Matroxusers team for CPDN BOINC

[Umfriend]

Oh, OK, thought you simply did not care and wanted to rub that in...

To bad though, we miss ya

[VJ]

OK, one week after I rang the alarm bells, the only thing that changed is that AZ resumed crunching again

As for the rest, where you guys? Whasup? Need help?

I have upgraded the client on my Xeon to the latest one, and resumed it... (finally).

With the new client, I can see the s/TS for each job (previous client crashed when trying this), and it is clear the dual xeon is seriously underperforming. The new client allows individual work units to be suspended, and even when running a single work unit it yields 5.55 s/TS (xeon 2.4 Ghz, 533 FSB), compared to the 2.4 s/TS (P4 2.4 Ghz, FSB 133) of my work PC.


Jörg

[Umfriend]

What results do you get from the CPU benchmarks? How did you get the 5.55 s/TS? You may remeber that what it shows is the average since starting the model, so you need to tally it as of when you change something until your measurement point.

[VJ]

What results do you get from the CPU benchmarks?

Quite similar to the P4...

How did you get the 5.55 s/TS? You may remeber that what it shows is the average since starting the model, so you need to tally it as of when you change something until your measurement point.

Oops... I forgot...
All my work units have ran for some time (2 are finished, 1 is close to 60%, the other close to 40, IIRC), so bearing in mind the average, it is hard to tell...

Guess I'll just have to see if the number goes down... (I have changed the settings quite often: once running 4 units in parallel, then running only 2 units, ...; so as I didn't keep track of what time it ran in any configuration, I can't calculated the current speed just from the average)



Jörg

[Umfriend]

If you run one unit now, you can use the two trickles shown after the change to runing on unit. Simply take the difference in CPU secs between the 2 and divide by 10802. I'l be waitin for a full report.

[VJ]

If you run one unit now, you can use the two trickles shown after the change to runing on unit. Simply take the difference in CPU secs between the 2 and divide by 10802. I'l be waitin for a full report.

Is it possible to see the CPU secs between 2 trickles without uploading (remember, this machine is offline for almost all of the time) ?


Jörg

[Rattledagger]

Is it possible to see the CPU secs between 2 trickles without uploading (remember, this machine is offline for almost all of the time) ?


Jörg

Yes, just look under BOINC\projects\climateprediction.net
Any unsent trickle-files should be here, and these includes cpu-time.

[gt40]

looks like I'm going to get my beer back from GNEP

[GNEP]

Hmmm, don't be so quick off the mark gt... one machine hasn't connected to the network in several days due to wireless issues... didn't have time to fix them but it's all sorted now so expect me to get a little bit of a boost v. soon

[Umfriend]

#23

[VJ]

Yes, just look under BOINC\projects\climateprediction.net
Any unsent trickle-files should be here, and these includes cpu-time.

Ok, my dual Xeon is crunching at 3.8 s/TS when running two units in parallel (compared to 2.5 s/TS for my P4). I haven't tried just one, but I doubt I'll see benefits in running 1 at a time (I doubt it will run at 1.9 s/TS ).

So the 5.5 s/TS average shows that running 4 units in parallel is taking a serious performance hit. Still, doesn't it also show that running 4 in parallel is still faster than running 2x 2 in parallel?



Jörg

[Umfriend]

Yes, two at 3.8 s/Ts equals one at 1.9 s/Ts and 4 at 5.5 s/ts equals 1 at 1.375 s/Ts. Now that wasn't to hard, was it?

Your gain is approx. 38%, which is quite high. Could it be the OS sometimes schedules both units at one physical CPU? I gain some 23% using HT.

[VJ]

Yes, two at 3.8 s/Ts equals one at 1.9 s/Ts and 4 at 5.5 s/ts equals 1 at 1.375 s/Ts. Now that wasn't to hard, was it?

The 5.5 is what it averages now, after crunching with 4 units for quite some time, and with 2 units for quite some time. So the real speed of 4 units will be lower (according to some previous post of mine, it crunched at 7.5 s/TS with 4 units).


The only reason to run one unit, is to test if running 2 units is slowed down by some bottleneck, or whether it effectively runs twice as fast. But it isn't that important: there is a speed gain.



Jörg

[Umfriend]

OK, so if I get it right you are saying:
2 units & 3.8 s/Ts for each unit -> 1.9 s/Ts -> 4.21 trickles/day
4 units & 7.5 s/Ts for each unit -> 1.875 s/Ts -> 4.27 trickles/day

If I remember correctly, we concluded earlier that 3 trickles was you optimum, but at the end of the day, the differences are small.

[VJ]

OK, so if I get it right you are saying:
2 units & 3.8 s/Ts for each unit -> 1.9 s/Ts -> 4.21 trickles/day
4 units & 7.5 s/Ts for each unit -> 1.875 s/Ts -> 4.27 trickles/day

If I remember correctly, we concluded earlier that 3 trickles was you optimum, but at the end of the day, the differences are small.

Yep.
I might try running just one unit, just to see if running stuff in parallel is suffering some limits (bandwith, ...). Client 4.25 allows one to individually suspend running units .


Jörg