3 ms·

KML: A machine learning framework for operating systems and storage systems

43 points by i_umit 5y ago

gigalord 5y ago
Whenever I see buzz around “ML for Computer Systems” my first question is always “how debuggable is this?” A major benefit of the sophisticated heuristics and algorithms that power modern systems is the legion of experts who deeply understand how they work. When we see poor performance, a bit of digging and asking around will usually unearth some principled reasoning about how we can modify our programs to work well with the existing heuristics (think branch prediction). Unless we are careful, we lose this with an ML-based approach
- Q6T46nT668w6i3m 5y ago
  For sure, I agree that interpretation is critical for computing systems like operating systems and compilers. I have been extremely pleased that most of the recent popular SysML research readily trades correctness for understandability. Likewise, I have also been pleased that correctness is often traded for performance.
- servytor 5y ago
  Eh. I remember reading the furor about the Linux scheduler[0], and that was by the hands of man, my man. [0]: https://people.ece.ubc.ca/sasha/papers/eurosys16-final29.pdf https://people.ece.ubc.ca/sasha/papers/eurosys16-final29.pdf
- i_umit 5y ago
  I think this is a fair concern. I agree that the ML approaches can cause explainability and traceability issues but not all of them. We also support ML approaches like decision trees which can be considered more "debuggable". However you might not be able to solve every single problem with decision trees just because it is "debuggable". Besides, there are a lot of researchers are working on ML explainability. Our vision is simply building self-adaptive tuning systems which may not require constant development efforts to optimize OS and storage components for ever-changing workloads and new devices. You can find more discussion about this topic in the paper too.
omneity 5y ago
Remember, Skynet is Genisys. Joke aside, I’m not sure whether the tradeoff between “automatically tune parameters” and “nightmare to debug” is worth it. On the other hand it seems patterns like this are already deployed in hardware so maybe it will work out fine.
i_umit 5y ago
You can find the related paper in arxiv: https://arxiv.org/abs/2111.11554 https://arxiv.org/abs/2111.11554

43 points by i_umit 5y ago

Whenever I see buzz around “ML for Computer Systems” my first question is always “how debuggable is this?” A major benefit of the sophisticated heuristics and algorithms that power modern systems is the legion of experts who deeply understand how they work. When we see poor performance, a bit of digging and asking around will usually unearth some principled reasoning about how we can modify our programs to work well with the existing heuristics (think branch prediction). Unless we are careful, we lose this with an ML-based approach

Q6T46nT668w6i3m 5y ago
For sure, I agree that interpretation is critical for computing systems like operating systems and compilers. I have been extremely pleased that most of the recent popular SysML research readily trades correctness for understandability. Likewise, I have also been pleased that correctness is often traded for performance.
servytor 5y ago
Eh. I remember reading the furor about the Linux scheduler[0], and that was by the hands of man, my man. [0]: https://people.ece.ubc.ca/sasha/papers/eurosys16-final29.pdf https://people.ece.ubc.ca/sasha/papers/eurosys16-final29.pdf
i_umit 5y ago
I think this is a fair concern. I agree that the ML approaches can cause explainability and traceability issues but not all of them. We also support ML approaches like decision trees which can be considered more "debuggable". However you might not be able to solve every single problem with decision trees just because it is "debuggable". Besides, there are a lot of researchers are working on ML explainability. Our vision is simply building self-adaptive tuning systems which may not require constant development efforts to optimize OS and storage components for ever-changing workloads and new devices. You can find more discussion about this topic in the paper too.

omneity 5y ago

Remember, Skynet is Genisys. Joke aside, I’m not sure whether the tradeoff between “automatically tune parameters” and “nightmare to debug” is worth it. On the other hand it seems patterns like this are already deployed in hardware so maybe it will work out fine.

i_umit 5y ago

You can find the related paper in arxiv: https://arxiv.org/abs/2111.11554 https://arxiv.org/abs/2111.11554