BOINC, Condor, CUDA and multiple tasks per GPU

In the old days this called for an app_info, so not so easy. Now, however, just go to your account at Einstein@home, select Einstein@home preferences, edit for the location (a.k.a. venue) of your host the following parameter:

GPU utilization factor of BRP apps

The value of 1.0 gives one active WU, the value of 0.5 gives two, 0.33 gives three. You are not likely to find more than three helpful (the big leap is often from one to two), and may not find it possible.

The current Einstein BRP ap requires rather a lot of support from the associated CPU task. Many of us have found total system output to be higher if we restrict BOINC to fewer than the maximum number of CPU tasks when running BRP on a GPU. The reduction in latency-imposed waiting of the GPU task in these cases increases productivity more than enough to pay for the loss of CPU work. The detailed tradeoffs are quite dependent on host characteristics, so experimentation is key. Happily the Einstein BRP jobs seem to be very consistent in computation requirement--so a small sample can be enough in many cases.