MacOS 10.4 G5 App 4.23 available
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4312
Credit: 250179637
RAC: 34740
2 Aug 2006 13:12:10 UTC
Topic 191646
(moderation:
Hi!
I've gone again over the 'vectorizing' code of the first S5 PPC Mac Apps and modified it a bit. The results that have been found invalid with the previous code now successfully pass validation (so far I could re-run them), so the validation problem should have been reduced significantly. The speed has been reported to be close to the 4.06 App.
Building a fully compatible App from this code is much of a hazzle, so this App version was built on MacOS 10.4 and for G5 only. It probably doesn't run at all on older OS versions or other CPUs.
Therefore I published it on our new Power User Apps page. Test it only if you are reasonably sure of what you are doing.
BM
BM
MacOS 10.4 G5 App 4.23 available
)
My 2.3-GHz dual-core G5 has just had its first WU crunched with v4.23 validated: no problems to report.
This system has been crunching from a batch of long “h1_0927.0_…� WUs pretty much since it was attached, so I presume the work involved to be comparable. Taking an average of about 19.5 hours and earning 176 cobblestones each, its production began at about 9 CS/h. Its last result before I installed v4.23 was apparently validated under the new credit system and therefore granted 146 credits, 17% less than before (suggesting an average rate of 7.5 CS/h, although that particular result came in at 7.1 CS/h). Now, using v4.23, it’s taken a little over 15 h to earn 120 credits, for nearly 8 CS/h.
Not that I’m complaining, considering that it crunches some 25% faster, but is the ‘credit rating’ for this version 18% lower than v4.12’s, on top of the 17% general reduction? Or has the general figure been reduced a second time, merely coincidental with my switch to the beta app? Or could that most recent WU be 18% shorter than the others from the same data file, so it’s likewise just a coincidence that it earned less under v4.23?
There were two credit
)
There were two credit reductions, for a net decrease of slightly under 1/3rd. The new x86 windows/linux apps have a 35% speedup, and since they make up the vast majority of crunching boxes the credit scaling was very close to what they earned. the two step reduction was probably to avoid howls if the initial cut was too small. Since older x86 (4+ for athlons, several more for pentiums) boxes didn't support the code where the speedup was made a simple survey of what percentage of boxes were running windows/linux wouldn't be able to accurately measure the project wide gain, and the per cpu page is a royal mess.
RE: There were two credit
)
Thanks; I just hadn’t realized there had been two steps. 0.83 * 0.82 ~= 0.68, which makes a 32% reduction. Considering I’m working with a sample of one, that seems quite consistent!
I used your sample to get the
)
I used your sample to get the .82 and .83. Much easier than going through my old results for the same. :)
I’ve crunched five WUs with
)
I’ve crunched five WUs with v4.23 now, with no errors. All validated fine except for the most recent, which is still pending. That’ll be it for my G5, as I installed v4.27 today.
I got a bit of a pleasant surprise, earning 176 cobblestones for one long WU amidst the now-usual 120-CS grants. That one had been reissued when someone missed a deadline, being sent to me for ‘cleanup’, and I guess it held its originally assigned value through the recent decreases. Nice work if you can get it, as they say … ;)
Thanks. A new App 4.27 is
)
Thanks. A new App 4.27 is available for G5, that should also run on 10.3. Read about it here. Should be even faster than 4.23.
BM
BM