As you know the problem with the current system pls give credits based on a reference system (PC).

I didn't do that since the work units for a particular factorization can vary as much as 20-30% in the time they take to complete on a given reference system. I could simply choose an average and specify that as the credit per WU. Why would that be better than the current system based on FLOPS?

I didn't do that since the work units for a particular factorization can vary as much as 20-30% in the time they take to complete on a given reference system. I could simply choose an average and specify that as the credit per WU. Why would that be better than the current system based on FLOPS?

Overclaimed benchmarks (cheating).

Why not fix them on say 1.5 times what an average host claims? After randomly selecting 20 units I come up wih 12.5 credits / WU. In almost all cases higher than what the (from the looks of them non cheating) hosts claims, but not so high as to be unreasonable. Or to attract way too many crunchers or even the credit cops :).

My last 10 units completed at an average of 2959 sec/unit (a tad over 49 minutes each) and 9.78 credits. That's a measly 11.9 credits an hour. That's pretty low (compared to most other projects) for a quad-core w/ plenty of RAM running at 100%.

I'll finish up my current batch of WUs and check back every now and then to see if the credit system has improved.

It's currently set to grant what the client reports, within set minimum and maximums. The limits are set wide right now, minimum of 6 and max of 14. I'll take a look at the cross project stats and compare.

I didn't do that since the work units for a particular factorization can vary as much as 20-30% in the time they take to complete on a given reference system. I could simply choose an average and specify that as the credit per WU. Why would that be better than the current system based on FLOPS?

Besides the cheating problem as mentioned above, the BOINC benchmark system is badly broken. Some types of CPUs benchmark way too high and some are way too low compared with their actualy performance. There is also a large variance based on OS and BOINC version. The problem has gotten worse with time as the BOINC benchmark is determined by a short loop run on a single core, while in actual use most CPUs are now multi core. The inaccuracies get pretty ridiculous when HT is thrown into the equation since on many projects machines with HT activated slow to around 2/3 speed while the BOINC benchmark is run at full speed. If the variance is only 20-30% it seems the average should even out pretty quickly with a fixed credit system.

As WimTea suggested if you set an average and multiply it by some reasonable factor pretty much all the users will be happier, cheating will be nearly impossible and the BOINC credit guidelines should be fine. You'll get more users too as quite a few will not run projects that use the BOINC benchmark to determine credits. Everyone wins.

Thanks for listening...

Thanks for instituting the fixed credit system!