There is a proposal to replace the existing benchmark as run by the client with a 'standard BOINC computation' - to address "cheating" sometimes done by manual alteration of existing benchmark results. My recent experience suggests that even a 'standard BOINC computation' might not achieve "fairness" (i.e., same credit given for same work unit processed) among different hardware / software configurations.


I participate in several BOINC projects, but run only one such project per system. My 'rank' (i.e., total credit given) among project participants is not important to me. However, I had joined the Rosetta project using the "official" 5.2.13 Linux client. By a combination of circumstances, the same work unit was processed by me and by another participant, who had the SAME hardware as me, but who received four times as much credit for his work. [The suggested explanation was that he was using Windows and running an optimized client - therefore his benchmark figure was that much better.] When I joined the QMC project (different computer!), I deliberately used an "optimized" Linux 5.2.14 client. (This *does* give me an "unfair" advantage over unoptimized clients having a lower benchmark figure!)

My reason for posting here is that recently I saw "non-linear" QMC project computation times when comparing my system with someone elses. With a previous set of work units (I believe they ALL need a similar amount of work done) his system consistently took 11 hours to complete, but my system took 13 hours. With the latest set of work units, his system consistently took 33 hours to complete, but mine took 56 hours (which, combined with my higher benchmark figure, gave me MUCH more credit than him!).

Now, irrespective of the benchmark figure, in each system the ratio of benchmark figure (however calculated) to 'elapsed time to do the work' __ought__ to be consistent. Yet in his system (AMD X2, Windows) the newer work units took THREE times as long as the previous work units, whereas in my system (AMD Barton, Linux) they took more than FOUR times as long.

This experience leads me to believe that doing a 'credit per minute' calculation using a 'standard BOINC computation' would __still__ give non-linear results, depending on how_long/how_hard that computation took on a particular hardware / software system.
.

At this moment Seti are running a Beta project, with the initial aim of enhancing the seti application. But it is also testing a method of counting the floating point operations (Fpops) performed in crunching a unit and baseing claimed credit on that figure. This testing is not complete but so far each computer that crunches the same unit of work claims the same credits +/- 1% if it uses the Fpops method, (requires BOINC 5.2.6 or higher).

When this work is complete it is hoped that most of the projects will use a similar system, I believe Einstein and Rosetta are watching with interest, not sure on other projects.

Andy