Advanced search

Message boards : Number crunching : Would this app_info.xml work?

Author Message
Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23533 - Posted: 19 Feb 2012 | 7:02:49 UTC

I wrote an app_info.xml for long tasks.

<app_info>
<app>
<name>GPUGRID</name>
</app>
<file_info>
<name>acemdlong_6.15_windows_intelx86__cuda31</name>
<executable/>
</file_info>
<file_info>
<name>cudart32_31_9.dll</name>
<executable/>
</file_info>
<file_info>
<name>cufft32_31_9.dll</name>
<executable/>
</file_info>
<file_info>
<name>tcl85.dll</name>
<executable/>
</file_info>
<app_version>
<app_name>Long runs (8-12 hours on fastest card) (cuda31)</app_name>
<version_num>615</version_num>
<avg_ncpus>0.36</avg_ncpus>
<max_ncpus>1.000000</max_ncpus>
<plan_class>cuda31</plan_class>
<coproc>
<type>CUDA</type>
<count>.5</count>
</coproc>
<file_ref>
<file_name>acemdlong_6.15_windows_intelx86__cuda31</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>cudart32_31_9.dll</file_name>
<open_name>cudart32_31_9.dll</open_name>
<copy_file/>
</file_ref>
<file_ref>
<file_name>cufft32_31_9.dll</file_name>
<open_name>cufft32_31_9.dll</open_name>
<copy_file/>
</file_ref>
<file_ref>
<file_name>tcl85.dll</file_name>
<open_name>tcl85.dll</open_name>
<copy_file/>
</file_ref>
</app_version>
</app_info>

Would this work? I'm running a Gianni task and the lowest reading in the last hour is 30% and the highest is 79%, with an average of 73%, checked with GPU-Z. The memory used is about 400MB. This way I could run like a SETI or MilkyWay wu with it to fully use the GPU. I know it slows down both WU's but my GPU gets fully utilized. And while running a Nathan which utilizes the GPU above 90% I would just change the <count> variable and restart BOINC.

And how would the project administrators/developers/scientists feel about getting results from an 'anonymous platform'.

Anthony.
____________

The longer I live, the more reasons I develop for wanting to die.

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23534 - Posted: 19 Feb 2012 | 7:21:33 UTC
Last modified: 19 Feb 2012 | 7:35:14 UTC

Is there a setting somewhere so that I only download 1 WU at a time. Or can I add a variable to the app_info so it will do that, for instance a very high flops number in the <flops> variable?

Anthony
____________

The longer I live, the more reasons I develop for wanting to die.

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1947
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 23536 - Posted: 19 Feb 2012 | 8:30:07 UTC - in response to Message 23533.

HI,
in principle there is no problem, but I don't understand why you did that?

gdf

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23537 - Posted: 19 Feb 2012 | 9:24:13 UTC - in response to Message 23536.

I want my GPU's to be fully utilized. If they are that means more work is done.

But the Gianni task for instance only utilizes my GPU for 50%-80%, so 20%-50% of my GTX-570 isn't used. As far as I can tell only Nathan tasks use more than 90% of my GPU. The GTX-570 is dedicated to crunching. I crunch also on my GTX-550TI and my TV is connected to it (I disabled screen output on my GTX-570 with Nvidia's control panel). It also crunches 2 wu's at any given moment, SETI wu's mostly and some MilkyWay wu's (MilkyWay only until I build my 3 GPU's system next week wich will have 2 Nvidia's and an ATI HD6950 especially for MilkyWay as MW wu's crunch faster on ATI's).

While I crunch 2 wu's at a time on my GTX-550TI I don't loose time while one WU finishes and has to reload another. So my crunching time doubles/wu but I do get more work done in the same amount of time.

An example: crunching 2 SETI WU's takes for instance 1 hour/wu for 2 wu's, while crunching 1 might take 32 minutes/wu. So doing 2 at once saves me 4 minutes in this example because my GPU is being 99% utilized all the time. And with those SETI and MilkyWay wu's I can just go watch TV with my TV-card or a movie/dvd without any stutter in my TV or movie/dvd, because I have enough video memory left over on my 1 GB card, and Media Center runs at a higher priority than BOINC tasks.

So I did write the app_info.xml to get my GTX-570 at 99% utilization so I can get more work done. That's how BOINC works. Using 'idle time' to do science work. But with most of the GPUGrid WU's I still have (a lot of) 'idle time' on my GTX-570, even with Nathan wu's but that's only about 6% (so that would be acceptable I guess).

I have app_info.xml's for SETI (comes with the optimized app), MilkyWay and Einstein and I changed all of them to do at least 2 WU's at a time on 1 GPU. 2 months ago I ran 3 at a time, so there was 1 SETI wu with 1 Einstein wu and 1 MilkyWay wu crunching together on my GTX-550TI.

I hope I explained it enough for you why I'm doing this. To get 99% utilization on my GPU's instead of 'idle time'.

Anthony.
____________

The longer I live, the more reasons I develop for wanting to die.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 23542 - Posted: 19 Feb 2012 | 13:08:52 UTC - in response to Message 23537.
Last modified: 19 Feb 2012 | 13:14:51 UTC

Optimizing is the way of the cruncher!
The unmentioned issue is that you are using Win7, which is at least 11% slower than XP for crunching at GPUGrid.
While using the anonymous platform should be fine in itself, it may not lead to increased performance overall. Other factors should be considered. The likely increase in task failures, especially when running mixed projects, is often detrimental overall. Mixed projects often results in gains for one at the expense of another (in the past I found GPUGrid task performance was terrible when running another non-GPUGrid task). I also found that when running at 80% or more there is no gain from running another task, and over 90% meant a loss. Temps rose as did power consumption, so increased running costs for no gain was a loser in my books. This was all on a GTX470, some time ago, and inspired by some tasks running at ~50% GPU utilization. The lowest I'm seeing on my setups is 75%, so definitely not worth it for me. Since then the apps have changed and you might be using a CC2.1 card, or mixed cards which may perform differently. You would need to re-build the app_info file every time a new app appeared, running the risk of repeated failures, possibly on multiple projects, and consequential outages.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23544 - Posted: 19 Feb 2012 | 13:35:01 UTC - in response to Message 23542.
Last modified: 19 Feb 2012 | 13:36:49 UTC

I run a GTX-570, a GTX-550TI and next week also 2 GTX-560TI-448's and an AMD HD6950 (all 3 in 1 system).
____________

The longer I live, the more reasons I develop for wanting to die.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 23547 - Posted: 19 Feb 2012 | 14:52:46 UTC - in response to Message 23544.

It will be interesting to see how you get on.

Good luck,
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23548 - Posted: 19 Feb 2012 | 15:10:35 UTC - in response to Message 23547.
Last modified: 19 Feb 2012 | 15:14:20 UTC

It's not all for GPUGrid. Maybe I'll try to get to 5M on GPUgrid, I want SETI and MilkyWay to get above 1M, maybe Einstein too, And then I'll see what I do.

Last oktober when a started crunching again I had about around 350K total BOINC credits (check my signature pictures for BoincStats and the other for FREE-DC if their site is up again; they're tinyurl links otherwise it wouldn't fit in the signature).

Begin oktober I bought the GTX-550TI and at the end of januari I bought the GTX-570. Both cards are factory OC-ed. So I went up fast since oktober and made a jump since 30-31 januari. If Free-DC is up you can check my milestones for BOINC total and projectwise.

It won't be for long that I can use my new system because I will be committing suicide within 2 or 3 to 5 or 6 weeks, when I receive the medication I ordered to do that. It's been shipped last monday but it might take a while to get here from the other side of the Atlantic. Sorry for bringing this up.

Edit: the Boincstats sig might contain the link for Free-DC, the other link (the team stats) is for my teams homepage. But in BoincStats you can look me up by my username.

Anthony.
____________

The longer I live, the more reasons I develop for wanting to die.

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23553 - Posted: 19 Feb 2012 | 18:32:41 UTC

There's an error in the app_info.xml file. When I update GPUGrid I get these messages:

GPUGRID Feb/19/2012 19:25:05 update requested by user
GPUGRID Feb/19/2012 19:25:10 Sending scheduler request: Requested by user.
GPUGRID Feb/19/2012 19:25:10 Not reporting or requesting tasks
GPUGRID Feb/19/2012 19:25:12 Scheduler request completed

It doesn't even request for tasks, while it should as I don't have enough tasks on my system. So what's wrong in the app_info.xml file so that the server doesn't recognize it?

Anthony.
____________

The longer I live, the more reasons I develop for wanting to die.

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23574 - Posted: 20 Feb 2012 | 22:09:29 UTC

With help from some guys at the BOINC forum I changed my app_info.xml to this:

<app_info>
<app>
<name>acemdlong</name>
<user_friendly_name>Long runs (8-12 hours on fastest card)</user_friendly_name>
</app>
<file_info>
<name>acemdlong_6.15_windows_intelx86__cuda31</name>
<executable/>
</file_info>
<file_info>
<name>cudart32_31_9.dll</name>
<executable/>
</file_info>
<file_info>
<name>cufft32_31_9.dll</name>
<executable/>
</file_info>
<file_info>
<name>tcl85.dll</name>
<executable/>
</file_info>
<app_version>
<app_name>acemdlong</app_name>
<version_num>615</version_num>
<platform>windows_intelx86</platform>
<avg_ncpus>0.36</avg_ncpus>
<max_ncpus>1.000000</max_ncpus>
<flops>167186362363.818660</flops>
<plan_class>cuda31</plan_class>
<api_version>6.7.0</api_version>
<coproc>
<type>CUDA</type>
<count>.5</count>
</coproc>
<gpu_ram>402653184.000000</gpu_ram>
<file_ref>
<file_name>acemdlong_6.15_windows_intelx86__cuda31</file_name>
<main_program/>
</file_ref>
<file_ref>
<file_name>cudart32_31_9.dll</file_name>
<open_name>cudart32_31_9.dll</open_name>
<copy_file/>
</file_ref>
<file_ref>
<file_name>cufft32_31_9.dll</file_name>
<open_name>cufft32_31_9.dll</open_name>
<copy_file/>
</file_ref>
<file_ref>
<file_name>tcl85.dll</file_name>
<open_name>tcl85.dll</open_name>
<copy_file/>
</file_ref>
</app_version>
</app_info>

Now I need to test if this works. I'm not sure about the .dll files either. I don't now if they are executable.

Anthony.
____________

The longer I live, the more reasons I develop for wanting to die.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 23578 - Posted: 21 Feb 2012 | 0:07:41 UTC - in response to Message 23574.
Last modified: 21 Feb 2012 | 0:08:51 UTC

Your GPU has 1GB and some of the long tasks use more graphics memory.
I think the server checks that it's at least 384MB GDDR. Rounding up/down errors could in theory prevent you getting work so I would suggest 512MB (536870912):


<gpu_ram>402653184.000000</gpu_ram>
to

<gpu_ram>536870912.000000</gpu_ram>
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23579 - Posted: 21 Feb 2012 | 0:18:11 UTC
Last modified: 21 Feb 2012 | 0:20:58 UTC

I've got the number from the client_state.xml file. I changed it in my own app_info.xml to 512 MB as you suggested. If anyone wants me to post the file with the changed value they can ask.

Remember I didn't test the file yet. So it could still contain errors and not work.

Edit: Your right about the memory on my GTX-550TI, the GTX-570 has 1.25 GB.

Anthony.
____________

The longer I live, the more reasons I develop for wanting to die.

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23583 - Posted: 21 Feb 2012 | 3:03:53 UTC

My app_info.xml works. Now I have to do some more testing to be sure utilization is 90%-99%, and test which application can do that best without errors.

Anthony.
____________

The longer I live, the more reasons I develop for wanting to die.

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 23590 - Posted: 21 Feb 2012 | 12:48:01 UTC - in response to Message 23583.

To test performance you will need to compare completed 'same-type-tasks' that previously ran without the app_info file in place. You will also need to have identical system and Boinc settings.

Then you would have to check error rates, outages.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Profile S@NL - Mellowman
Send message
Joined: 5 Oct 11
Posts: 47
Credit: 10,002,948
RAC: 0
Level
Pro
Scientific publications
watwatwatwat
Message 23595 - Posted: 21 Feb 2012 | 14:13:54 UTC - in response to Message 23594.

With a Nathan unit I only see MilkyWay take the GPU to 99% utilization, with SETI and Einstein it stays lower. Don't remember the exact figures cause I tested it last night. It's running with a SETI unit now, and in the short moment I checked utilization stays within 77%-90%.
____________

The longer I live, the more reasons I develop for wanting to die.

Operator
Send message
Joined: 15 May 11
Posts: 108
Credit: 297,176,099
RAC: 0
Level
Asn
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29217 - Posted: 22 Mar 2013 | 17:00:56 UTC - in response to Message 23578.

So the ram amount is how much each GPU card actually has or how much you're willing to allocate? (in the case of multiple identical GPUs...)

If it's how much each GPU has....a GPU with 6144mb then would be written as 6442451000.000000 correct?

I'm thinking of creating an app_info.xml to try on my system with the two Titans as well as the one with the two 590's.

The current long Nathans are showing that they are only utilizing 30-40% on the 590 box.

So I'm thinking of tinkering (which is old school for "optimizing") ;-}

What does the

<type>CUDA</type>
<count>.5</count>

relate to? Is that how many WU you want each GPU to process at a time?

I wonder if anyone has given any thought to an option like Einstein has where you set this preference on the site (the so-called 'Dangerous' BRP option) and your system then just picks it up that way and runs with it. I'm running x3 WUs per GPU in Einstein that way now on the two Titans with no apparent issues and only a slightly longer run time per WU.

Operator
____________

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29234 - Posted: 24 Mar 2013 | 10:34:46 UTC - in response to Message 29217.
Last modified: 24 Mar 2013 | 10:36:18 UTC

Yes <count>.5</count> means you are telling Boinc your card has 2 GPU's/to run two task on it.

Probably easier to use app_config.xml,
http://boinc.berkeley.edu/wiki/Client_configuration#Application_configuration
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Richard Haselgrove
Send message
Joined: 11 Jul 09
Posts: 1451
Credit: 3,575,929,351
RAC: 345,202
Level
Arg
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29235 - Posted: 24 Mar 2013 | 11:10:04 UTC - in response to Message 29234.

Yes <count>.5</count> means you are telling Boinc your card has 2 GPU's/to run two task on it.

Probably easier to use app_config.xml,
http://boinc.berkeley.edu/wiki/Client_configuration#Application_configuration

Not quite. It's telling BOINC that each task will run on half a card. So, if you have one card, you can run two tasks on it: if you have two GPUs, you can run four tasks and so on.

App_config.xml is certainly the easiest way to do this. You need at least BOINC v7.0.42 to use it: if you use v7.0.54 or later (the current alpha test version is v7.0.58) you can change the values and apply them ('read config file') while BOINC is running.

ExtraTerrestrial Apes
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 17 Aug 08
Posts: 2705
Credit: 1,311,122,549
RAC: 0
Level
Met
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29241 - Posted: 25 Mar 2013 | 16:01:56 UTC - in response to Message 29217.

I wonder if anyone has given any thought to an option like Einstein has where you set this preference on the site (the so-called 'Dangerous' BRP option)

Yes :)

It hardly provides higher throughput at GPU-Grid and increases the averaged turn-around time - which is not something the project desires. Getting tasks back quickly is much more important to GPU-Grid than most (if not all) other projects.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile skgiven
Volunteer moderator
Volunteer tester
Avatar
Send message
Joined: 23 Apr 09
Posts: 3968
Credit: 1,995,359,260
RAC: 0
Level
His
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 29250 - Posted: 25 Mar 2013 | 21:39:31 UTC - in response to Message 29241.

I think it's more likely the project would want one task to run over two GPU's than the other way round.
____________
FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Post to thread

Message boards : Number crunching : Would this app_info.xml work?

//