Message boards : Number crunching : Sixtrack (notag/sse2/pni/sse3)
Message board moderation

To post messages, you must log in.

AuthorMessage
Carsten Milkau

Send message
Joined: 3 Oct 06
Posts: 2
Credit: 1,769,561
RAC: 0
Message 27798 - Posted: 16 Aug 2016, 7:27:32 UTC
Last modified: 16 Aug 2016, 7:31:49 UTC

I recently noticed that credit on sixtrack is surprisingly low, compared to other projects (about a factor 3-4 below expectation). I checked the WU and noticed two differences[1] to other users:

1) WU take 10-15x longer to complete on my machine
2) Other users run "Sixtrack (sse2)", while I run "Sixtrack" (without tag)

My CPU is 5y old, so that might already explain everything, but I'm worried about that SSE2 tag. If I'm using a non-SSE implementation, that would fit very well with that factor 3-4 performance gap. My CPU supports SSE2, SSE3, SSE4.1 and SSE4.2 (it's an Intel i7 920). If the (notag) Sixtrack is without SSE2, how do I make boinc use the SSE2 version?

[1] http://lhcathomeclassic.cern.ch/sixtrack/workunit.php?wuid=46101520
ID: 27798 · Report as offensive     Reply Quote
Harri Liljeroos
Avatar

Send message
Joined: 28 Sep 04
Posts: 585
Credit: 33,087,916
RAC: 18,016
Message 27800 - Posted: 16 Aug 2016, 9:13:12 UTC - in response to Message 27798.  
Last modified: 16 Aug 2016, 9:24:15 UTC

If you go to LHC@home website and view the details of your computer and select there the Application details, you will see that your computer has run both variations of application (with and without of sse2) and the server has rated the application without sse2 to be faster on your computer (Average processing rate). The number of finished and validated tasks with sse2 support is quite low (seven) and may be that those tasks were just of long running type (sometimes the simulated beam crashes to the wall very soon and the task is finished very fast).

From time to time the server will send you work also for the sse2 application to check if the situation (speed) has changed. So with a little luck you may see the server selected application changed.

It may be possible to use the "Reset project" selection in Boinc manager to force the speed evaluation process to start from the beginning but it would probably delete all unfinished tasks from your computer as well unless you clear your cache before that. It also resets the estimated time left calculations until 10 new tasks has been finished and validated.
ID: 27800 · Report as offensive     Reply Quote
Profile Bryan

Send message
Joined: 30 Jan 14
Posts: 4
Credit: 25,027,828
RAC: 0
Message 27810 - Posted: 10 Sep 2016, 16:03:20 UTC

SSE2 was introduced by Intel in 2001 and AMD followed suit in 2003. If people are running processors made prior to those dates then SSE2 isn't available.

The WU w/o extensions take MUCH longer to complete and therefore waste precious computer time.

Could you please quit sending them
-or
Put a check field in preferences that says:

I'm still running an abacus and the extensions are too hard on my fingers :)
ID: 27810 · Report as offensive     Reply Quote
Luigi R.
Avatar

Send message
Joined: 7 Feb 14
Posts: 99
Credit: 5,180,005
RAC: 1,090
Message 27823 - Posted: 16 Oct 2016, 14:24:09 UTC - in response to Message 27810.  

Hello, I think there is something wrong with server rating process.

I'm an i7-4770k owner and I have more than one BOINC client on the same host.
The server thinks one of those clients is faster to run (no tag) workunits, guessing cause of short tasks. Totally wrong. Usually I get sse2 tasks. I don't know why I don't get sse3 workunits, but I will try to crunch them through anonymous platform.
ID: 27823 · Report as offensive     Reply Quote
Uffe F

Send message
Joined: 9 Jan 08
Posts: 66
Credit: 727,923
RAC: 0
Message 27838 - Posted: 6 Nov 2016, 0:55:07 UTC
Last modified: 6 Nov 2016, 0:56:46 UTC

According to https://en.wikipedia.org/wiki/X86-64 SSE2 are an arcitectural feature of the 64 bit compatible processors, meaning that all 64 bit processors does have it (if I'm reading it correctly).

So right now on the applications page, it says that "Microsoft Windows running on an AMD x86_64 or Intel EM64T CPU" with "no-tag" is using 494 GigaFLOPS and "Linux running on an AMD x86_64 or Intel EM64T CPU" with "no-tag" is using 40 GigaFLOPS.

These could be run much faster as SSE2 applications, hence a lot of wasted cycles, that could be avoided by removing the 64 bit version of the "no-tag" program and only let the SSE2, SSE3 and PNI versions remain on 64 bit. Then the 64 bit client would always choose a minimum of SSE2 and wouldn't have to waste cycles testing if the "no-tag" version is faster.
ID: 27838 · Report as offensive     Reply Quote
Eric Mcintosh
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 12 Jul 11
Posts: 857
Credit: 1,619,050
RAC: 0
Message 27840 - Posted: 9 Nov 2016, 14:41:29 UTC - in response to Message 27823.  

Well SSE2 is just fine; I have compared with SSE3 and there is no significant
difference in performance. SSE2 gives at least a factor of 2 speed up.Note we run 32-bit code even on 64-bit systems. Eric.
ID: 27840 · Report as offensive     Reply Quote
Eric Mcintosh
Volunteer moderator
Project administrator
Project developer
Project tester
Project scientist

Send message
Joined: 12 Jul 11
Posts: 857
Credit: 1,619,050
RAC: 0
Message 27841 - Posted: 9 Nov 2016, 14:43:51 UTC - in response to Message 27838.  

Agh I think you are correct, and we must really use SSE2.
I'll ask my colleagues to delete the 64-bit "no tag" as you
suggest. Eric.
ID: 27841 · Report as offensive     Reply Quote

Message boards : Number crunching : Sixtrack (notag/sse2/pni/sse3)


©2022 CERN