Closed rzinsly closed 2 years ago
I rebased on top of develop and added code to decrease avg_delay when we choose software, otherwise after the switch we stop measuring time and never go back to NX.
I split the threshold values between compression and decompression as decompression with software is faster. Also added a multiplier for baremetal as the latency is higher there. I choose higher thresholds trying to be on the safe side and just avoid starvation instead of trying to get the best performance. The PR is open for review but I'll wait for the resolution of #169 before merging it.
The code looks good, but I agree the second commit is not necessary with the changes from #169.
I rebased the code and removed commit 621bac4.
Measure the average time spent on NX jobs and fallback to software if it is greater than our threshold. I'm keeping it as draft for now because I'm not sure which value should be used as threshold yet. This seems to solve #152 as well.