At present, if a host has GPUs from multiple vendors installed, it will request work from both vendors. If the client has a cache set such that the it is larger than twice the max_wus_in_progress_gpu setting, the user will only get tasks for one GPU. The other will remain idle.
In the example below, the server is set to use max_wus_in_progress_gpu of 60. The host has 120 WUs downloaded already but all 120 are for CUDA. So, even though the AMD GPU is idle, it won't download any work. Collatz Conjecture | 3/7/2014 10:22:43 AM | Sending scheduler request: Requested by user. Collatz Conjecture | 3/7/2014 10:22:43 AM | Requesting new tasks for NVIDIA and ATI Collatz Conjecture | 3/7/2014 10:22:43 AM | [sched_op] CPU work request: 0.00 seconds; 0.00 devices Collatz Conjecture | 3/7/2014 10:22:43 AM | [sched_op] NVIDIA work request: 14315.74 seconds; 0.00 devices Collatz Conjecture | 3/7/2014 10:22:43 AM | [sched_op] ATI work request: 86400.00 seconds; 1.00 devices Collatz Conjecture | 3/7/2014 10:22:47 AM | Scheduler request completed: got 0 new tasks As a temporary "fix" I've increased the max_wus_in_progress_gpu and also told the user to reduce his cache. A permanent fix would be to enforce the max_wus_in_progress_gpu by GPU type so that if a user has two nVidia and one AMD GPU in a host, he would get 66% CUDA and 33% AMD WUs so that no device would be left idle. _______________________________________________ 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.
