4 ms·
One of the point in the article is that despite scheduler's importance and complexity there were no tools that allowed to evaluate the performance. This work co
by fpoling 11y ago
One of the point in the article is that despite scheduler's importance and complexity there were no tools that allowed to evaluate the performance. This work contributed such tools and those immediately allowed to identify quite a few bugs. Fixing those improves performance on NUMA systems rather substantially across various loads.
- sitkack 11y agoYou can't do science without measurement.
- kevan 11y ago>there were no tools that allowed to evaluate the performance. This reminds me of Brian Cantril's remarks in a talk (sorry, can't remember which one) about Solaris and DTrace. For the first time ever, DTrace gave them the ability to look into the live performance of low-level OS code and they found a lot of pathological worst-case behavior that no one suspected beforehand. No matter how good your team is it's really hard to accurately predict how a system will behave, measurement is key.