New Modfit Runs
log in

Advanced search

Message boards : News : New Modfit Runs

1 · 2 · Next
Author Message
Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist
Send message
Joined: 25 Feb 13
Posts: 440
Credit: 10,409,099
RAC: 203,541

Message 62811 - Posted: 9 Dec 2014, 21:36:18 UTC
Last modified: 9 Dec 2014, 21:36:44 UTC

Hey Everyone,

Just a heads up, I will be starting new runs over the next couple days that will have a different step size in a computationally intensive section of our code. Through some work we did, we found that this step size could be optimized and reduced to improve the speed at which our code runs without sacrificing accuracy. We expect to see these runs complete in about 1/5th the time depending on the platform. Credits will hopefully still be calculated correctly, but if you think they are not for these runs please post here. If these runs cause you any trouble with these runs or you notice you stop receiving work units please post that here too. These runs will all include the tag _fast_ in the work unit name so it should be easy to track.

For now, these optimizations will only be used on the Modfit code, but in the near future we plan on also using them on original separation code, too.

If you have any questions about this optimization or problems, please let me know,

Jake

Profile Mumak
Avatar
Send message
Joined: 8 Apr 13
Posts: 89
Credit: 515,960,170
RAC: 3,220

Message 62822 - Posted: 11 Dec 2014, 11:47:05 UTC

Any chance to look at the high error rate of current Modfit units ?
I have no issues with the default MW application, only a few errors during several days.
Today I gave Modfit another chance, but after ~1 hour I have disabled them again - 10 invalids during such a short period and that's probably not the end.

aad
Send message
Joined: 30 Mar 09
Posts: 51
Credit: 249,256,626
RAC: 344,051

Message 62823 - Posted: 11 Dec 2014, 15:13:40 UTC
Last modified: 11 Dec 2014, 15:14:09 UTC

My first "fast" http://milkyway.cs.rpi.edu/milkyway/result.php?resultid=903753428
12 secs and 26,74 credits
Normal modified 52 secs and 213,76 credits

Sooooooooo..............

Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist
Send message
Joined: 25 Feb 13
Posts: 440
Credit: 10,409,099
RAC: 203,541

Message 62824 - Posted: 11 Dec 2014, 18:41:49 UTC

Looks like the credits aren't scaled exactly right. Pretty sure we neglect some constant overheads when we are scaling things so I will talk to Travis about fixing this.

Thanks,

Jake

Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist
Send message
Joined: 25 Feb 13
Posts: 440
Credit: 10,409,099
RAC: 203,541

Message 62831 - Posted: 12 Dec 2014, 20:57:59 UTC

After some checking, I think the credits might actually be scaled properly based on the number of flops required to complete the work unit. The problem is that now the runs might not fully utilize your GPU. It might be a good idea to try configuring your GPU to run 2 of these at the same time and see if that makes a difference.

Jake

aad
Send message
Joined: 30 Mar 09
Posts: 51
Credit: 249,256,626
RAC: 344,051

Message 62832 - Posted: 12 Dec 2014, 23:23:02 UTC - in response to Message 62831.
Last modified: 12 Dec 2014, 23:32:42 UTC

It might be a good idea to try configuring your GPU to run 2 of these at the same time and see if that makes a difference.

Jake


That might be a problem because the normal modfit takes 98% of the GPU (with a CPU core to feed it)
If I can make a choice to do only 'fast' modfits it might work with 3 at the same time because it only take about 33% of the GPU (sometime less)

Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist
Send message
Joined: 25 Feb 13
Posts: 440
Credit: 10,409,099
RAC: 203,541

Message 62844 - Posted: 15 Dec 2014, 16:17:32 UTC
Last modified: 15 Dec 2014, 16:26:35 UTC

Modfit will be switching over to using only fast runs as these slow ones finish up. You can expect them to be all the faster runs by the end of the week.

Thank you for the feedback,

Jake

Profile Cliff
Avatar
Send message
Joined: 28 Nov 14
Posts: 45
Credit: 54,184,495
RAC: 117,647

Message 62845 - Posted: 15 Dec 2014, 16:27:36 UTC - in response to Message 62844.

We will be switching over to using only fast runs as these slow ones finish up. You can expect them to be all the faster runs by the end of the week.

Thank you for the feedback,

Jake

On a slightly different topic, since yesterday nearly every single unmodded ask I've processed has ended up as a computational error, some 30 odd before I realised what was going on, nothing has changed on my rig...
Has there been an alteration at MW@H end?
I've stopped processing the affected WU and am now only doing modified, seemingly without problems other than delayed validation.

Regards,
Cliff

Tom*
Send message
Joined: 4 Oct 11
Posts: 33
Credit: 268,813,511
RAC: 397,537

Message 62848 - Posted: 15 Dec 2014, 21:38:36 UTC

Cliff All your failed tasks are also failed by the other two users doing them.

In other Words Bad Tasks. "May Have Bug"

Profile Cliff
Avatar
Send message
Joined: 28 Nov 14
Posts: 45
Credit: 54,184,495
RAC: 117,647

Message 62850 - Posted: 16 Dec 2014, 0:26:22 UTC - in response to Message 62848.

Hi Tom,
Yup, trouble is, 'every' task of that type fails the same way.. I cannot see any point in processing them, if they are all going to fail.
All they do is utilise my GPU and block other WU that can be processed.

I'll continue to process the modfitts until the others are sorted out.

Thanks for your assistance.

Regards,
Cliff

Jake Weiss
Volunteer moderator
Project developer
Project tester
Project scientist
Send message
Joined: 25 Feb 13
Posts: 440
Credit: 10,409,099
RAC: 203,541

Message 62852 - Posted: 16 Dec 2014, 16:19:53 UTC
Last modified: 23 Dec 2014, 17:38:52 UTC

Hey Guys,

Jeff put up some new runs with a bad parameter file. He said they are are fixed now so it should be good to run them now. There may still be a couple in the crunch queue but that should clear itself soon.

Sorry for the trouble.

Jake


[Edit]

The "Slow" runs are still running because it seems I am still getting good results from them. I would hate to kill them before they were done. Sorry for the delay, but these runs are still being useful and will be run a little while longer.

Enjoy your holidays

[\Edit]

hans dorn
Send message
Joined: 6 Apr 13
Posts: 8
Credit: 215,331,753
RAC: 231

Message 62914 - Posted: 30 Dec 2014, 20:33:39 UTC
Last modified: 30 Dec 2014, 20:34:38 UTC

I frequently run out of work for short periods of time when there's a high percentage of "fast" WUs.

My GPU goes through 2 of them in about 20 seconds.

Would it be possible to increase the amount of work done in the fast WUs, so that the required amount of communication with the server drops?

10esseeTony
Send message
Joined: 31 Aug 11
Posts: 19
Credit: 367,069,952
RAC: 49,256

Message 62922 - Posted: 1 Jan 2015, 14:36:31 UTC

My 7970 graphics cards can each tear through a "fast" workunit in 8-12 seconds, but then there is a 2-6 second pause while BOINC does whatever it does between workunits before the next one starts. This is highly inefficient from a processing point of view. Even my poor 6 year old Radeon HD4650 can crunch these units quickly (70 seconds).

WE NEED BIGGER WORKUNITS FOR GPU's.

WE NEED BIGGER WORKUNITS FOR GPU's.

WE NEED BIGGER WORKUNITS FOR GPU's.

That being said, de_nbody_08_05 has tied up 7 CPU cores for 37 hours. 8(

10esseeTony
Send message
Joined: 31 Aug 11
Posts: 19
Credit: 367,069,952
RAC: 49,256

Message 62923 - Posted: 1 Jan 2015, 14:44:20 UTC - in response to Message 62914.

I frequently run out of work for short periods of time when there's a high percentage of "fast" WUs.

My GPU goes through 2 of them in about 20 seconds.

Would it be possible to increase the amount of work done in the fast WUs, so that the required amount of communication with the server drops?



You may need to update/increase the "Minimum Work Buffer" under the Tools/Computing Preferences/Network settings tab. Add .5 to 1 day to whatever it currently is set at.

Profile Pooh Bear 27
Avatar
Send message
Joined: 17 Apr 10
Posts: 14
Credit: 5,122,519
RAC: 0

Message 62924 - Posted: 1 Jan 2015, 15:37:14 UTC

Problem with adding buffer, it adds buffer for CPU and GPU and you cannot separate. Since I want my CPU not to buffer the work it is doing, the short units are causing a lot of non-work time because of the time it takes to process the finishing and downloading the next unit.

10esseeTony
Send message
Joined: 31 Aug 11
Posts: 19
Credit: 367,069,952
RAC: 49,256

Message 62925 - Posted: 1 Jan 2015, 17:01:40 UTC - in response to Message 62831.

....the runs might not fully utilize your GPU. ...try configuring your GPU to run 2 of these at the same time...

Jake


Uhm, is there a tutorial for that?

swiftmallard
Avatar
Send message
Joined: 18 Jul 09
Posts: 289
Credit: 302,980,648
RAC: 0

Message 62927 - Posted: 1 Jan 2015, 21:42:26 UTC - in response to Message 62925.

....the runs might not fully utilize your GPU. ...try configuring your GPU to run 2 of these at the same time...

Jake


Uhm, is there a tutorial for that?

Copy this in Notepad (not Word):


<app_config>
<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
<app>
<name>milkyway_separation__modified_fit</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
</app_config>


Save the file as "app-config.xml" and place the file here:
(Unhide your folders if you haven't already)
Program data\Boinc\Projects\MilkyWay
Restart Boinc and you'll be running two at a time.

Profile mmstick
Avatar
Send message
Joined: 23 Nov 09
Posts: 29
Credit: 17,119,258
RAC: 0

Message 62931 - Posted: 2 Jan 2015, 13:54:58 UTC - in response to Message 62927.

....the runs might not fully utilize your GPU. ...try configuring your GPU to run 2 of these at the same time...

Jake


Uhm, is there a tutorial for that?

Copy this in Notepad (not Word):


<app_config>
<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
<app>
<name>milkyway_separation__modified_fit</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
</app_config>


Save the file as "app-config.xml" and place the file here:
(Unhide your folders if you haven't already)
Program data\Boinc\Projects\MilkyWay
Restart Boinc and you'll be running two at a time.


Is this app not available on Linux? I simply get an error that it cannot find an app by that name, so it only discovers the normal milkyway app.

Profile arkayn
Avatar
Send message
Joined: 14 Feb 09
Posts: 999
Credit: 74,932,619
RAC: 0

Message 62935 - Posted: 2 Jan 2015, 17:38:26 UTC - in response to Message 62931.

....the runs might not fully utilize your GPU. ...try configuring your GPU to run 2 of these at the same time...

Jake


Uhm, is there a tutorial for that?

Copy this in Notepad (not Word):


<app_config>
<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
<app>
<name>milkyway_separation__modified_fit</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
</app_config>


Save the file as "app-config.xml" and place the file here:
(Unhide your folders if you haven't already)
Program data\Boinc\Projects\MilkyWay
Restart Boinc and you'll be running two at a time.


Is this app not available on Linux? I simply get an error that it cannot find an app by that name, so it only discovers the normal milkyway app.


Make certain that you have mod fit checked in your preferences

CPU: http://milkyway.cs.rpi.edu/milkyway/download/milkyway_separation__modified_fit_1.36_x86_64-pc-linux-gnu
AMD: http://milkyway.cs.rpi.edu/milkyway/download/milkyway_separation__modified_fit_1.36_x86_64-pc-linux-gnu__opencl_ati_101
____________

10esseeTony
Send message
Joined: 31 Aug 11
Posts: 19
Credit: 367,069,952
RAC: 49,256

Message 62938 - Posted: 3 Jan 2015, 4:52:28 UTC - in response to Message 62927.

....the runs might not fully utilize your GPU. ...try configuring your GPU to run 2 of these at the same time...

Jake


Uhm, is there a tutorial for that?

Copy this in Notepad (not Word):


<app_config>
<app>
<name>milkyway</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
<app>
<name>milkyway_separation__modified_fit</name>
<gpu_versions>
<gpu_usage>0.5</gpu_usage>
<cpu_usage>0.05</cpu_usage>
</gpu_versions>
</app>
</app_config>


Save the file as "app-config.xml" and place the file here:
(Unhide your folders if you haven't already)
Program data\Boinc\Projects\MilkyWay
Restart Boinc and you'll be running two at a time.


I'm afraid that did nothing on neither my dual GPU machine nor my single GPU box. :( Not even after triple checking I'd followed the instructions.

So, having successfully (but unintentionally) 'hijacked' a thread, I'd like to try and return the focus to these fast work units.

These fast units, due to the standard GUI interface issues of the BOINC client, don't keep the GPU's busy enough. I've dropped from almost a million credits per 24 hours to 650k a day. I don't give a crap about the credits, the point is that these efficient fast WU's have resulted in almost a third less processing power, apparently.

Again, even my lowly 6 year old ATI HD4850 can crunch these WU's very quickly. And if it's possible from a computer science perspective, I think the project would better benefit from a denser, more prolonged set of WU's, that help to minimize/offset the 2-6 seconds of 'downtime' that occurs between WU's.

1 · 2 · Next
Post to thread

Message boards : News : New Modfit Runs


Main page · Your account · Message boards


Copyright © 2017 AstroInformatics Group