Lynn W. Taylor wrote:
> There is a small catch-22. > > If you base cache size on speed, and you measure speed using only the > live application, then you can't fill the cache until you've done some > work. > > If you assume the machine is fairly fast, and it isn't, you run the risk > of grabbing an initial task that couldn't possibly finish on time. > > If you assume the machine is slow, and it's really fast, you run the > risk of being denied work because BOINC thinks the machine is too slow. > > So you need the benchmark as a rough initial measurement of speed. > > Emphasis on the word "rough." It can be off by a factor of five and > still work. There would actually be some benefit in re-thinking this. At the moment, on attaching to a new project, the BOINC client makes an initial request for a nominal 1 second of work, and then immediately proceeds to fill the predicted cache on the basis of benchmark and TDCF = 1.000000 But as you pointed out to Martin, it would be better if the project could dial-in automatically (fpops_est, in this case, rather than credit - the scientists don't want to, and already don't, figure that one manually either). If, on initial connection, the client requested 1 second of work (repeated if necessary for additional CPUs or GPUs), and then stayed with that number of tasks until something like 95% complete / 5 minutes remaining, future work requests could be based on a reasonable first stab at a full-run estimate. By that stage in processing, too, it should have some clues about TDCF: it's crazy that when AQUA recently issued tasks which turned out to need a TDCF of over 100, BOINC doggedly stuck to its 'new project' default of 1.000000 right up until the first task completed. Then panic set in! Of course, this scheme would be broken by badly-behaved non-linear science apps, but they would break the current 'fill cache based on extimate' scheme too. _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
