not handle it. I tried going out with one of them and then the other the next day; it was awesome! This was great, then they told me they never do anything apart and want to be together with one guy and one guy only (1.0). I mean how could I turn it down. So I have tried going out with both of them at the same time and things have been great(0.5).
They agreed at first to just be with me and not do other work at the same time. They have been really patient with me, however I know they have been yearning for more. I know they can do more so we have slowly introduced me going out with them both at the same time and surprisingly enough they can even take on 30 other tasks while still giving me all the attention I could handle.
It came up the other day that they suggested I try to do a little multitasking of my own and they would help me; however it was a disaster (0.33). I kinda concentrated on them, then I could not focus on the additional task. So we thought it was because the time was being split 3 ways so we went to 4-way (0.25).
Just to be sure they dropped all their other tasks and I still could not focus on anything but the two of them at the same time (0.5). On a hunch of their best friend, The Operator; (Who I thought for sure lead an alternative lifestyle since he doesn't pay hardly any attention to these hot twins) he suggested I stretch out my horizons more and just see what an 8-way would feel like (0.12). It felt the same as anything more than just the two of them (0.5). The Operator it turns out is just incredibly busy working two jobs, juggling two living arrangements, and spends time with his family locally. Oh, and after talking with the Operator a few times he is as straight as a flagpole.
Now the twins have 32 Gigs of playground to work with in their place but they share it all (and would share with me if I needed it). I have a smaller pad at 8 Gigs. However mine has been recently remodeled and is very modern. They spend more time at my place than I do at theirs.
My question is: I know I have this killer view off the penthouse that looks across all the city lights and on to top it off the building sign "XFX" sits right on top of my roof and can been seen for miles. Why can I not focus on more than the two of them at the same time (0.5)? They keep telling me I can do more. What am I doing wrong?
I have been reading on https://einsteinathome.org/content/looking-get-started-gpu-processing and reading all the links, etc. and have not been able to come up with a solution to my problem. To make matters worse I have two older brothers that basically have the opposite situation as me. They are R7 265s' and their lady is closer in age (i5-2400K). They tell me they can do two tasks each (four total between the two of them) and their girl can still take on a few tasks if she wants do but she likes watching them workup a sweat. Now she has a big pad at 32 Gigs, but each of my brothers have a small studio at 2 Gigs each. But they can do more than I can? How does that work? Now granted they are in a little older building (Sapphire) but they don't have a huge sign that lights up the sky like I do!
...
It is not the authors intention to offend anyone with the aforementioned metaphors.
I have read about other users getting multiple GPU tasks on their workstations. I have tried following the directions, links, made sure my preferences at set for each workstation appropriately. Made sure to keep my workload low (0.25 days), and always refresh if I change a preference.
If it was not apparent; I have a Dual Xeon E5-2670 v1, 32G ECC RDIMM, XFX RX480 8GB with air cooling and nothing gets over 65C except the video card which stays an even 72C. This system will run an all core clock of 3.0 GHz. I read THAT same article from TechCrunch earlier this year and thought, "Ooooohhhh" $700 later I am in it up to my eyeballs of amazement. I am running Windows 10 Pro (don't laugh too hard) Anniversary update blah blah, and have the latest drivers for the RX480 installed but am running Dec 2015 drivers I think on R7 265 cards.
Looking forward to some sage advice. Its great that there are so many down to earth guys backing BOINC and the Einstein@Home project.
Thanks,
Rick.
The website settings for GPU
)
The website settings for GPU utilization factor only take effect after new work for that GPU-app is downloaded. A forced update from Boinc will not help. If I remember correctly the new setting effects all downloaded tasks but the numbers shown in Boinc will not changed for the currently running tasks unless Boinc is restarted. Might even be that only newly downloaded tasks will show the right numbers.
Other than that I seem to remember that there is/was problems with using different venues/locations after the website upgrade, I'm not sure if those problems has been solved yet. If your using any other location than "Default" try changing it and see if that helps.
There is always the option to create an app_config.xml to control how many GPU-tasks to run concurrently, the biggest advantage is that you can issue a "read config files" command in Boinc and changes apply directly, the drawback is that it needs updating if new searches are launched, although it only controls what's in it so searches not configured in the file will use the website settings. If you would like to test this approach I can provide an example and some guidelines about it.
Holmis wrote:Other than that
)
FYI, the preferences venue problem should be solved. That being said the project preferences (where GPU settings are kept) should have not been affected by that problem.
Einstein@Home Project
Hi Holmis, Thanks for the
)
Hi Holmis,
Thanks for the info. I thought that might be the case if the client thought it had to finish its tasks before starting new ones. So I waited until it finished its current tasks and it initiated more GPU tasks just like I had asked it to. The problem was anything over two (2) GPU tasks and the progress comes almost to a screeching halt.
With two (2) tasks it usually does about 70% per hour on each one give or take 10%.
I am using the default location for this machine and it appears to get every change I initiate on the web to the client. For example I can tell it to use 92% of its processors and it will limit itself to thirty (30) processes. If I change it to 50% then its sixteen (16), etc. etc. The CPU tasks crunch along just fine it is only when the GPU tasks >= two (2).
I understand there should be some slowdown, but we are talking from 70%/hour to 0.433%/hour? [Yes less than 1%]
I did try suspending all the current tasks and let it start new ones with the new GPU setting (0.25 and 0.12) and it fired up all the tasks like it said it was but was painfully slow.
I read up on making an app_config.xml file but I thought that was to manually control how much processor time is given to each GPU task. It does not appear to be a processor time limitation though. I mean there are 16 cores to go around and if the OS wants to 32 threads.
I guess I could just try and set the GPU tasks to 0.25 and just ride it out for a few days. I just hate loosing the progress. I guess it does not matter as much since my older home machine is beating the pants off this server machine with the new RX 480 8GB card.
I have read others having a similar Xeon machine, has anyone else gotten more than two (2) GPU tasks per card to work?
Thanks,
Rick.
app_config can immediately
)
app_config can immediately change the # of tasks per GPU after the config files are read. I run 3x BRP6 WUs on my 1070 and 970 all on the same 3770k. I'll have to switch that over to something else when my queue runs out.
I know others run more than a single task per card.
A single task on my 1070 runs at 61% util and 3x run at 85-90% util so there is more work be done. Each task is slower but the total time for 3 tasks is less.
Think I kind of misunderstood
)
Think I kind of misunderstood your first post and thought you couldn't get more than 2 to run but if I understand correctly the tasks slow down to almost no progress when running more than x2 GPU tasks.
It's been my understanding that OpenCL based tasks here at Einstein do need one CPU core/thread free to support the GPU tasks else they slow down considerably, for the most part I think people leave on CPU core/thread free per GPU tasks.
What happens if you suspend all CPU tasks and try to run x3 GPU tasks? Are they still slow? It might due to memory bandwith as the applications seem to require a lot of it. So running 16 or even 32 CPU tasks will require quite a lot of memory bandwith and that may starve the GPU.
Using an app_config.xml when doing these kind of experiments is most convenient as you only need to change the parameters in the file and then issue a "Read config files" command in Boinc Manager and the changes take effect immediately as to updating the settings on the webpage and having to wait for new tasks to be downloaded for the setting to take effect, this only applies to the GPU utilization setting, all other settings are propagated and take effect when Boinc contacts the server.
To use a app_config.xml file open Notepad and copy the following into it:
Then save the file as app_config.xml in boinc\projects\einstein.phys.uwm.edu
To find out where that path and folder is restart Boinc and open the Event Log (Ctrl+Shift+E) in Boinc Manger and look for something similar "30/09/2016 00:02:25 | | Data directory: G:\Boinc" in the first few lines. The standard path is C:\Program Data\Boinc\projects\einstein.phys.uwm.edu and that's a hidden folder, if you get "Data directory: C:\Program Data\Boinc" then you can copy the above path and paste it in File Explorer to get to the right folder. Make sure the name of the file is exactly app_config.xml and not app_config.xml.txt
You can accomplice this by making sure to select "All files" as the file format when saving instead of "Text-document (.txt)".
The example above will instruct Boinc to run x2 Binary Radio Pulsar Search (BRP4G) GPU tasks at once and "reserve" one CPU core/thread for support, it actually instructs Boinc that 2 GPU tasks will use 1 CPU so don't run as many CPU tasks. To run three change the gpu_usage to 0.33. You can also play with the cpu_usage setting to instruct Boinc to start even fewer CPU tasks by setting it to 1.0 so that you get one less CPU task for every GPU tasks started.
And next time feel free to ask in the Cruncher's Corner where most of the performance discussions take place and quite a few really clever people hang out!
Hi Holmis, Well I was a
)
Hi Holmis,
Well I was a little disappointed that I had a few suggestions to try out and then all the GPU tasks were gone for a few days. I see they are back now.
So I left my machine to run 30 processes and I came back after a long day at work today and it was set to 0.25. In my work list I had eight (8) GPU processes running. Each for 0.5 CPU and 0.25 GPU this was before I created the XML file you stated.
Most of the tasks stated they had been working for about 15 hours and had an estimated completion time of ~92 days give or take 15 days.
Even cutting off all the other processor tasks they still remained the same speed. Finally I was able to get some movement if I suspended all but three (3) tasks.
I am going to give each process a whole processor (I have lots to spare) and leave it at the 0.33 for a while to see how it does. I doubt it is a memory bandwidth unless the program is trying to move the whole queue in and out for each compute cycle of the 590,000. This would be a terrible design and horribly inefficient use of resources.
I did see a few articles about people with the R9 chips not being able to run more than two (2) tasks per GPU.
What I did find odd is that BOINC sees my GPU as having Device0 and Device1 even though there is only one (1) RX480 card installed. Along those same lines it is interesting that running two (2) tasks is fine, three (3) is the new high for simultaneous tasks, however four (4) slows things to a glacial pace.
Off to bed, its late and work is only a handful of hours away.
Is the RX480 a double GPU
)
Is the RX480 a double GPU card? I know Nvidia has had a few models with 2 in 1 cards.
If not then somethings not right here. From your description it sounds as you're actually running 6 or 8 tasks concurrently on one GPU and I can't remember reading about anyone who have managed that efficiently. Normally people run 2-4 task / GPU.
Could you post the startup messages from Boinc's event log? Especially the part with the GPU detection.
Holmis wrote: Is the RX480 a
)
No, it is a single GPU card. I am running 2 tasks quite comfortably on the RX-480. although i have had 8 running (slowly) on BRP6
OP has two hosts the second does have a dual GPU card. The Xeon seems to have, when last contacted server only one GPU, the RX-480.
Also setting the debug_coproc flag will be revealing.
Hi Holmis and AgentB, I have
)
Hi Holmis and AgentB,
I have not had a chance to sit down and post a debug log. I will hopefully get to that in the next few days.
Thank you for the ideas!
Hello, Here is what I
)
Hello,
Here is what I gathered from the LOG after suspending the client and restarting it. I took AgentB's advice and turned on the coproc_debug flag.
I noticed that the log refers to the video card as ATI instance 0.
This is with the following settings configured on E@H webpage. Which I presume are being overridden by the APP_CONFIG.XML.
BRP = 0.5
FGRP = 1.0
GW = 1.0
If I change the # of processors up to four (4) (which is 13% on this system) it simply adds more CPU tasks as expected.
Because I have an APP_CONFIG.XML I have to change it there instead of the E@H preferences correct?:
BRP = 0.25 <-- Should do four (4) tasks at a time correct? (I changed it to 0.25 on the web preferences and in the .XML file)
This is what the log looks like when I changed it via the E@H preferences:
Changing the APP_CONFIG.XML to reflect the BRP = 0.25 looks like:
Not much change so I will wait an hour or so to see what the GPU tasks to when they are up. They were running about 96% each per hour when it was assigned BRP = 0.5.
It is when BRP < 0.5 that it has always freaked out and gone glacially slow....