climateprediction.net home page
A performance oddity.

A performance oddity.

Message boards : Number crunching : A performance oddity.
Message board moderation

To post messages, you must log in.

AuthorMessage
Profile Dave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4529
Credit: 18,661,594
RAC: 14,529
Message 70637 - Posted: 11 Mar 2024, 12:00:07 UTC
Last modified: 11 Mar 2024, 12:02:45 UTC

And an oddity:

Running tasks from the same batches using WINE seems to be about 20% faster than running windows in VB. No great surprise there.
However with identical settings for memory use and number of CPUs in use etc. running tasks in Ubuntu24.04 in VB tasks are finishing noticeably faster than natively under Ubuntu 23.10 BOINC in both cases is 7.25.0 though the one in VB is a more recent incarnation compiled from master branch in git-hub. When more work from CPDN appears I will test for consistency in this across batches. With luck the main site tasks running under WINE and Windows in VB that I had paused while the Linux testing branch work was going will be finished.
Run time average
Average run time native: 345212.255
Average run time VB: 193168.265

Fastest task Native
11630 10938 6 Mar 2024, 8:11:26 UTC 11 Mar 2024, 5:21:34 UTC Completed 241,850.73 200,118.90 7,518.32
Slowest task VB
11635 10943 6 Mar 2024, 8:49:25 UTC 11 Mar 2024, 6:31:18 UTC Completed 200,549.47 199,507.80 7,518.32

CPU times for VB are all faster though the fastest native is close to the slowest VB one.
ID: 70637 · Report as offensive     Reply Quote
Glenn Carver

Send message
Joined: 29 Oct 17
Posts: 1044
Credit: 16,196,312
RAC: 12,647
Message 70638 - Posted: 11 Mar 2024, 13:09:29 UTC - in response to Message 70637.  
Last modified: 11 Mar 2024, 13:10:54 UTC

Couple of obvious questions. Is this all on the same physical hardware? Are the timings taken from a single task on a quiet machine with nothing else running?

I am surprised by a 20% difference between wine and Windows in a VM. Both have to use an emulation layer.

If it's not a quiet machine, processes might be switching between cores and flushing cache lines, affecting performance.

I have measured a factor of two in task completion time between a busy machine and a same quiet one.
ID: 70638 · Report as offensive     Reply Quote
Profile Dave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4529
Credit: 18,661,594
RAC: 14,529
Message 70639 - Posted: 11 Mar 2024, 14:42:53 UTC - in response to Message 70638.  
Last modified: 11 Mar 2024, 14:54:42 UTC

The same physical machine, with the tasks running concurrently. It is the machine I use for web browsing, email etc. but only rarely for anything more intensive than listening to audio on BBC sounds.

Edit: Over 20 hours a day nothing apart from BOINC running on the machine.
ID: 70639 · Report as offensive     Reply Quote
Glenn Carver

Send message
Joined: 29 Oct 17
Posts: 1044
Credit: 16,196,312
RAC: 12,647
Message 70640 - Posted: 11 Mar 2024, 14:49:01 UTC - in response to Message 70639.  

It's not possible to get reliable numbers on a busy machine. Too many processes competing for resources with different priorities.
ID: 70640 · Report as offensive     Reply Quote
Profile Dave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4529
Credit: 18,661,594
RAC: 14,529
Message 70641 - Posted: 11 Mar 2024, 14:59:23 UTC - in response to Message 70640.  

It's not possible to get reliable numbers on a busy machine. Too many processes competing for resources with different priorities.



That is why I am going to look at it over the next few batches. I will know after getting work for another three or four batches if the differences are consistent or statistically significant. It may be something to do with the priority given by default to the VM? (I haven't changed the priority given to anything from defaults.)
ID: 70641 · Report as offensive     Reply Quote
wujj123456

Send message
Joined: 14 Sep 08
Posts: 127
Credit: 40,873,378
RAC: 56,718
Message 70642 - Posted: 12 Mar 2024, 0:38:07 UTC - in response to Message 70637.  

Does CPDN use system libraries extensively for calculation? If so, that could simply be 24.04 is actually faster. IIRC, Ubuntu 24.04 is experimenting with x86-64-v3 target while anything older is using baseline x86-64. AVX2 can make this kind of difference under the right conditions, though 20% does sound a bit too good to be true. However, I haven't tried 24.04 so I'm not sure if they've rolled x86-64-v3 out to test images already.
ID: 70642 · Report as offensive     Reply Quote
Profile Dave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4529
Credit: 18,661,594
RAC: 14,529
Message 70643 - Posted: 12 Mar 2024, 6:53:12 UTC - in response to Message 70642.  

Good point. If so, next month we could see a credit jump for those who upgrade right away.
ID: 70643 · Report as offensive     Reply Quote
Glenn Carver

Send message
Joined: 29 Oct 17
Posts: 1044
Credit: 16,196,312
RAC: 12,647
Message 70645 - Posted: 12 Mar 2024, 9:44:43 UTC - in response to Message 70643.  

The wah models are compiled to use up to sse4.1 & sse4.2 but not avx. The code includes all the maths functions it needs like FFTs. I think it links against lib math, forget now.
I doubt it's anything to do with libraries. See what you get on a quiet system.
---
CPDN Visiting Scientist
ID: 70645 · Report as offensive     Reply Quote
Profile Dave Jackson
Volunteer moderator

Send message
Joined: 15 May 09
Posts: 4529
Credit: 18,661,594
RAC: 14,529
Message 70646 - Posted: 12 Mar 2024, 10:04:05 UTC - in response to Message 70645.  

The wah models are compiled to use up to sse4.1 & sse4.2 but not avx. The code includes all the maths functions it needs like FFTs. I think it links against lib math, forget now.
I doubt it's anything to do with libraries. See what you get on a quiet system.


Out of interest, once the last of my WAH tasks are finished, I will try some WCG or other work to compare if there is no CPDN Linux work by then. While a completely quiet machine will test for absolute speed, the results I get in practice are more important to me. I can always work out the standard deviation on the native and virtual results and check for significance once I have a bit more to work with.
ID: 70646 · Report as offensive     Reply Quote

Message boards : Number crunching : A performance oddity.

©2024 cpdn.org