MS Windows Beta Test App 4.62 available
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4312
Credit: 250715741
RAC: 35696
5 Jun 2006 12:29:36 UTC
Topic 191326
(moderation:
I put a new Windows App for testing on the Beta Test Page.
We are still working on incorporatinging Akos' optimizations into an App that meets the requirements for our "official" Apps, but it turned out to take somewhat longer than we expected.
As we still need a new App to start the next run with, I made this intermediate release of a standard Windows App from the same codebase as the 4.58 Linux App.
Short tests showed a speedup of 3-4 compared to our current official Window App 4.37. This is not as fast as Akos' Apps with architecture-specific optimizations, but as this will probably be the official Windows App we start the next run with (if this test turns out to be successful), I strongly encourage you to test it beforehand.
BM
BM
MS Windows Beta Test App 4.62 available
)
Is there the new version of charging the credits involved? And then yes, how does it function?
The credit is claimed by the
)
The credit is claimed by the Core Client (and the scheduler), and granted by the validator, so primarily this doesn't have much to do with the App.
There is a way to tell the Client the amount of FLOPs that's actually used from the App, instead of using the client's benchmark result, but it's hard to implement a counting that works on all CPUs and it would just balance for the different benchmark results of different Clients, but does not much about the different efficiencies of the Apps on different CPUs.
So in the next run, starting in about two weeks, we will grant a rather fixed credit for each (valid) Task done. But that, again, is rather related to the scheduler and the validator, not to the App.
There will be no change regarding crediting for the rest of the current run.
BM
BM
Does this mean that those of
)
Does this mean that those of us currently using a CPU specific akos app will be able to continue using it for S5?
RE: Does this mean that
)
I simply can't say that right now. It is covered in internal tests of the new setup which are not finished yet.
Akos' Apps are still based on the 4.37 App, which is pretty old code, especially the BOINC part. If it turns out we need e.g. fixes or features from newer BOINC, then you won't be able to continue using these Apps.
BM
BM
Got my first result with 4.62
)
Got my first result with 4.62 on a P4 2GHz SSE machine with XP SP2. Stock 4.37 client times in the past were about 40,000 seconds with some variation. 4.62 ran ok in 13760 seconds, around 3 times faster. Akos S41.07 times around 6070 seconds for similar WUs, a little over 2 times faster again. This looks like a great speed up for the stock client, and leaves Akos some scope to apply a useful boost again eventually if he feels like it for those of us who enjoy that sort of thing. Well done Bernd and the team.
My first result with 4.62 is
)
My first result with 4.62 is similar (more than twice as long as my typical for S41.07).
Hi Bernd! RE: Short
)
Hi Bernd!
I think this speedup is really good. You could achieve the "common divisor" and the "combined precision" parts. I think the only slow-points of the application are the type conversions at the sinus-cosinus calculations.
Greetings, on my Pentium
)
Greetings,
on my Pentium III 1100@1233 MHz (now with W2ksp4) are no problems so far.
Times:
Std 4.37: 34,501.00 sec
Beta 4.62: 12,649.08 - 13,109.71 sec
S41.07: 5,379.56 - 5,499.02 sec
see here - (the Client error was my fault - forgot to replace a file).
All betas are still waiting for validation.
Beta is ca. 2.68 faster
S41 speedup is ca. 6.34
thanks for your work
MB
P4 2.8GHz, 1024MB DDR333,
)
P4 2.8GHz, 1024MB DDR333, Windows 2000 SP4, BOINC 5.4.9
Akosf's S39L 4.37 would run results in 4,600 - 5,200 seconds.
First result with 4.62 runs in 9,288 seconds.
Results list.
Not bad, so far. I'll keep it on my system.
Slooooow!!! But work
)
Slooooow!!!
But work -:)
CPU time 9729.300046 app 4.62
http://einsteinathome.org/task/33209298
CPU time 4128.486472 app 4.37 D41.14 (akosf)
http://einsteinathome.org/task/33131853
As above seems that 4.62 delays more than double time than 4.37 askof (d4114)
to crunch a WU -:(
So,
I reverted back to the faster version
Thanks
Click signature for global team stats
