Athlon XP's SSE
9 Nov 2011 11:32:08 UTC
Topic 196049
(moderation:
Hi,
I will give a bit of background first as I think it is pertinent. Note - at the time this occurred I was running gravity work exclusively.
I noticed that two of my computers were sometimes giving invalid results. One was running ~50% invalid, the other ~70-80% invalid. I did extensive diagnostic work on the worst of these. When I could not fix the "problem" I posted in Problems and Bug Reports: http://einsteinathome.org/node/195970
The answer was that the Athlon XP's with only sse were running a "compatibility" application and did not come up with results close enough to sse2 results to count. I have since shut-down the worst of the two computers and allowed the other to do binary pulsar work. I am unable to stop this computer from getting gravity wave work thru the E@H preferences. The binary work has had a 100% success rate on this computer as would be expected based on what Bernd said in this thread: http://einsteinathome.org/node/195853
So my wish(s)/suggestion(s)are:
If the compatibility mode produces invalids compared to sse2 on gravity work it should not be an option - disable the compatibility mode.
OR
Widen the computational difference tolerance if it doesn't hurt the science.
OR
Let folks opt-out their sse only computers from gravitational work.
Thanks
Joe B
Athlon XP's SSE
)
OR
4th option - run your app of choice under AP.
I actually have two old Athlon XP hosts still running and I chose to let them work on the new Gamma Ray Pulsar work rather than the BRP work. They don't seem to be having any issue with invalid results. My thinking was to leave the BRP work to the GPUs which do those tasks so much more efficiently.
Cheers,
Gary.
Hi Gary, Your #4 is
)
Hi Gary,
Your #4 is sort of the same as my #3. The difficulty is that I cannot unselect gravity wave work in "Edit Einstein@Home preferences for ...". So I presume that these sse machines might still get gravity work. So my #3 was supposed to suggest making it possible to unselect gravity wave work. If the project doesn't make that possible I'll try to abort any gravity wave work on my sse machine, but I don't always check that often and some computing time and wingman patience may be wasted.
Joe B
I previously thought that the
)
I previously thought that the few SSE-only hosts are not worth to justify the additional maintenance work of a pure SSE App for the GW search (currently S6Bucket).
However looking at a tool I originally wrote for tracking validation problems with the new FGRP search I found that the "compatibility apps" (x87 and Mac PPC w/o AltiVec) produce >50% invalid results.
This certainly needs investigation and maintenance. I'll probably build an SSE app and disable the "compatibility" ones (as for the Mac OS PPC app building is a collection of hacks which makes it a nightmare to maintain anyway).
Thanks for bringing this up again.
BM
Edit: I'm not fully sure when I'll have the time to get back to this, so feel free to remind me by posting here, say, every week.
BM
RE: The difficulty is that
)
which wouldn't be necessary with Anonymous Platform (AP), where only tasks are sent for applications listed in the app_info.xml
Gruß,
Gundolf
Computer sind nicht alles im Leben. (Kleiner Scherz)
Thank you Bernd & Gubdolf.
)
Thank you Bernd & Gubdolf.
Joe B
I built and published
)
I built and published SSE-only Apps for Windows and Linux. I'll keep an eye on validation.
BM
BM
Since you put the new app out
)
Since you put the new app out I have run 6 wu on 2 sse computers (4273844 & 4278906).
3 have validated & 3 are pending. So far so good.
Thanks
Joe B
Hi Bernd, I've started
)
Hi Bernd,
I've started seeing invalid wu's on my 2 sse computers.
4273844: 9 wu complete
6 valid
2 invalid
1 pending
4278906: 5 complete
2 valid
1 invalid
2 pending
I presume that if a wu is completed by 2 healthy sse computers the result will will be valid and it is only a mixture of computer capabilities that causes this problem. Is the valid/invalid tolerance set too tightly? Or do the sse computers sometimes produce results that are not scientifically useful?
Thanks
Joe B
Not sure yet. What I can
)
Not sure yet.
What I can say is that the statistics differ largely between the Windows and teh Linux SSE Apps. While the Windows SSE App has significantly less invalid results than the "compatibility" App (<50%), the Linux SSE App is almost as bad as the "compatibility" App there. The only advantage of the SSE App is that it should be much faster than the "compatibility" App.
I haven't looked into validation details of individual results yet. My initial assumption that the validation issues arise from x87 vs. SSE arithmetics, however, doesn't seem to hold here. Pitty.
BM
BM
Hi Bernd, Given the
)
Hi Bernd,
Given the differences - Linux vs windows and sse vs what I'll call sse+ - I still have a question. To try to explain it I propose a gedanken. If two sse/Linux computers run the same wu I expect that they would produce valid results. Also if two sse+ computers using either windows or Linux run the same wu (as the sse Linux computer) they also would produce valid results. If either set is accurate/precise enough to produce "good scientific results", why does running the same wu with one sse/Linux computer and one sse+ computer often produce invalid results?
It seems to me that either the sse/Linux computers are producing a high percentage of scientifically bad results and should be excluded from running gravitational work at all OR if sse/Linux computers are producing "good" results the valid/invalid tolerances are too tight and should be widened to accommodate results from both sets of computers when the wu is run with one of each.
Thank you,
Joe B