Hi!

I have a strange behaviour of one of my machines to report.

I am running Boinc 5.10.13 on a Win XP Home AMD Althlon64 X2 4200+ CPU with 1GB RAM on a Asus A8V-VM SE socket 939 mainboard with an external Graphics card.

Usually it reports a integer speed of 4100 and a floating point speed of 2200.

Now since the last restart ( due to the Windows Patch Day ) the scores are totally different.
I noticed that on the benchmark when the second benchmark run begins one CPU-core stays idle. So it seems that the processes are not sticking to its core anymore.

Overall performance was roughly 33% normal with 765 integer and 465 float points, but varied widely on several runs in a row.

http://climateapps2.oucs.ox.ac.uk/cpdnboinc/show_host_detail.php?hostid=701301

CPU temperature was a moderate 44° C as I checked.

I ask myself what is the cause of the bahaviour?

• Windows not being sticky with the processes per core?
  
• BOINC not being sticky with the processes per core?
  
• Something hardware related?
  

Cheers

PJ

Hi PJ

I've had a look at your model results on this computer

http://climateapps2.oucs.ox.ac.uk/cpdnboinc/results.php?hostid=701301

The computer specs look fine to be running 2 models side-by-side. Only the second model on the list has trickled since you reported the problem. Its speed is slightly lower than before, but that could perhaps be accounted for by you exiting boinc repeatedly to check on performance (this would take the model back each time to its last checkpoint and then having to repeat calculations). Or by you previously not running the graphics/screensaver, but running it now.

http://climateapps2.oucs.ox.ac.uk/cpdnboinc/result.php?resultid=6606528

The sec/TS running speed given in the trickle list is a cumulative figure since you started the model. It would be useful if you could tell us the real current crunching speed of each model. To do this, for each model you need to

*Open the model's graphics window using the button in boinc manager
*Press Z then 8 on the keyboard to show more details
*Note down on paper the timestep reached and the exact time by your watch
*Close down the model graphics window
*Exactly 30 minutes or an hour later open the graphics again, look at the timestep reached now and do the calculation. We need to know how many seconds are needed to crunch each timestep.

Best not to do the calculation with the graphics or screensaver running because this slows down the model crunching speed.

By the way, I don't think there's anything we can do to make a particular workunit always crunch on the same core. In any case, on a dual-core, this same type of cpdn model should crunch at an almost identical speed on either core.

---

As you have an AMD processor, make sure 'cool & quiet' is turned off in the bios. This will reduce your clock speed to 1GHz if it thinks that the PC is idle. Because Boinc runs at Idle priority, it can trigger this behaviour.

Also try running the benchmarks again, sometimes if something is happening on the PC (for example, if you move the mouse), the benchmark scores will be affected.

---

Hi PJ

I've had a look at your model results on this computer

http://climateapps2.oucs.ox.ac.uk/cpdnboinc/results.php?hostid=701301

The computer specs look fine to be running 2 models side-by-side. Only the second model on the list has trickled since you reported the problem. Its speed is slightly lower than before, but that could perhaps be accounted for by you exiting boinc repeatedly to check on performance (this would take the model back each time to its last checkpoint and then having to repeat calculations). Or by you previously not running the graphics/screensaver, but running it now.

http://climateapps2.oucs.ox.ac.uk/cpdnboinc/result.php?resultid=6606528

The sec/TS running speed given in the trickle list is a cumulative figure since you started the model. It would be useful if you could tell us the real current crunching speed of each model. To do this, for each model you need to

*Open the model's graphics window using the button in boinc manager
*Press Z then 8 on the keyboard to show more details
*Note down on paper the timestep reached and the exact time by your watch
*Close down the model graphics window
*Exactly 30 minutes or an hour later open the graphics again, look at the timestep reached now and do the calculation. We need to know how many seconds are needed to crunch each timestep.

Best not to do the calculation with the graphics or screensaver running because this slows down the model crunching speed.

By the way, I don't think there's anything we can do to make a particular workunit always crunch on the same core. In any case, on a dual-core, this same type of cpdn model should crunch at an almost identical speed on either core.

At the moment one model has a speed of 6.9s / TS and the other roughly 6.7s / TS so roughly double what the average is.

PJ
You have 4 models listed on the server.
Do you have 4 listed in the Tasks tab?

It could be that each of your 2 processors is running 2 models, which will slow things down.

PJ, have you recently changed your general preferences regarding the % of CPU time the models are allowed to use? Most of us allow 100%. Check your preferences in your cpdn account from the menu here:

http://climateapps2.oucs.ox.ac.uk/cpdnboinc/index.php

Or have you recently changed your preferences in your account for any other boinc project?

I think you should also check what your Task manager shows as current CPU usage. Right-click on the digital clock, bottom right of screen & select Task manager. In the Performance tab, each processor should show near 100% usage. In the Processes tab, each model process is called hadcm3transum...... You should have two with this name, each running at 49-50%. Is anything else in the processes list using more than 1 or 2%?

---

PJ
You have 4 models listed on the server.
Do you have 4 listed in the Tasks tab?

It could be that each of your 2 processors is running 2 models, which will slow things down.

Hi Les,

2 of these models, those with 3 trickles or so were terminated by pebkac...

So actually only 2 models are trickling and on the machine.

I originally had my memory preferences set to use at most 50% of RAM and, although each process uses only about 100 MB of RAM, I am curious if the BOINC client tries to interpret the settings so that half the RAM on the machine (i.e. 500 MB) is actually free?

PJ, have you recently changed your general preferences regarding the % of CPU time the models are allowed to use? Most of us allow 100%. Check your preferences in your cpdn account from the menu here:

http://climateapps2.oucs.ox.ac.uk/cpdnboinc/index.php

Or have you recently changed your preferences in your account for any other boinc project?

I think you should also check what your Task manager shows as current CPU usage. Right-click on the digital clock, bottom right of screen & select Task manager. In the Performance tab, each processor should show near 100% usage. In the Processes tab, each model process is called hadcm3transum...... You should have two with this name, each running at 49-50%. Is anything else in the processes list using more than 1 or 2%?

Hi m.ov,

no, all processes are below the 2% watermark. Screensaver is inactive.

I actually went that far and associated the processes by hand to a single core in the task manager.

That improved the situation a bit and now the slower model has 6s/TS instead of 7s/TS and the faster one is at 6.3s/TS instead of 6.4s/TS.

Cool'n'Quiet is deactivated in the BIOS.

Hmm.

Is the machine overclocked or overheating?

If you run [http://www.mersenne.org/]Prime95's torture test[/url] for 24 hours, does it finish without errors?


One reason for a model suddenly slowing down is if it has 'rewound' because it thinks that there is something wrong.


Another reason for it slowing down is if the 'controller' thread (HadCM3_5.44_...) starts burning CPU time. The 'worker' thread (HadCM3_um_5.44...) is the one which is running the model. Sometimes this happens after you display graphics and the 'stop showing graphics' message from Boinc gets mislaid. If you reboot that usually fixes that particular problem.

---

Is the machine overclocked or overheating?

If you run [http://www.mersenne.org/]Prime95's torture test[/url] for 24 hours, does it finish without errors?


One reason for a model suddenly slowing down is if it has 'rewound' because it thinks that there is something wrong.


Another reason for it slowing down is if the 'controller' thread (HadCM3_5.44_...) starts burning CPU time. The 'worker' thread (HadCM3_um_5.44...) is the one which is running the model. Sometimes this happens after you display graphics and the 'stop showing graphics' message from Boinc gets mislaid. If you reboot that usually fixes that particular problem.

Hello MikeMarsUK,

the machine is running at the stock speed of 2.2 GHz per core. I think I´ll try the Mersenne test out.

Rebooting did not help. I already threw out some old drivers and installed a new CPU driver ( the delivered one with XP did not work well and the BOINC benchmark was broken as it seems. )

Right now the performance is at roughly 4.2s/ TS still around 33% lower than the average so far.

I tie each process to one of the CPU-cores in the task manager.

Right now the performance is at roughly 4.2s/ TS still around 33% lower than the average so far.

I tie each process to one of the CPU-cores in the task manager.

Peak performance was roughly 2.3s / TS on August 14th.

Right now the performance is at roughly 4.2s/ TS still around 33% lower than the average so far.

I tie each process to one of the CPU-cores in the task manager.

Peak performance was roughly 2.3s / TS on August 14th.

Looks like the effects are due to thermal throttling in the end.