Welcome to MilkyWay@home

Very Strange Time To Competion

Message boards : Number crunching : Very Strange Time To Competion
Message board moderation

To post messages, you must log in.

Previous · 1 · 2

AuthorMessage
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
Message 12142 - Posted: 21 Feb 2009, 20:01:54 UTC - in response to Message 12079.  

I'm noticing some really long wu's on my older machines but not many. I posted the details of one here:

http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=642&nowrap=true#12070

but it seems like it maybe just a random super long wu


The WUs aren't doing anything nondeterministic, so they all do the same amount of work (for a particular stripe).

If there are some out there that are taking 2x as long or so, this might be a problem with the alpha == delta == 1 optimization, some older machines might not be able to do the comparison as effectively. Other than that I honestly don't see what would make a WU take twice as long, unless there's something function going on with the processor or other running processes on that machine.

I think I'm going to update the code to do an initial check to see if alpha == delta == 1 with some low tolerance instead and see if this fixed the problem.

Travis, have you read my post in the other "unusual time to completion" thread?
I've tested it. The problem is not that alpha==delta==1 condition. It is really the underflow condition at the stream parameter bounds for the particle swarm searches.
ID: 12142 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Travis
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 30 Aug 07
Posts: 2046
Credit: 26,480
RAC: 0
Message 12147 - Posted: 21 Feb 2009, 20:15:02 UTC - in response to Message 12142.  

I'm noticing some really long wu's on my older machines but not many. I posted the details of one here:

http://milkyway.cs.rpi.edu/milkyway/forum_thread.php?id=642&nowrap=true#12070

but it seems like it maybe just a random super long wu


The WUs aren't doing anything nondeterministic, so they all do the same amount of work (for a particular stripe).

If there are some out there that are taking 2x as long or so, this might be a problem with the alpha == delta == 1 optimization, some older machines might not be able to do the comparison as effectively. Other than that I honestly don't see what would make a WU take twice as long, unless there's something function going on with the processor or other running processes on that machine.

I think I'm going to update the code to do an initial check to see if alpha == delta == 1 with some low tolerance instead and see if this fixed the problem.

Travis, have you read my post in the other "unusual time to completion" thread?
I've tested it. The problem is not that alpha==delta==1 condition. It is really the underflow condition at the stream parameter bounds for the particle swarm searches.



It seems like people are having some other problem though. The one you were talking about was a much smaller increase in runtime.
ID: 12147 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile speedimic
Avatar

Send message
Joined: 22 Feb 08
Posts: 260
Credit: 57,387,048
RAC: 0
Message 12150 - Posted: 21 Feb 2009, 20:28:15 UTC

Do those evercrunch WUs only jump on Windows machines?
mic.


ID: 12150 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
Message 12155 - Posted: 21 Feb 2009, 20:36:18 UTC - in response to Message 12150.  
Last modified: 21 Feb 2009, 20:36:58 UTC

Do those evercrunch WUs only jump on Windows machines?

As it is most likely a problem specific to thoses machines, one can't say that. I've never experienced such a WU on my computers (all running Windows).

The somewhat smaller runtime deviations related to the underflows at the bounds of some stream parameters are visisble on all types of machines and also with the stock application.
ID: 12155 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile speedimic
Avatar

Send message
Joined: 22 Feb 08
Posts: 260
Credit: 57,387,048
RAC: 0
Message 12170 - Posted: 21 Feb 2009, 21:14:29 UTC - in response to Message 12155.  
Last modified: 21 Feb 2009, 21:14:37 UTC

Do those evercrunch WUs only jump on Windows machines?

As it is most likely a problem specific to thoses machines, one can't say that. I've never experienced such a WU on my computers (all running Windows).

The somewhat smaller runtime deviations related to the underflows at the bounds of some stream parameters are visisble on all types of machines and also with the stock application.


I was just wondering, cause I haven't seen any on my machines...
mic.


ID: 12170 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
XB-STX

Send message
Joined: 9 Aug 08
Posts: 18
Credit: 56,863,533
RAC: 0
Message 12185 - Posted: 21 Feb 2009, 21:36:25 UTC - in response to Message 12170.  

Several additional interesting observations.

1) Variations in WU run times (by stripe) are greatest on my Intel-based machines. The AMD-based machines are all within a couple percentages points of completion time; the Windows based machines (Q6600 Vista 32, P4 XP Pro) can sometimes vary by more than two to three-fold.
2) Problems have been most pronounced with the opti-apps; last two versions I have run were Gispel's. When I detached from the project, and re-attached, and ran a couple dozen off the stock app, the run times (by stripe) again were within a couple percentage points of each other (Q6600).
3) Later, re-loaded v0.19 opti app (Gispel), fresh download, and again, variations between run times went way up again (Q6600).
4) Further, time to read the parameter files on the stock app were <1 second, however, back under the opti app, they varied wildly - sometimes to over 90 seconds, which is ridiculous (Q6600).

Again, these are merely observations, and in no way should be considered judgements. We all know that the unique combination of HW + OS + SW can often give different results for some. In other words, your mileage may vary.

Perhaps due to the continuing frustration of all of the above, I decided to give the GPU app a try. Still have a couple of kinks to work out, but judging by the stability and consistency (and, to be honest, speed of credit!), I'm gonna put my frustrations aside, and stay with the GPU app instead.

Now, if only we could get a CUDA version so I can run it on three of my other boxes!

Rgds
XB
ID: 12185 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Cluster Physik

Send message
Joined: 26 Jul 08
Posts: 627
Credit: 94,940,203
RAC: 0
Message 12193 - Posted: 21 Feb 2009, 21:45:30 UTC - in response to Message 12185.  

Perhaps due to the continuing frustration of all of the above, I decided to give the GPU app a try. Still have a couple of kinks to work out, but judging by the stability and consistency (and, to be honest, speed of credit!), I'm gonna put my frustrations aside, and stay with the GPU app instead.

Ater some errors in the beginning, it appears to be running quite well now on your Q6600 and the HD38?0, isn't it?
ID: 12193 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
XB-STX

Send message
Joined: 9 Aug 08
Posts: 18
Credit: 56,863,533
RAC: 0
Message 12216 - Posted: 21 Feb 2009, 22:18:14 UTC - in response to Message 12193.  

That it is, CP, that it is. :)

XB
ID: 12216 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile Paul D. Buck

Send message
Joined: 12 Apr 08
Posts: 621
Credit: 161,934,067
RAC: 0
Message 12228 - Posted: 21 Feb 2009, 23:02:22 UTC - in response to Message 12193.  

Perhaps due to the continuing frustration of all of the above, I decided to give the GPU app a try. Still have a couple of kinks to work out, but judging by the stability and consistency (and, to be honest, speed of credit!), I'm gonna put my frustrations aside, and stay with the GPU app instead.

Ater some errors in the beginning, it appears to be running quite well now on your Q6600 and the HD38?0, isn't it?


Well, Dual Xeon and HD4870 after 22 hours seems to be running Ok ... still have not figured out how to keep it running one task at all times. I get some pauses at times...

Color me reasonably happy for the moment ...

Has anyone tried the last version of BOINC out to see if it picks up the ATI cards? I sent them the notes on GPU detection but have not seen any feedback on it ...
ID: 12228 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Profile GalaxyIce
Avatar

Send message
Joined: 6 Apr 08
Posts: 2018
Credit: 100,142,856
RAC: 0
Message 12233 - Posted: 21 Feb 2009, 23:29:44 UTC - in response to Message 12228.  
Last modified: 21 Feb 2009, 23:30:22 UTC

Has anyone tried the last version of BOINC out to see if it picks up the ATI cards? I sent them the notes on GPU detection but have not seen any feedback on it ...

Yes I have and I don't think so. I installed BOINC 6.4.5 after installing a HD4850 ATI card today;

21/02/2009 23:24:27||Starting BOINC client version 6.4.5 for windows_intelx86
21/02/2009 23:24:27||log flags: task, file_xfer, sched_ops
21/02/2009 23:24:27||Libraries: libcurl/7.19.0 OpenSSL/0.9.8i zlib/1.2.3
21/02/2009 23:24:27||Data directory: C:\boinc-data
21/02/2009 23:24:27||Running under account ix
21/02/2009 23:24:27|Milkyway@home|Found app_info.xml; using anonymous platform
21/02/2009 23:24:28||Processor: 2 GenuineIntel Intel(R) Pentium(R) 4 CPU 3.20GHz [x86 Family 15 Model 4 Stepping 3]
21/02/2009 23:24:28||Processor features: fpu tsc pae nx sse sse2 mmx
21/02/2009 23:24:28||OS: Microsoft Windows XP: Professional x86 Editon, Service Pack 3, (05.01.2600.00)
21/02/2009 23:24:28||Memory: 2.00 GB physical, 7.85 GB virtual
21/02/2009 23:24:28||Disk: 74.53 GB total, 29.55 GB free
21/02/2009 23:24:28||Local time is UTC +0 hours
21/02/2009 23:24:28||Not using a proxy
21/02/2009 23:24:29||No CUDA devices found
21/02/2009 23:24:29||No coprocessors
21/02/2009 23:24:29|Milkyway@home|URL: http://milkyway.cs.rpi.edu/milkyway/;


ID: 12233 · Rating: 0 · rate: Rate + / Rate - Report as offensive     Reply Quote
Previous · 1 · 2

Message boards : Number crunching : Very Strange Time To Competion

©2024 Astroinformatics Group