I have noticed GPU utilization around 70% (reported by nvidia-smi) while running Long Runs workunit. Also the gpu was not heating properly and duration correction factor rised above 2. In addition to gpugrid I run WCG on my 4CPUs.

Solution is to raise priority of the gpugrid process, using command:

sudo chrt --rr -pa 5 <pid>

This command moves the process to SCHED_RR scheduling class, which is, in simple terms really high prioriy but lower than realtime. Unlike SWAN_SYNC this uses around 3-5%CPU on my machine. I was able to achieve only small improvement by decreasing the nice(2) value alone.

The application also causes considerably more PCIe load than other boinc gpu projects.