- The aMOF criterion was isol-01 being above 90%. This turned out to be too strict for many devices leading to overly pessimistic solutions or no solutions on lower end devices. Another problem is that it made the benchmark too sensitive to disturbances which may not be coming from the IO device. Relax the criterion to isol-05, which still provides a pretty strong guarantee while making the benchmark more robust.
- aMOF is one of the most important metrics that iocost-tune uses in determining the parameter solutions. However, because of the limited line fitting that it could do, it wasn't well utilized - e.g. some, usually high performing, devices show aMOF data points which peak in the middle, which fits the isolated-bandwidth solution perfectly. However, due to the limitations in line fitting, we couldn't detect this mid peak and instead relied on lat-imp which is a lot less reliable. Line fitting is improved to address this shortcoming.
- Fix bench merge mode rejecting sources spuriously on ignored sysreqs.
- Improve visibility on merge fails.
- Merge now only cares about the major and minor semantic versions and will try merging results from e.g. 2.1.0, 2.1.1 and 2.1.2 together by default.
- CHANGELOG added.
- bench study mode fix.
- iocost-tune solution criteria updated.
- resctl-bench added.
- resctl-demo initial release.