Computation Error With Optimized Apps

[Lysia]

Had Memtest running last night (about 10 hours), no errors detected.

And http://setiathome.berkeley.edu/workunit.php?wuid=417591604 is finished, found 30 single pulses and is inconclusive.

Nevertheless if you want to test the stability of your machine run prime95 ( http://www.mersenne.org/freesoft ) for at least 8h (some like to run at least 24h).

Running right now.

T2300 should support sse3. You can gain a slight increase in computation speed switching from sse2 to sse3 app.

That's the information in Wikipedia and intels datasheet, but /proc/cpuinfo just tells me

Code: [Select]

flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe nx constant_tsc arch_perfmon bts pni monitor est tm2 xtpr pdcmI guess the "ss" should be sse3 or ssse3

[sunu]

That's the information in Wikipedia and intels datasheet, but /proc/cpuinfo just tells me

Code: [Select]

flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe nx constant_tsc arch_perfmon bts pni monitor est tm2 xtpr pdcmI guess the "ss" should be sse3 or ssse3

nx is sse3.

[Josef W. Segur]

That's the information in Wikipedia and intels datasheet, but /proc/cpuinfo just tells me

Code: [Select]

flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe nx constant_tsc arch_perfmon bts pni monitor est tm2 xtpr pdcmI guess the "ss" should be sse3 or ssse3

nx is sse3.

I disagree. SSE3 was originally called Prescott New Instructions, and is shown as pni in that list (Windows Vista may also use the pni code). The nx code means an execution protection mechanism which allows marking memory regions for data only, not executable instructions. Early Prescott chips didn't have nx but obviously did have pni/SSE3, early AMD Athlon 64 chips had nx but not SSE3.
                                                                              Joe

[sunu]

I disagree. SSE3 was originally called Prescott New Instructions, and is shown as pni in that list (Windows Vista may also use the pni code). The nx code means an execution protection mechanism which allows marking memory regions for data only, not executable instructions. Early Prescott chips didn't have nx but obviously did have pni/SSE3, early AMD Athlon 64 chips had nx but not SSE3.
                                                                              Joe

I stand corrected. 

[Lysia]

mprime didn't find any problem in 20 hours, my system seems fine. What's left is a buggy app or there is a bug in the CPU that is triggered by an instruction which is used in the optimized AP but not in mprime.

[sunu]

Lysius, what tests did you run? small ffts, large ffts, blend?.

Let this machine do a few more WUs and we'll see how it goes. Also, as we've said earlier you can switch to a SSE3 app.

[Lysia]

Lysius, what tests did you run? small ffts, large ffts, blend?.

Blend ran for 20 hours, large ffts runs for 10 hours now without an error. I think I will do small ffts next.

EDIT:
I'm running http://setiathome.berkeley.edu/workunit.php?wuid=417924453 now an my Desktop PC with the SSE2 and the SSE3 AP.
I guess pulse.out is what would be send to the server. I have attached these for both apps (after 40 minutes of running). The SSE3-app doesn't seem to have found anything by now, the SSE2-app found 30 pulses already (almost immediately after start).

EDIT2: I have downloaded the app from http://calbe.dw70.de/linux32.html in case that matters. The link at http://lunatics.kwsn.net/optimized-applications-release-news/optimized-astropulsev5-5-03-r112-for-linux-osx-transitional-package.msg15103.html;topicseen#msg15103 does not work.

[attachment deleted by admin]

[sunu]

I don't think we can compare the two results in the middle of computation. Wait for them to end and then we can compare the results.

Yes, the Linux 32 bit AstroPulse_v5 5.03_SSE2 + 5.00 (INTEL/AMD) (transitional package) link is broken.

[Josef W. Segur]

...
EDIT:
I'm running http://setiathome.berkeley.edu/workunit.php?wuid=417924453 now an my Desktop PC with the SSE2 and the SSE3 AP.
I guess pulse.out is what would be send to the server. I have attached these for both apps (after 40 minutes of running). The SSE3-app doesn't seem to have found anything by now, the SSE2-app found 30 pulses already (almost immediately after start).
...

The 30 single pulses found by the SSE2 app are all at DM 911 and less. The app starts at DM 896 and goes to 15103, so if the SSE3 app hasn't found any and has run at least 1% of the work, it's clear the results won't match.

Lysius, would you upload that WU as an attachment here? Archive it with Zip, 7Zip, or RAR so it has an acceptable extension, not that it will compress much. I think the evidence indicates there's a flaw in the SSE2 app, having a WU which demonstrates the flaw would help.
                                                                     Joe

[Lysia]

The SSE3-App now has found something at DM 7023.

I have attached the WU.

[attachment deleted by admin]

[Urs Echternacht]

Rerunning your wu, Lysius. Using AP SSE2 application from Crunch3r's page on  my T7200, openSuSE 11.1, KDE4.2. If there is something wrong with this application this host should fail as well. It's now half an hour into that wu with ca. 20 hours to go.

[Josef W. Segur]

The SSE3-App now has found something at DM 7023.

I have attached the WU.

Thanks, and I apologize for listing Windows archivers rather than gzip and bzip2 which would have made sense here.

Because the WU showed signals very early on the SSE2 build I changed the <dm_high> value to 1023 [1] and ran it with the Windows ap_5.03r112_SSE.exe app with my Pentium-M system. No reportable pulses were found, the best_pulses all had peak_power well below threshold.

[1] We use changes to <dm_low> (which needs to be a multiple of the <dm_chunk_large> value) and <dm_high> (which can be any value higher than <dm_low>) to make AP WUs for standalone testing which run fairly quickly. If the difference is less than the <dm_chunk_large> value the crunch time is 1/111 of a full run.
                                                                        Joe

[sunu]

I'm also running  this wu with sse2 and sse3 32bit apps from crunch3r's page. Sometime on Tuesday I should have the results.

I don't know if it is of importance or no but these 2 apps are different from the ones I, and probably also Urs, have tested during the testing phase where there was no problem found. On the other hand the 32bit sse and 64bit sse3 apps from crunch3r's page are the same with the ones we tested. I don't know what crunch3r changed.

[Lysia]

Both are finished now. The SSE3-App actually took longer than the SSE2-App.

SSE3:

Code: [Select]

real    1475m25.372s
user    1363m40.233s
sys     86m3.655sSSE2:

Code: [Select]

real    1429m15.220s
user    1317m58.278s
sys     86m18.336s
[attachment deleted by admin]

[sunu]

They are different. Lysius can you also run the two clients attached below?. They are the ones we've tested before release.

[attachment deleted by admin]