On standard PC, BOINC assign credits based on the average between the floating-point and integer performance of the machine according to a set of benchmarks performed by the client, regardless of the real performance of the application on the machine.

These benchmarks on the Cell processor are of course wrong because they do not use the SPEs. The way we assign credits takes into account these facts. We have benchmarked a scalar version of Cell MD on a reference machine based on a Opteron 2Ghz (see this paper).

This machine returns the following benchmark by the BOINC client:
CPU type AuthenticAMD
AMD Opteron(tm) Processor 146 [Family 15 Model 5 Stepping 10][fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 syscall nx mmxext lm 3dnowext 3dnow up]
Number of CPUs 1
Measured floating point speed 1707.25 million ops/sec
Measured integer speed 2969.64 million ops/sec

The average is therefore 2338.4 MIPS (million instruction per second) or equivalently 9.7 Cobblestone/hour (each unit of BOINC credit, the Cobblestone, is 864,000 MIPS). Cell MD is currently 16 time faster than this reference machine on the PS3, therefore we will credit 155.9 Cobblestone/hour on the PS3 for Cell MD.

This is in line with the SUSTAINED floating-point performance measured by us which for Cell MD is between 25 and 30 GFLOPS. In fact, the ratio between average number of operations and floating-point operations is 1.27 which gives between 132 and 158 Cobblestone/hour.

This credit system is in place only for the NEW workunits.

This sounds reasonable to me.

Looking at my log,
I had 15 finish between 18:52:48 and 19:08:23. So fixed credit is fine if they all finish within about the same time. But I did have 2 finish in 20:02:02 and 20:13:16 (1 hour longer). So if the credit is based on per hour I would get more on the longer running ones, which is only fair.

I have one that shows only 11:33:43 hours, but I know that was when I powered down my PS3 to add some fans and it had about 7:30 at the time, so total would be 19 hours.

So my question is, have you solved the problem so that when work is interrupted that the total time be reflected and not just the time since last startup ?

This sounds reasonable to me.

So my question is, have you solved the problem so that when work is interrupted that the total time be reflected and not just the time since last startup ?

No, in fact I don\'t know why it does that, but we would like to fix it. Any idea?

Regarding the amount of credit/workunit. Our simulations should always take the same time to finish. We estimate this before giving the wu out.

In principle the credits should not depend on the time used to compute it, but on the number of operations performed. If you are using your machine while computing results it will take longer to complete. However, the total number of operations performed does not change.

For the calculation time trouble, I think it\'s a bug in application. I\'d already suspended calculation before restart under Sony OS, and when I came back, I launched BOINC always under suspended calculation mode. I clearly view the good time on the current WU line. When I released calculation, the first screen refresh shown a restarted from 0 time.

---

No, in fact I don\'t know why it does that, but we would like to fix it. Any idea?

Not exactly.

I suspect you need to save the time spent when it checkpoints or is suspended to disk, then read that data back in and start from there on a resume.

I kind of remember something like this on the developer email list.

Your best choice would to be to ask the question there, I\'m sure someone already knows the answer.

Regarding the amount of credit/workunit. Our simulations should always take the same time to finish. We estimate this before giving the wu out.

In principle the credits should not depend on the time used to compute it, but on the number of operations performed. If you are using your machine while computing results it will take longer to complete. However, the total number of operations performed does not change.

OK, I understand, but I was thinking the time reported was CPU time for the application and not \'wall clock time\' which would include any delays caused by other processing.

I must have had something else running at the time, but not usually. I mostly just let the PS3 run untouched.

...therefore we will credit 155.9 Cobblestone/hour on the PS3 for Cell MD.

This credit system is in place only for the NEW workunits.

I\'m not sure if there is a problem or I do not quite understand this.

I assume this meant, the new work units (5.17) would be granting more credit, because you recalculated the equivelant, You also mentioned they were longer running, so I thought this also meant more based on per hour.

I processed a 5.17 result ID#5574 which ran 21:44 hours and received the same credit ?

5574 4276 13 Jun 2007 6:53:36 UTC 14 Jun 2007 9:54:24 UTC Over Success Done 78,243.82 78.14 1,875.00

So I am confused.

So I am confused.

The wu your are referring to is still an old one. You probably had in your BOINC client queued. The one you are running now is the new one. Other users are receiving the right new amount of credits (3375)

gdf

OK Thanks

Now I\'m not confused. ;)