i think it'd be better to work on an SSE2 version, until SSE3 is more widely used, unless there isn't that much benifit from SSE2 over SSE(1)
It sounds well. Could you tell me more about your idea?
I'm thinking on a SSE2(3) / 3DNow! overlapped code, but it needs some new programming technique to eliminate dataconversions.
e.g.: I'm testing a new rounding method.
well, i like the fact that you're trying to use the latest version of SSE (which is also supported by AMD)
however, only the newest processors support SSE3, and there are many that only support SSE2
now, unless there's not much speed gain by using SSE2 rather than SSE(1), then fair enough, but if there is a significant speed gain from using SSE2, then i feel an SSE2 app would help even further
another reason for me asking was that over at SETI, Crunch3r's SSE3 app isn't that much faster than his SSE2 version (only a few minutes quicker, not a whole lot), so if you only have time for either an SSE2 or an SSE3 app, i personlly feel your time would be best spent on an SSE2 version, because i think that would have the most benifit :)
an example:
I have a host with an intel Xeon CPU (Prestonia), which only supports MMX, SSE and SSE2
so, my best choice is the S series (SSE(1)), becuase the U series (SSE3) won't work
however, if there were an SSE2 version, i could crunch even faster, and i think many people are in this position, because there are a lot of CPUs (most pentium 4's) that only support SSE2, but don't support SSE3, so if SSE2 is faster than SSE(1) then we're losing a lot of processing speed here
----
I have no issue with a combinded app which supports multiple platforms (eg, auto-detects which platform it's running on, and uses the best code for that particular one, eg SSE on a P3 and SSE2 on a P4 (or SSE3 if it's supported) but 3DNow! on AMDs), my point is just that we're lacking an SSE2 app, which is causing a lot of CPUs to not be crunching as fast as they could be (because of only being able to use SSE(1))
...Some people propose [HT off]... yes, that's the thing: They think, but they don't know for sure!
What I noticed so far is: With HT on and crunching two einstein at the same time makes the time for one wu longer than crunching one einstein and one other project like seti simultaneous. The crunching time for einstein is about 10 to 20 percent lower than with two einstein at the same time.
But I havn't a proof yet.
AndyK
I don't need proof. Last year I used to run two Einstein WUs in 13 hours, while 1 Einstein WU opposite 1 Rosetta WU would run in 11 hours. So I would always set the resource share 50/50 and the "connect every.." to 0.01. This guaranteed that I always had 1 WU from each project onboard at any given time. However, after the Albert WUs came out and Akosf developed his S39L I have dedicated my machine 100% to Einstein (and also because Rosetta@Home was/is plagued with problems). -- NBIT
SSE2 needs some type conversions so that would be slower than the current versions.
Only the 3DNow! + SSE2 (SSE3?) combination would be faster with some percentages.
Using these opptimized programs, am i too assume that the results are ok?
I have cut my computing time down approx. 75%. I have tweaked my machine and
never could get fast results with Einstein.
I know the credits have gone down, which is secondary to me. But I have noticed
other fast machines that get higher credit witht the same time of computing.
Using these opptimized programs, am i too assume that the results are ok?
I have cut my computing time down approx. 75%. I have tweaked my machine and
never could get fast results with Einstein.
You can't expect all of your work to be validated OK.
Sometimes your optimized client gets a different result from the official app, and your result is not validated. (Eventough your result could be more accurate)
Quote:
I know the credits have gone down, which is secondary to me. But I have noticed
other fast machines that get higher credit witht the same time of computing.
Is the credit system that screwed up?
The credit system is somewhat useless with optimized apps.
You should use a BOINC client, which calibrate your credits.
I can see you use BoincStudio. It has the abillity to calibrate credit.
As far as I know it tweaks your result duration factor to 1 hence your claimed credits are "normalized".
I don't know how to activate the calibration with your version of BoincStudio.
With the mod version (which uses the original client) you add something like:
I noticed that a link was posted today to a new app, U41.05, on the New Optimised Executables Links thread with the comment "fixed memory prefetching fault". Should I replace U41.04 which seems to be working ok?
RE: RE: i think it'd be
)
well, i like the fact that you're trying to use the latest version of SSE (which is also supported by AMD)
however, only the newest processors support SSE3, and there are many that only support SSE2
now, unless there's not much speed gain by using SSE2 rather than SSE(1), then fair enough, but if there is a significant speed gain from using SSE2, then i feel an SSE2 app would help even further
another reason for me asking was that over at SETI, Crunch3r's SSE3 app isn't that much faster than his SSE2 version (only a few minutes quicker, not a whole lot), so if you only have time for either an SSE2 or an SSE3 app, i personlly feel your time would be best spent on an SSE2 version, because i think that would have the most benifit :)
an example:
I have a host with an intel Xeon CPU (Prestonia), which only supports MMX, SSE and SSE2
so, my best choice is the S series (SSE(1)), becuase the U series (SSE3) won't work
however, if there were an SSE2 version, i could crunch even faster, and i think many people are in this position, because there are a lot of CPUs (most pentium 4's) that only support SSE2, but don't support SSE3, so if SSE2 is faster than SSE(1) then we're losing a lot of processing speed here
----
I have no issue with a combinded app which supports multiple platforms (eg, auto-detects which platform it's running on, and uses the best code for that particular one, eg SSE on a P3 and SSE2 on a P4 (or SSE3 if it's supported) but 3DNow! on AMDs), my point is just that we're lacking an SSE2 app, which is causing a lot of CPUs to not be crunching as fast as they could be (because of only being able to use SSE(1))
Want to search the BOINC Wiki, BOINCstats, or various BOINC forums from within firefox? Try the BOINC related Firefox Search Plugins
There was at least one app
)
There was at least one app available in SSE1/2 variants, SSE2 provided a negligable gain. The SSE3 enabled U41 series smokes s31 on amd.
RE: ...Some people propose
)
I don't need proof. Last year I used to run two Einstein WUs in 13 hours, while 1 Einstein WU opposite 1 Rosetta WU would run in 11 hours. So I would always set the resource share 50/50 and the "connect every.." to 0.01. This guaranteed that I always had 1 WU from each project onboard at any given time. However, after the Albert WUs came out and Akosf developed his S39L I have dedicated my machine 100% to Einstein (and also because Rosetta@Home was/is plagued with problems). -- NBIT
RE: There was at least one
)
was the SSE2 app the S series?
if SSE2 won't give that much gain, then nevermind, i'm no programmer, so can't speak with any authority, it was just an idea :)
Want to search the BOINC Wiki, BOINCstats, or various BOINC forums from within firefox? Try the BOINC related Firefox Search Plugins
Only one S variant was sse2,
)
Only one S variant was sse2, it had a negligable gain over it's sse1 equivilant.
RE: was the SSE2 app the S
)
SSE2 needs some type conversions so that would be slower than the current versions.
Only the 3DNow! + SSE2 (SSE3?) combination would be faster with some percentages.
some results for you... any
)
some results for you...
any faults have been caused by me, so no fault of the clients..
Sempron 2800+ @1760 mhz, u41.04 2,900 av
compared to
Celeron-D 2.66 ghz, u41.04 3,400 av
the sempron likes this project, but the celeron-D impress's me..
the results are the other way around on some projects..
Just got rid of a celeron 2.4ghz, U41.07 but was extremely slow, compared to these other two..
that was 6,000 av...
Like the optimizing akosf...
Kath..
Couple of questions
)
Couple of questions
Using these opptimized programs, am i too assume that the results are ok?
I have cut my computing time down approx. 75%. I have tweaked my machine and
never could get fast results with Einstein.
I know the credits have gone down, which is secondary to me. But I have noticed
other fast machines that get higher credit witht the same time of computing.
Is the credit system that screwed up?
http://einsteinathome.org/host/534343
"Don't take life to serious .....after all none of us will make it out alive."
RE: Couple of
)
You can't expect all of your work to be validated OK.
Sometimes your optimized client gets a different result from the official app, and your result is not validated. (Eventough your result could be more accurate)
The credit system is somewhat useless with optimized apps.
You should use a BOINC client, which calibrate your credits.
I can see you use BoincStudio. It has the abillity to calibrate credit.
As far as I know it tweaks your result duration factor to 1 hence your claimed credits are "normalized".
I don't know how to activate the calibration with your version of BoincStudio.
With the mod version (which uses the original client) you add something like:
Einstein@Home
http://einstein.phys.uwm.edu/
1
0
1
To your bs_opts.xml file.
I noticed that a link was
)
I noticed that a link was posted today to a new app, U41.05, on the New Optimised Executables Links thread with the comment "fixed memory prefetching fault". Should I replace U41.04 which seems to be working ok?
Get BOINC WIKIed