speed predictions...

[VJ]

Originally posted by Umfriend
Sis is saying the room where her 4 PC's are was getting warm She turned them off, and now onl run when actually used. This w/end you'll have passed me I'd think.

Well, I turned my heating down...

Luckily, it is winter, for in summer running my PC would make it too hot in the room (it is a very hot building). I noticed it last summer.


Jörg

[Umfriend]

I've been thinking about starting a water-cooled crunching farm and use th e water to heat the floor (or the floor to cool the water).

It's a tuff choice, either start something like that or be able to send the kids to college.....as yet undecided.

[VJ]

I have been thinking of making my system more silent. I was planning on exchanging my fans (Adda) by more quiet ones (Noiseblocker, Papst or Verax) and adding sound proofing mats. However, the possible influence of the mats on the temperature scares me...
(lack of good reviews of Verax fans is also holding me back)

I also thought of watercooling (i.e. Exos II, Zalman reserator), but it would yield other problems: cooling 10K scsi drives, cooling my mainboard chipset, ... And as I have no overclocking intentions, it is not that big a deal. And finally, I also gave peltiers some thought (as a means of transferring heat from the inside of the case to the outside; not mounted as a cpu cooler), and found some applications of it...


Jörg

[VJ]

Just uploaded again...
Running 3 units in parallel results in about 2000 credits after one week. So in number of credits, it is the exact same as running 2 units in parallel.
I average 5.4 sec/TS now...

Well, it was enough to temporarily hop over Umfriend again...


Jörg

[VJ]

Weird, with 2 units in parallel, I also have 5.4 s/TS, and a similar number of credits...

05 Dec 2004 15:43:42	hidden	380333	3j2t_100186348_0	1	226842	1234774	5.4433	
05 Dec 2004 15:43:42	hidden	380790	3jfd_100186805_0	1	237644	1287968	5.4197	
05 Dec 2004 15:43:42	hidden	380790	3jfd_100186805_0	1	226842	1230693	5.4253	
05 Dec 2004 15:43:42	hidden	380333	3j2t_100186348_0	2	32406	1587417	5.4428	
05 Dec 2004 15:43:42	hidden	380333	3j2t_100186348_0	2	21604	1528906	5.4438

Jörg

[Rattledagger]

Letting Excel do the math, after removing the trickle then changing #cpu the averages are:

result 380333:
2 instances: 3,667 s/TS
3 instances: 5,441 s/TS
4 instances: 7,554 s/TS

result 380790:
2 instances: 3,632 s/TS
3 instances: 5,339 s/TS
4 instances: 7,500 s/TS

result 380791:
3 instances: 5,349 s/TS
4 instances: 7,564 s/TS


This gives your theoretical daily production:
2 instances: 4,383 trickles/day
3 instances: 4,463 trickles/day
4 instances: 4,244 trickles/day

In other words, 2900 Cobblestone/week if runs all the time...

[VJ]

Cool... euhm, does this last 'theoretical daily production' mean that 3 instances is the fasted, 2 is second fastest and 4 is the slowest?

How did you calculate this?


Just for comparisons (i.e. with single processor, I want to check if I'm memory bandwith limited), I'm currently running just 1 instance.

Expect an upload sometime later this week...


Jörg

[Rattledagger]

Look on the overview of all trickles for a result, example http://climateapps2.oucs.ox.ac.uk/cp...esultid=380333

If you wants to calculate many timesteps, the easiest is to copy this table to Excel, and you should now get cpu-time in column G. Use this formula:
I2 = (G2-G3)/10802 and you gets s/TS for this last trickle.

Since s/TS isn't constant, can't really say if it's better to run 2 or 3 instances of CPDN.


BTW, it's also possible to calculate s/TS based on the info in the trickle-files before they're uploaded.

[VJ]

ha, I see...


Jörg

[VJ]

Just a thought: suppose a PC is not working 24/7, does it impact the s/TS ?
i.e. suppose a PC is doing 2.4 s/TS, but is only on half of the day. Will it report 4.8 s/TS ?
(in other words: the seconds: is it 'seconds computing' time, or 'seconds real time')


Jörg

[Rattledagger]

If you're not running win9x that doesn't know what cpu-time is, you're using cpu-time so doesn't matter if you're down for some days or is pre-empted some hours or not.

Since you're running NT, the s/TS isn't dependent on you using 1 day or 1 month for a trickle.

[VJ]

Now this is odd...

To see whether the computations on my dual Xeon were bandwith limited, I ran only one instance. And this is the result:

08 Dec 2004 08:11:55	hidden	380333	3j2t_100186348_0	2	75614	1712682	5.1146	
08 Dec 2004 08:11:55	hidden	380333	3j2t_100186348_0	2	64812	1684552	5.1983	
08 Dec 2004 08:11:55	hidden	380333	3j2t_100186348_0	2	54010	1655778	5.2857	
08 Dec 2004 08:11:55	hidden	380333	3j2t_100186348_0	2	43208	1627579	5.3812

How come a Xeon 2.4 GHz (FSB 533) performs well below a P4 2.4 GHz (it does somewhere between 2 and 3 s/TS) ?
(hyperthreading was enabled, but nothing else was running on the PC)

It did show that running multiple instances in parallel don't slow it down much, so we can rule out bandwith limitations.



Jörg

[Rattledagger]

Hmm, after the latest numbers is added, the table for result 380333 looks like this:

4 instances: 7,554 s/TS
3 instances: 5,441 s/TS
2 instances: 3,667 s/TS
1 instances: 2,626 s/TS

This gives your theoretical daily production:
4 instances: 4,24 trickles/day
3 instances: 4,46 trickles/day
2 instances: 4,38 trickles/day
1 instances: 3,05 trickles/day


In other words, you only get 40-45% more production from running 2 or more instances compared to only running 1...

[VJ]

I'm really confused with those calculations... How do you get 2.6 s/TS ? And why isn't this value reported in the trickle information?
That value does more or less seem to match my P4.

The fact that multiple instances don't double the performance could be explained by e.g. a memory bottleneck (it is DDR266, registered, ECC).


Jörg

[Rattledagger]

Trickle-31; At phase 2, timestep 75614 you had used 1712682 s cpu-time.
Trickle-28; At phase 2, timestep 43208 you had used 1627579 s cpu-time.

1712682 s - 1627579 s = 85103 s
75614 TS - 43208 TS = 32406 TS
85103 s / 32406 TS = 2,626 s/TS

This shows my calculation.

As for the displayed calculation in the table:
1 phase = 24 trickles, every trickle is 10802 TS.
31 trickles = 334862 TS
1712682 s / 334862 TS = 5,1146 s/TS as displayed in the table.

The table shows the average speed for a trickle from start of a wu, so to get the speed of an individual trickle you must calculate this yourself.