Advanced search

Message boards : Number crunching : swan_sync = 0 ... little effect.

Author Message
Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 251
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 21502 - Posted: 18 Jun 2011 | 19:03:35 UTC
Last modified: 18 Jun 2011 | 19:05:31 UTC

Hello: A view of the results of using the variable "swan_sync = 0" for months, I think it is not interesting.

Platform Linux-64bits Ubuntu 11.04 -i7-930-6GB of Ram. GTX295, tasks "acemdlong."

The yield on the GPU is increased by 10%. on average.
The CPU load is increased by 400%, over temperature, fan noise and expense.

In Windows7-64Bits (and the same hardaware than on Linux) the results are equal, with the aggravating circumstance that the overall performance of Windows GPUGRID is 15% lower on average. Regards

Snow Crash
Send message
Joined: 4 Apr 09
Posts: 450
Credit: 539,316,349
RAC: 0
Level
Lys
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 21504 - Posted: 18 Jun 2011 | 19:20:56 UTC - in response to Message 21502.

SWAN_SYNC was created to increase the efficiency on fermi cards, and it does.
It was not intended to be used on the older, but still great, 2xx series.
____________
Thanks - Steve

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 21505 - Posted: 18 Jun 2011 | 19:34:35 UTC - in response to Message 21504.

On Linux SWAN_SYNC=0 must be in CAPITALS!

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 251
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 21507 - Posted: 18 Jun 2011 | 19:57:16 UTC - in response to Message 21505.
Last modified: 18 Jun 2011 | 20:00:45 UTC

On Linux SWAN_SYNC=0 must be in CAPITALS!


Hi, The variable on Linux and is capitalized and it works perfectly.

I just say the results do not offset (in GTX295) and show me what my numbers after several months of work.

Please who believes otherwise simply showing their numbers ... comparing the performance increase (reduction of time on a task 10%) with the increased load on the CPU, which is four times with all that entails. Greetings.

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 21516 - Posted: 19 Jun 2011 | 8:26:10 UTC - in response to Message 21507.

Some people might think a 10% improvement is worthwhile, others will think not. As Steve said SWAN_SYNC works best on Fermi's. It does still bring some improvement on other cards, but usually less.

On Linux you are about 5% faster than under W7, just going by the one task:

4082104 2531823 15 Jun 2011 14:07:36 UTC 16 Jun 2011 10:50:27 UTC Completed and validated 39,739.04 3,011.96 35,067.36 52,601.04 Long runs (8-12 hours on fastest card) v6.14 (cuda31) Linux

4059957 2515436 6 Jun 2011 13:46:10 UTC 7 Jun 2011 11:03:34 UTC Completed and validated 43,268.86 42,590.44 35,067.36 52,601.04 Long runs (8-12 hours on fastest card) v6.13 (cuda31) W7

Both TONI_AGGsoup tasks with the same credit.

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 251
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 21521 - Posted: 19 Jun 2011 | 14:15:17 UTC - in response to Message 21516.
Last modified: 19 Jun 2011 | 14:30:57 UTC

Some people might think a 10% improvement is worthwhile, others will think not. As Steve said SWAN_SYNC works best on Fermi's. It does still bring some improvement on other cards, but usually less.


Hello: The comment is only to be valued if winning that 10% (one hour) really pays off when one considers the large increase in load on the CPU which means assigning a CPU + GPU core SWAN_SYNC = 0, as we is 10 or 14 times, as we compare Linux and Windows.


On Linux you are about 5% faster than under W7, just going by the one task:

4082104 2531823 15 Jun 2011 14:07:36 UTC 16 Jun 2011 10:50:27 UTC Completed and validated 39,739.04 3,011.96 35,067.36 52,601.04 Long runs (8-12 hours on fastest card) v6.14 (cuda31) Linux

4059957 2515436 6 Jun 2011 13:46:10 UTC 7 Jun 2011 11:03:34 UTC Completed and validated 43,268.86 42,590.44 35,067.36 52,601.04 Long runs (8-12 hours on fastest card) v6.13 (cuda31) W7

Both TONI_AGGsoup tasks with the same credit.


Taking the numbers of these tasks I have done recently, I play the following, correct if I'm wrong.

The task in Linux I see it is 8.9% faster (or slower Windows 8.15%) almost 10% not 5% but this task in Linux also have not enabled SWAN_SYNC = 0. it shows how little CPU time. If selected would be 15% faster Linux than Windows.

In short we are talking about one hour less in Linux to complete the task and enabled again without SWAN_SYNC = 0 so with a load on the CPU much lower.

The task in Windows is enabled SWAN_SYNC = 0 and the CPU load was 14 times greater than that needed in Linux, frankly very bad deal if you also 8.15% has been slower.

Comparing CPU time Linux-Linux and Windows-Windows, the difference is 10 times as I say in the remarks, with or without SWAN_SYNC = 0.

Remember that these tasks are the same hardware. Greetings.

Profile Beyond
Avatar
Send message
Joined: 23 Nov 08
Posts: 1112
Credit: 6,162,416,256
RAC: 0
Level
Tyr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 21554 - Posted: 29 Jun 2011 | 13:38:27 UTC - in response to Message 21521.

The task in Linux I see it is 8.9% faster (or slower Windows 8.15%) almost 10% not 5% but this task in Linux also have not enabled SWAN_SYNC = 0. it shows how little CPU time. If selected would be 15% faster Linux than Windows.

In short we are talking about one hour less in Linux to complete the task and enabled again without SWAN_SYNC = 0 so with a load on the CPU much lower.

The task in Windows is enabled SWAN_SYNC = 0 and the CPU load was 14 times greater than that needed in Linux, frankly very bad deal if you also 8.15% has been slower.

Comparing CPU time Linux-Linux and Windows-Windows, the difference is 10 times as I say in the remarks, with or without SWAN_SYNC = 0.

Remember that these tasks are the same hardware. Greetings.

One thing to keep in mind. You're comparing to Win7 which is very slow here. XP is considerably faster.

Profile Carlesa25
Avatar
Send message
Joined: 13 Nov 10
Posts: 328
Credit: 72,619,453
RAC: 251
Level
Thr
Scientific publications
watwatwatwatwatwatwatwatwatwatwatwatwatwatwatwatwat
Message 21555 - Posted: 29 Jun 2011 | 14:09:47 UTC - in response to Message 21554.
Last modified: 29 Jun 2011 | 14:11:01 UTC


One thing to keep in mind. You're comparing to Win7 which is very slow here. XP is considerably faster.


Hi. True XP is faster than Win7 (but not as much as Linux), but this is not the issue.

What we are discussing is whether to use the variable SWAN_SYNC = 0 is profitable or not by comparing the cost of CPU time and energy costs and temperatures.

My result is simply not worth it. Greetings.

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 21576 - Posted: 2 Jul 2011 | 20:04:15 UTC - in response to Message 21555.

CC 1.3 cards do not benefit as much as CC 2.0 GPU's.
As well as using SWAN_SYNC you MUST free up a CPU Core. If you do not also free up a CPU core you will not see much improvement.
To some extent the present Priority settings are obfuscating the normal situation; using a higher priority under some circumstances improves performance albeit at the expense of lag. Perhaps you are demonstrating this.

Profile Saenger
Avatar
Send message
Joined: 20 Jul 08
Posts: 134
Credit: 23,657,183
RAC: 0
Level
Pro
Scientific publications
watwatwatwatwatwat
Message 21761 - Posted: 28 Jul 2011 | 15:11:54 UTC

swan_sync is a giant waste of resources on a GT240.
Crunching gets slower and CPU-use up for absolutely no gain.
Recommending swan_sync to non-Fermis seems to be something completely rubbish.
____________
Gruesse vom Saenger

For questions about Boinc look in the BOINC-Wiki

Post to thread

Message boards : Number crunching : swan_sync = 0 ... little effect.

//