Is s/TS a running average since the beginning of the WU or since the start of the current TS?

It seems to me that the s/ts measurement is averaged two different ways: as a running average between trickles, and also as a total average for the entire WU. This is only what I've inferred from observation, though, so I couldn't tell you for sure. Carl could probably shed some light on it for us.

well it's basically a running average, when you see it in the program or on a finished WU it would be the CPU time of the result divided by the maximum timestep reached. But in the "trickle detail" listing it is simply the "running average" up to that trickle point, so it can rise or fall on later trickles etc.

That's Carl

Well, at least I was half right :-D

*AMD Athlon XP 3000+ 2.17 Ghz, crunching @ ~2.5 s/ts
*Intel Pentium III 500 Mhz, crunching @ ~10.3 s/ts

I also have a question on the s/ts.
I have a p4 3.00GHz with HT running 2 models/Wu's and there is a big difference in s/ts:
First wu: between 1.9 s/ts and 2.2s/ts
Second wu: between 3.0 s/ts and 3.3 s/ts

I don't know if this matters but then first wu is running all the time, the second gets pre-empted (but remains in memory), sharing CPU time with seti

Is this normal?

> I also have a question on the s/ts.
> I have a p4 3.00GHz with HT running 2 models/Wu's and there is a big
> difference in s/ts:
> First wu: between 1.9 s/ts and 2.2s/ts
> Second wu: between 3.0 s/ts and 3.3 s/ts
>
> I don't know if this matters but then first wu is running all the time, the
> second gets pre-empted (but remains in memory), sharing CPU time with seti
>
> Is this normal?

It's probably normal if you think about hyperthreading --- two CPDN's running at once are probably really "competing" for memory bandwidth and FPU, but perhaps when it's one cpdn running with one seti there's not as much "competition" for the same resources (probably more so when it's memory bw as I imagine SETI is FPU-intensive! :-)

So when it's 1 CPDN WU running along with 1 SETI WU they "get along" better. I guess to prove this you'd have to see if CPDN WU#1 gets better sec/timstep avg in the intervals when it's running alongside SETI; but when it's two CPDN WU's at once they both are probably 3s/ts.

But this is all nice when you think that we different BOINC projects live in peace and harmony, and these nasty CPDN climate models just fight amongst themselves! ;-)

> > I also have a question on the s/ts.
> > I have a p4 3.00GHz with HT running 2 models/Wu's and there is a big
> > difference in s/ts:
> > First wu: between 1.9 s/ts and 2.2s/ts
> > Second wu: between 3.0 s/ts and 3.3 s/ts
> >
> > I don't know if this matters but then first wu is running all the time,
> the
> > second gets pre-empted (but remains in memory), sharing CPU time with
> seti
> >
> > Is this normal?
>
> It's probably normal if you think about hyperthreading --- two CPDN's running
> at once are probably really "competing" for memory bandwidth and FPU, but
> perhaps when it's one cpdn running with one seti there's not as much
> "competition" for the same resources (probably more so when it's memory bw as
> I imagine SETI is FPU-intensive! :-)
>
> So when it's 1 CPDN WU running along with 1 SETI WU they "get along" better.
> I guess to prove this you'd have to see if CPDN WU#1 gets better sec/timstep
> avg in the intervals when it's running alongside SETI; but when it's two CPDN
> WU's at once they both are probably 3s/ts.
>
> But this is all nice when you think that we different BOINC projects live in
> peace and harmony, and these nasty CPDN climate models just fight amongst
> themselves! ;-)
>

Thaks for the reply

I've trying to force the scheduler in doing what I want, but playing with the resource share doesn't help. Any help would be welcome.

What I've been able to do is changing the scheduler period and now there is one CPDN and one Seti running, before that there were 2 seti wu's running:
2 seti wu's: between 3:00 and 3:10/WU (or maybe just short wu)
1 seti to 1 CPDN => seti between 3:30 and 3:40/WU

For now I conclude that CPDN is dominant over seti

Update:

those seti wu were just short wu's, but still, if 2 seti wu are running they get a time close to 3:30, one seti and one cpu get a time close or over 3:40

With the changed scheduler period, now the projects get 4 hours of work before switching and for over 24 hour the scheduler work on one project at a time (4 hours seti then 4 hours CPDN)and the s/ts have remain (very) different:


Latest Trickles For This Host
Result ID | Phase | Timestep |CPU Time (sec)| Average (sec/TS)

139759 | 2 | 64812 | 810458 | 2.5010
201550 | 1 | 54010 | 175261 | 3.2450

> > I also have a question on the s/ts.
> > I have a p4 3.00GHz with HT running 2 models/Wu's and there is a big
> > difference in s/ts:
> > First wu: between 1.9 s/ts and 2.2s/ts
> > Second wu: between 3.0 s/ts and 3.3 s/ts
> >
> > I don't know if this matters but then first wu is running all the time,
> the
> > second gets pre-empted (but remains in memory), sharing CPU time with
> seti
> >
> > Is this normal?
>
> It's probably normal if you think about hyperthreading --- two CPDN's running
> at once are probably really "competing" for memory bandwidth and FPU, but
> perhaps when it's one cpdn running with one seti there's not as much
> "competition" for the same resources (probably more so when it's memory bw as
> I imagine SETI is FPU-intensive! :-)
>
> So when it's 1 CPDN WU running along with 1 SETI WU they "get along" better.
> I guess to prove this you'd have to see if CPDN WU#1 gets better sec/timstep
> avg in the intervals when it's running alongside SETI; but when it's two CPDN
> WU's at once they both are probably 3s/ts.
>
> But this is all nice when you think that we different BOINC projects live in
> peace and harmony, and these nasty CPDN climate models just fight amongst
> themselves! ;-)
>
>
>Thanks Carl, that almost answered my own question, or rather one that was asked in our teams forum. http://www.wwseti.net/forum/viewtopic.php?p=51258#51258

---