Linux S5R2 App 4.27 available for Beta test
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4312
Credit: 250192356
RAC: 35035
16 Jul 2007 9:53:49 UTC
Topic 192966
(moderation:
A Linux App (4.27) is available from our Beta Test page.
This App is part of a new generation of Apps that follow our attempts to track down and fix the remaining problems of the new software we're using since S5R2 (There will soon be a MacOS PPC App once I've fixed a CPU-specific problem there).
This may remain an intermediate release that will never make it to become official Apps if we get enough feedback from running the Beta Test Apps that we can fix some more problems for the next "official" Apps.
Following a suggestion from Gary these Beta test packages feature an app_info.xml that should allow for continuing a Task that has been assigned to and possibly partially crunched with an older App version with this Beta App (keeping the old App version number, though). Note that there are some clients with which this doesn't work as desired, i.e. that reset the project when switching the platform (I've seen this behavior with the first Universal (PPC+Intel) MacOS Clients). So this Beta Test also covers this app_info.xml scheme. If in doubt, wait at least until the currently running Task has finished, report it manually and install the Beta App then.
gcc-compiled Apps such as the ones from MacOS and Linux apparently get a little speedup from using ftol() instead of modf() (which we originally did to fix problems of the Windows App), but this release is not primarily meant to improve speed. We'll work on this not before we have the validation problems found solved.
Please test&report here.
BM
BM
Linux S5R2 App 4.27 available for Beta test
)
Thanks, Bernd and Gary
This new "app_info.xml" works very well on my AMD linux box, using BOINC 5.8.16.
I have a WU partially computed with the 4.21 app. I've simply added the new files and restarted BOINC, and now it is running with the new 4.27 app without any trouble.
BoincView still shows the 4.21 version number, but with the "top" command I can see it's really the 4.27 version.
So switching to the Beta test is much easier now. A very nice feature.
RE: So switching to the
)
Yes, it is a very nice feature, particularly for people who maintain a significant cache of work. It is so simple and convenient to avoid the dead time of running down the cache with the old app or having to jump through hoops to make the transition in a variety of other, less safe ways.
Because of the convenience, I think a lot more people could be attracted into running the beta test than otherwise would do so.
Cheers,
Gary.
Just switched over now,
)
Just switched over now, conveniently just after 4 WU's started, the app_info.xml approach made it much easier with a large cache. The first 4 WU's should be done in about 11.5 hours from now so I'll post back then.
Thanks guys.
There are 10^11 stars in the galaxy. That used to be a huge number. But it's only a hundred billion. It's less than the national deficit! We used to call them astronomical numbers. Now we should call them economical numbers. - Richard Feynman
I have finished my firsts WUs
)
I have finished my firsts WUs with 4.27 beta app and it seems slower than the previous version (4.21)
My machine is an AMD Duron 1600 (Applebred) running OpenSUSE 10.2
here are my results:
[link]http://einsteinathome.org/account/tasks[/link]
All of them have 394.70 granted credits, so I suppose they are equal in size.
With einstein's 4.21 app the results took about 113000 seconds. Now this completion time has increased to 138000.
The result completed in 130000 seconds is a mixed one, first 30% computed using 4.21 and the remaining using 4.27
I know this version is not intended for increasing speed, and I don't care about credits, it's the science work done what I like. I post this to check if this behaviour is normal.
Hi! I see the same effect.
)
Hi!
I see the same effect. The new beta app seems to be betweeen 25% and 30% slower than the previous official version on my Linux hosts so far.
CU
BRM
I inserted some sanity checks
)
I inserted some sanity checks at one place to track an access violation Client Error. On my machines the App didn't run noticeably slower. Leads me to wonder how representative the workunits are we currently use for testing.
Please continue to advise people suffering from Client errors (especially on Windows) to still try this Beta App. It should help to track the problem down.
I'll see what I can do about the speed then.
BM
BM
RE: I inserted some sanity
)
Hi Bernd,
It seems that gcc generates inline code for the "(INT8) x" truncation that makes heavy use of the dreaded "fldcw" instruction, stalling the pipeline and hurting performance quite badly. Going back to modf for gcc only could be an option, or using rounding instead of truncation here and then dealing with negative result of x - ROUND(x) separately. Something like that.
For the curious, an interesting short discussion on this effect can be found here: http://mega-nerd.com/FPcast/
CU
H-B
Thanks for the input.
)
Thanks for the input. Switching back to modf() is actually the easiest thing to do.
We're currently fixing some more x-platform differences (not in speed, but in result). After that we should be in the position to try the SIN/COS routine Akos and I implemented a while ago in larger scale - it won't be using any of modf(), ftol() or whatever.
BM
BM
I have completed 4 WU's with
)
I have completed 4 WU's with the beta app. 3 of the WU's have validated while the fourth is still pending. The loss in performance is rather large, going from an average of 42600s to 50300s required for the same data pack.
The windows app now performers better on my Intel C2Q with 49000s required for a 407 credit WU running XP, while 50000s is required for a 381 credit WU running Linux.
There are 10^11 stars in the galaxy. That used to be a huge number. But it's only a hundred billion. It's less than the national deficit! We used to call them astronomical numbers. Now we should call them economical numbers. - Richard Feynman
RE: It seems that gcc
)
Plaese try 4.31.
BM
BM