Windows S5R3 App 4.15 available for Beta Test

googloo
googloo
Joined: 11 Feb 05
Posts: 43
Credit: 13154292
RAC: 2624

Another one succeeded;

Another one succeeded; started and finished with 4.15. This one without EAH_NO_SYNC. Can't tell much difference, given the variation in completion times.

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7023404931
RAC: 1808138

4.15 has the same old

4.15 has the same old periodicities in CPU time vs. sequence number within frequency we have seen since the start of S5R3.

However, as Richard Haselgrove pointed out a while back, my initial claim that the period he saw first and I second was close to 60 is only true in the frequency range of those first observations.

As it happens, for 4.15 one of my hosts picked up a bunch of 368.70, which appears to have a period near 30, while another picked up some 138.05, which at first glance does not appear to follow the pattern, but may in fact show a period down near 4.

I urge caution on all those concluding from a couple of new results that the new ap is faster on their host than the old. It is quite common to get sequential results, so with these periodicities things look better or worse depending on where you are in the cycle.

Here is a quick summary of approximate periods I've noticed reported to date:

_freq_ period
138.05 4
368.70 25
464.35 46
522.10 55
542.70 61

Even though my period estimates have some slop in them, I think it likely that the apparent trend of increasing period with higher frequency is real.

Here is a graph of these estimates:

At a guess, it may be hard to see the periodicity even with sequential samples for frequencies much below 250, while at the high frequencies even a sparse sample should still show the signature easily.
[edited to revise the period estimate for 368.70 from 30 to 25]

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2139
Credit: 2752760842
RAC: 1457500

Have you collected enough

Have you collected enough data to comment on my other hypothesis - that result __0 seems to come at a maximum of runtime? The 138s on Stoll3 seem to match, and the 368s on Stoll4 do as well if we postulate a period of 28 rather than 30.

It was this question of phasing which had Bernd scratching his head, and which clearly would have to be confirmed (along with the periodicity) before any compensating tweaks could be build into the credit awards.

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7023404931
RAC: 1808138

RE: Have you collected

Message 74870 in response to message 74869

Quote:
Have you collected enough data to comment on my other hypothesis - that result __0 seems to come at a maximum of runtime?


I think that hypothesis is alive, and looking a bit better. For another look I tried aggregating extremely close frequencies. Initially I tried aggregating all 4.nn applications as well, but quickly saw that was a problem.

I think this graph is pretty strong evidence that the Windows 4.13/4.15 aps are nearly 10% faster than the 4.07 ap for these work units on this type of host.

This graph also supports your suggestion that the cyclic period for 368.nn is 28, and neither 30 nor 25. That is a bit dangerous, however, as the little cluster of points in the 90s all are 368.60, while the ones from 19 through 40 are all from 368.55. It thus adds a little support to the maximum at 0, but we'd need a significantly more diverse set of frequencies to add confidence on that point.

On the negative side for this style of presentation, I think some of the increased noise is coming from my combining of closely related frequencies.

Also I should confess, the 138.nn curves combine 137.nn and 138.nn frequencies, so my labels are not quite full disclosure

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3522
Credit: 686043288
RAC: 582127

Hi! See this recent

Hi!

See this recent message by one of the scientists for the reasons of the runtime variations. Basically , the different workunits cover different regions of the sky, and the algorithm used by E@H runs faster in certain regions of the sky than in others.

So the period of the variation is determined by the number of units it takes to cover the whole sky (after that, the workunits start over with a new pass over the sky). The phase is given by the starting point. Different frequency workunit series use slightly different grids to cover the sky, so that period and phase can differ.

The data collected here is very useful in understanding the runtime variations, and verifying the explanations for it, thanks to all who provide this data!!

CU
Bikeman

googloo
googloo
Joined: 11 Feb 05
Posts: 43
Credit: 13154292
RAC: 2624

Here's another successful

Here's another successful completion. I will report only failures after this.

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4266
Credit: 244924143
RAC: 16679

I published the 4.15 App to

I published the 4.15 App to be official.

I hope it soon gets followed by another App that has some bug fixed I recently found in the BOINC library.

BM

BM

Brian Silvers
Brian Silvers
Joined: 26 Aug 05
Posts: 772
Credit: 282700
RAC: 0

RE: I published the 4.15

Message 74874 in response to message 74873

Quote:

I published the 4.15 App to be official.

I hope it soon gets followed by another App that has some bug fixed I recently found in the BOINC library.

BM

Any chance at getting the SSE optimization that the Mac OS X app has? :-)

Bernd Machenschalk
Bernd Machenschalk
Moderator
Administrator
Joined: 15 Oct 04
Posts: 4266
Credit: 244924143
RAC: 16679

RE: Any chance at getting

Message 74875 in response to message 74874

Quote:
Any chance at getting the SSE optimization that the Mac OS X app has? :-)


I'm actually working on it. SSE seems to be less [edit:]sensitive to our famous "Input domain error"s than the x87 FPU code.

BM

BM

Annika
Annika
Joined: 8 Aug 06
Posts: 720
Credit: 494410
RAC: 0

Hey Bernd ;-) I think that

Hey Bernd ;-) I think that makes two "false friends" in a single sentence...

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.