Have runners execute on dedicated cores
morehouse opened this issue · 3 comments
We would like to minimize the impact of runner bookkeeping operations (copying corpus, rsync, etc.) on fuzzing times. One way to approach this is to limit fuzzing to a dedicated VM core and use the other core for bookkeeping operations.
Assigning to David for now, but it's probably more important for him to get an end-to-end system running before he leaves.
I've done maybe a dozen experiments so far, but I've found very few benchmarks which manage to even finish one trial in reasonable time on an n1-standard-1 single core machine. This is highly suspect, given that these are benchmark which we expect to finish in 5~10 minutes.
By contrast, two core n1-standard-2 machines finish benchmarks in a more reasonable time, so I'm using them for now in my experiments. I'm still hoping to investigate this further
This is highly suspect, given that these are benchmark which we expect to finish in 5~10 minutes.
This isn't too surprising since we run with -workers=8 but the VMs have less than 8 cores.
Have had dual-core runners for a while now. One core is used for fuzzing while the other is used for bookkeeping.