BOINC running on upcoming Nehalem system:question on HyperThreading

Message boards : BOINC client : BOINC running on upcoming Nehalem system:question on HyperThreading
Message board moderation

To post messages, you must log in.

AuthorMessage
Ivan

Send message
Joined: 7 Jul 08
Posts: 2
Message 18266 - Posted: 7 Jul 2008, 13:11:55 UTC
Last modified: 7 Jul 2008, 13:12:59 UTC

Hi people,
There is a benchmark from latest BOINC client running on single socket Nehalem system(2.933Ghz) on Xtremesystems:
http://www.xtremesystems.org/forums/showpost.php?p=3122412&postcount=414

My question is how many cores does the client support and will it work properly with HyperThreading enabled cores(like Nehalem is).
The above linked score is with HT turned off.Here is the one with HT -> ON:
http://www.xtremesystems.org/forums/showpost.php?p=3106435&postcount=350

Seems unrealisticly that 8 logical cores can score double than 4 physical ones...Is there a glitch somewhere and HT isn't actually working but showing the total number of cores as 8,or does it recognize HT enabled chip properly and works accordingly?From other benchmarks the HT brings around 20-40% increase in performance when enabled,in multithreaded applications.In BOINC,it's almost a double which is kinda hard to swallow.

Thanks in advance,
Ivan
ID: 18266 · Report as offensive
SekeRob

Send message
Joined: 25 Aug 06
Posts: 1596
Message 18267 - Posted: 7 Jul 2008, 13:30:44 UTC - in response to Message 18266.  
Last modified: 7 Jul 2008, 13:37:09 UTC

BOINC has been seen on BOINCstats with 96 cores and a teamy of mine managed 32. I've succeeded to force BOINC to run 4 jobs on a quad from a single core, so why not? The 4 did take more time to finish than running of 4 separate cores.

Like on a P4-HT don't expect that 4 virtual threads will double your performance. Overhead and resource restrictions will kick in.

Even when only running physical cores, don't expect 4 fold of a single core either. Depending on what the jobs do simultaneously, expect for a digressive curve. A good example is running SoB. 2 jobs in parallel is not 200% and 3 or 4 could even translate to getting less than 2 would do. In relation there's been discussion on these fora to limit the number of jobs per (sub)project, to optimize throughput. e.g. I'd never want more than 1 or 2 PrimeGrid or QMC to run simultaneous as they do elevate the temps several C i.e. rather intense work. Obviously some control could be exercised by setting lower resource shares for these projects so the scheduler is less likely to pull multiple jobs from these.

So, benchmarking with just 1 core enabled is bound to show a fair value. If multicore, expect an over claim on projects where credit is time driven.

cheers
Coelum Non Animum Mutant, Qui Trans Mare Currunt
ID: 18267 · Report as offensive
Keck_Komputers
Avatar

Send message
Joined: 29 Aug 05
Posts: 304
United States
Message 18268 - Posted: 7 Jul 2008, 13:31:16 UTC

It's been awhile since all the troubles trying to benchmark HT machines. I think the final solution was to run the benchmarks on each logical CPU and average the scores.

This still can give slightly high results in some cases. Particularly when half of the benchmark processes finish quickly.
BOINC WIKI

BOINCing since 2002/12/8
ID: 18268 · Report as offensive
Ivan

Send message
Joined: 7 Jul 08
Posts: 2
Message 18269 - Posted: 7 Jul 2008, 13:58:39 UTC
Last modified: 7 Jul 2008, 13:59:28 UTC

Thanks for the replies people.
So to recap,the HT enabled Nehalem(8 logical cores) system has no chance having(according to built in benchmark tool) double the score of 4 cores?How reliable is the built-in tool in terms of predicting the actual points per day?
So the system would need to run the actual projects in order to be seen if it can produce anywhere near the score the benchmark is showing(8x2800 vs 4x3100 and 8x7000 vs 4x7000).

@Sekerob

I'm a BOINC noob,i just started running my client, sorry if it was a noobish question i asked :).Thanks for the response and explanation ;).

Cheers guys
Ivan
ID: 18269 · Report as offensive
Sid
Avatar

Send message
Joined: 29 Jun 08
Posts: 30
United States
Message 18272 - Posted: 7 Jul 2008, 14:40:22 UTC - in response to Message 18269.  

Thanks for the replies people.
So to recap,the HT enabled Nehalem(8 logical cores) system has no chance having(according to built in benchmark tool) double the score of 4 cores?

How reliable is the built-in tool in terms of predicting the actual points per day?



Ivan:


Everything you find in the fan magazines and sites are mere speculation and guesses.

I believe that Nehalem will be as drastic a step forward as was P4 to C2D.

I have been following this development, with eager anticipation:

Intel Slates "Nehalem" for Q4 2008


If everyone is thinking alike, someone isn't thinking. - General George Patton



ID: 18272 · Report as offensive

Message boards : BOINC client : BOINC running on upcoming Nehalem system:question on HyperThreading

Copyright © 2022 University of California. Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.