With the coming on line of projects whose WUs take just minutes to complete (e.g., XtremLab, Leiden, HashClash, Tanpaku), I noticed an interesting behavior on MP systems.

As short WUs finish, the debts are recomputed often, resulting in other projects being paused. However, many projects check-point only every few tens of minutes, resulting in dilated processing times unless the WUs are left in memory.

As it's only a problem in MP systems, I suggest that each processor is treated individually. As a WU finishes, thus freeing up a processor, the debts are recomputed as usual, but rather to find a candidate project for the freed up processor only. The other processors are left alone processing whatever WU they were processing. The time to switch between applications would thus govern for how long a WU uses each processor individually.

Moreover, as NUMA systems proliferate, I'd also suggest that once a WU starts on a processor, it should not migrate to others unless there are no more WUs to process. Otherwise, if a WU starts on a processor and allocates memory on it, if it migrates to another processor, its processing time will be slower as it goes to another processor's memory.

HTH

---

Another solution to this problem is proposed here:

http://boinc.berkeley.edu/dev/forum_thread.php?id=739

but for some reason the idea seems not to be too popular (or hard to implement).

---