Advanced search

Message boards : Graphics cards (GPUs) : Application 6.44 for Linux

Author Message
Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1957
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 2260 - Posted: 11 Sep 2008 | 14:24:06 UTC
Last modified: 11 Sep 2008 | 14:27:07 UTC

A new application has been uploaded:
1) this application has drastically reduced CPU-usage
2) it reports the time per step in output, as now the cputime is not anymore a good estimate of elapsed time.

We are still requiring one CPU per WU until the new boinc client 6.3.11 is out.
Windows equivalent coming up.
gdf

FeuerKater
Avatar
Send message
Joined: 27 Aug 08
Posts: 9
Credit: 900,439
RAC: 0
Level
Gly
Scientific publications
watwat
Message 2261 - Posted: 11 Sep 2008 | 14:31:50 UTC - in response to Message 2260.
Last modified: 11 Sep 2008 | 15:21:52 UTC

A new application has been uploaded:
1) this application has drastically reduced CPU-usage
2) it reports the time per step in output, as now the cputime is not anymore a good estimate of elapsed time.


... wow ... great work ... now only the the new Boinc manager (4 Windows) and we are satisfied so we can use the GPU on PS3Grid an the CPU on other Boinc Projects :)
____________

Profile [SETI.USA]Tank_Master
Avatar
Send message
Joined: 8 Jul 07
Posts: 85
Credit: 67,463,387
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 2263 - Posted: 11 Sep 2008 | 15:52:45 UTC

unless I am mistaken, we need the new client for BOTH Linux and Windows to not need the CPU to run a WU....

FeuerKater
Avatar
Send message
Joined: 27 Aug 08
Posts: 9
Credit: 900,439
RAC: 0
Level
Gly
Scientific publications
watwat
Message 2264 - Posted: 11 Sep 2008 | 15:56:21 UTC - in response to Message 2263.

unless I am mistaken, we need the new client for BOTH Linux and Windows to not need the CPU to run a WU....


allright, when comes the new Client ? have you an plan then this is available ?

____________

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 2267 - Posted: 11 Sep 2008 | 17:51:18 UTC

The responsible guy is on holiday. I'd say Expect something between 1 and 3 weeks.

@Linux users: does the new run time correspond with the time reported previously?

MrS
____________
Scanning for our furry friends since Jan 2002

Profile koschi
Avatar
Send message
Joined: 14 Aug 08
Posts: 124
Credit: 792,979,198
RAC: 11,592
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 2272 - Posted: 11 Sep 2008 | 18:20:43 UTC

What do yo mean? The time that is shown in BOINC manager?

It is not. Right now the process is consuming only 7-8% CPU load on my C2D E7200@3,4GHz and the process took almost 5 minutes of CPU time up to now. But the progress is already at 8.x% That would give a calculation time of roughly one hour. While my CPU time was around 12hours for one unit before, this fits together.

Nice job GDF! Now we are only lacking BOINC 6.3.11, which lets us use the spare 90% of the CPU time. Woohoo, I'm so much looking forward to it =)

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 2275 - Posted: 11 Sep 2008 | 19:06:16 UTC - in response to Message 2272.

What do yo mean?


Let's say before you needed 12h per WU. I just wanted to know if the new time agrees with the old one, i.e. also 12h.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile koschi
Avatar
Send message
Joined: 14 Aug 08
Posts: 124
Credit: 792,979,198
RAC: 11,592
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 2277 - Posted: 11 Sep 2008 | 19:18:24 UTC

On the clock at my wall it will still take 12 hours, but when the process ends it has collected only ~1hour of CPU time. I think the same time will be displayed in each WUs listing on the web page.

This should be the reason for the new feature 2 which GDF describes...

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 2279 - Posted: 11 Sep 2008 | 19:26:17 UTC - in response to Message 2277.

This should be the reason for the new feature 2 which GDF describes...


Sure. I asked because I wanted to get a feeling for how inaccurate the previous method was. So far it seems like it was OK.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile koschi
Avatar
Send message
Joined: 14 Aug 08
Posts: 124
Credit: 792,979,198
RAC: 11,592
Level
Glu
Scientific publications
watwatwatwatwatwatwatwatwatwatwat
Message 2302 - Posted: 13 Sep 2008 | 5:38:46 UTC - in response to Message 2260.

2) it reports the time per step in output, as now the cputime is not anymore a good estimate of elapsed time.



stderr out now tells me:
Time per step: 53.181 ms
Is there any way I can calculate the wall time from this, lets say for credit calculation or stuff like that?

Profile Bender10
Avatar
Send message
Joined: 3 Dec 07
Posts: 167
Credit: 8,368,897
RAC: 0
Level
Ser
Scientific publications
watwatwatwatwatwatwat
Message 2346 - Posted: 15 Sep 2008 | 11:32:41 UTC


Linux box results:

Boinc 6.3.8
177.13 driver
Ubuntu 64
8800GS no OC

http://www.ps3grid.net/result.php?resultid=56032

____________


Consciousness: That annoying time between naps......

Experience is a wonderful thing: it enables you to recognize a mistake every time you repeat it.

Profile Stefan Ledwina
Avatar
Send message
Joined: 16 Jul 07
Posts: 464
Credit: 240,924,370
RAC: 4,586,911
Level
Leu
Scientific publications
watwatwatwatwatwatwatwat
Message 2421 - Posted: 17 Sep 2008 | 14:55:24 UTC

Like other users reported it for the Windows app, I see pretty different CPU times on my two Linux 64 boxes.

The one with CentOS 5.2, and the 9800GTX uses about 2550 CPU seconds per WU. That's pretty the same CPU usage on every WU.

The other one with Ubuntu 7.10 and a 8800GT is using between 99 and 55 CPU seconds per WU which is a really big difference compared to the CentOS host...

It doesn't really fit in the Linux thread, but on the Vista computer I noticed that the application is also using approximately 2500 CPU seconds per WU on a GTX 260.

So why is the Ubuntu box using much less of the CPU than the CentOS and Vista box?
Both Linux computers are full time crunchers without keyboard, mouse and monitor attached and are not used for anything else than crunching, and both use the same drivers (177.67 for Linux64). Both Linux boxes have a Q6600 CPU, both sligtly ocerclocked to 2.6 Ghz, both have DDR2 800 RAM, the only difference is that the CentOS box (the one which uses more CPU time) has 4 GB RAM and the other one only 2 GB.
____________

pixelicious.at - my little photoblog

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1957
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 2423 - Posted: 17 Sep 2008 | 16:27:56 UTC - in response to Message 2421.

I don't know why the CPU time is different in different system, but it is not so important. More important is if the new application is as fast as the old one on these boxes.

I think that this is difficult to see by looking at the cpu time of past workunits. Only the time per step is a reliable estimate of the elapsed time. Unfortunately, we did not have this on previous application.

gdf

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1957
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 2426 - Posted: 17 Sep 2008 | 17:18:27 UTC - in response to Message 2423.
Last modified: 17 Sep 2008 | 17:19:26 UTC

I don't know why the CPU time is different in different system, but it is not so important. More important is if the new application is as fast as the old one on these boxes.

I think that this is difficult to see by looking at the cpu time of past workunits. Only the time per step is a reliable estimate of the elapsed time. Unfortunately, we did not have this on previous application.

gdf



On my Linux box
# Device 0: "GeForce 8800 GT"
# Clock rate: 1512000 kilohertz
MDIO ERROR: cannot open file "restart.coor"
# Time per step: 70.266 ms

So, the new application takes exactly as before, as I knew the performance before. We will test on Windows vista in the next few days.

CPU time is 3404.313, but this has little meaning.

gdf

gdf

Profile Venturini Dario[VENETO]
Send message
Joined: 26 Jul 08
Posts: 44
Credit: 4,832,360
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwatwatwat
Message 2427 - Posted: 17 Sep 2008 | 17:55:55 UTC

My Linux box with 8800GT reports more or less the same time per step that yours do.

Profile [SETI.USA]Tank_Master
Avatar
Send message
Joined: 8 Jul 07
Posts: 85
Credit: 67,463,387
RAC: 0
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwat
Message 2428 - Posted: 17 Sep 2008 | 18:04:43 UTC

I have a server 2008 x64 box if you need some help testing.

Profile Venturini Dario[VENETO]
Send message
Joined: 26 Jul 08
Posts: 44
Credit: 4,832,360
RAC: 0
Level
Ala
Scientific publications
watwatwatwatwatwatwatwat
Message 2430 - Posted: 17 Sep 2008 | 18:43:10 UTC - in response to Message 2426.

On my Linux box
# Device 0: "GeForce 8800 GT"
# Clock rate: 1512000 kilohertz
MDIO ERROR: cannot open file "restart.coor"
# Time per step: 70.266 ms

So, the new application takes exactly as before, as I knew the performance before. We will test on Windows vista in the next few days.


Oh btw I forgot to say that, even if you report that it takes exactly the same as before, since the new application my GPU has run a lot cooler, let's say 4-5 degrees less (from 57° to 52°)

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 2431 - Posted: 17 Sep 2008 | 19:45:45 UTC - in response to Message 2430.
Last modified: 17 Sep 2008 | 19:47:20 UTC

since the new application my GPU has run a lot cooler, let's say 4-5 degrees less (from 57° to 52°)


That's what we observe under windows as well.

I took a look at your 4 results with 6.44: they took on average 61500s whereas previously you needed about 55700s. EDIT: that's a 10% performance hit. Under win we're dealing with ~23%.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1957
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 2439 - Posted: 18 Sep 2008 | 8:59:11 UTC - in response to Message 2431.

since the new application my GPU has run a lot cooler, let's say 4-5 degrees less (from 57° to 52°)


That's what we observe under windows as well.

I took a look at your 4 results with 6.44: they took on average 61500s whereas previously you needed about 55700s. EDIT: that's a 10% performance hit. Under win we're dealing with ~23%.

MrS


You must be comparing previous cpu time with current elapsed time. As the cpu was really doing nothing even before, it is likely that 1 h cpu time corresponds to 1.1 hours elapsed time.

GDF

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 2442 - Posted: 18 Sep 2008 | 19:04:17 UTC - in response to Message 2439.
Last modified: 18 Sep 2008 | 19:04:40 UTC

You must be comparing previous cpu time with current elapsed time.


Yes, that's what I'm doing.

As the cpu was really doing nothing even before, it is likely that 1 h cpu time corresponds to 1.1 hours elapsed time.


That's what I tried to find out in the beginning of the thread by asking. Didn't work very well, though.

I just took the time to chase one of my typical WUs through my log file. The 6.43-WU took 44085s of wall clock time and reported a CPU usage of 43703s. So the difference is not that big. My conclusion: comparing previous CPU-time with current elapsed time is not accurate, but if the machine was undisturbed the error can be <1%.

MrS
____________
Scanning for our furry friends since Jan 2002

Profile Kokomiko
Avatar
Send message
Joined: 18 Jul 08
Posts: 190
Credit: 24,093,690
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 2452 - Posted: 19 Sep 2008 | 8:31:27 UTC - in response to Message 2439.
Last modified: 19 Sep 2008 | 8:51:16 UTC



You must be comparing previous cpu time with current elapsed time. As the cpu was really doing nothing even before, it is likely that 1 h cpu time corresponds to 1.1 hours elapsed time.

GDF


Under Windows I have much less using CPU time with 0:43:38h CPU time and 9:26:48h elapsed time with the 6.45.

WU with 6.45

That's 1 hour CPU time to 12.99 elapsed time (or GPU-time).

With the 6.43 the CPU = elapsed time was 7:50h, so the WUs need much longer with the 6.45 by less CPU usage.

WU with 6.43

And btw: I use the free time for another task through the cc_config.xml and the ncpus-entry. Give one more than exist and 5 tasks are running on a quad.

____________

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1957
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 2453 - Posted: 19 Sep 2008 | 8:59:53 UTC - in response to Message 2452.



You must be comparing previous cpu time with current elapsed time. As the cpu was really doing nothing even before, it is likely that 1 h cpu time corresponds to 1.1 hours elapsed time.

GDF


Under Windows I have much less using CPU time with 0:43:38h CPU time and 9:26:48h elapsed time with the 6.45.

WU with 6.45

That's 1 hour CPU time to 12.99 elapsed time (or GPU-time).

With the 6.43 the CPU = elapsed time was 7:50h, so the WUs need much longer with the 6.45 by less CPU usage.

WU with 6.43

And btw: I use the free time for another task through the cc_config.xml and the ncpus-entry. Give one more than exist and 5 tasks are running on a quad.



It is possible that faster cards as yours suffers more with the new app. Could you try to run with just 3 other jobs to see if the GPU time reduces?

gdf

Profile Kokomiko
Avatar
Send message
Joined: 18 Jul 08
Posts: 190
Credit: 24,093,690
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwatwatwatwatwat
Message 2460 - Posted: 19 Sep 2008 | 12:05:35 UTC - in response to Message 2453.
Last modified: 19 Sep 2008 | 12:09:27 UTC


It is possible that faster cards as yours suffers more with the new app. Could you try to run with just 3 other jobs to see if the GPU time reduces?

gdf


The first 3 WUs with the 6.45 I've crunched without the change in the cc_config.xml. The CPU time was 2618.57 (34008.726 elapsed time) and 3084.76 (35277.078 elapsed time) with ncpus=5 and 3484.17 (36270.111 elapsed time), 3082.61 (39698.400 elapsed time) and 3681.14 (30449.058 elapsed time) without.

You can find this here, TaskIDs 59992, 60120 and 60200 without ncpus-entry and 60334 and 60539 with ncpus=5.
____________

Profile GDF
Volunteer moderator
Project administrator
Project developer
Project tester
Volunteer developer
Volunteer tester
Project scientist
Send message
Joined: 14 Mar 07
Posts: 1957
Credit: 629,356
RAC: 0
Level
Gly
Scientific publications
watwatwatwatwat
Message 2470 - Posted: 19 Sep 2008 | 15:33:45 UTC - in response to Message 2460.


It is possible that faster cards as yours suffers more with the new app. Could you try to run with just 3 other jobs to see if the GPU time reduces?

gdf


The first 3 WUs with the 6.45 I've crunched without the change in the cc_config.xml. The CPU time was 2618.57 (34008.726 elapsed time) and 3084.76 (35277.078 elapsed time) with ncpus=5 and 3484.17 (36270.111 elapsed time), 3082.61 (39698.400 elapsed time) and 3681.14 (30449.058 elapsed time) without.

You can find this here, TaskIDs 59992, 60120 and 60200 without ncpus-entry and 60334 and 60539 with ncpus=5.


I was referring to the GPU time. Anyway, your is approx 40 ms/step. Independently on how many cpu jobs you are running, which is good.

gdf

Post to thread

Message boards : Graphics cards (GPUs) : Application 6.44 for Linux

//