Chris Mesterharm wrote: > > Hello, > > I was hoping someone could suggest an area within the Linux kernel > that could benefit from research in machine learning. People have suggested VM, but frankly, I think the issues with the Linux VM subsystem are fairly well-understood, and the main reason that they aren't dealt with is that dealing with them will require significant changes to the VM architecture, changes that Rik and others are working on. For example, when the machine starts to thrash, there's just no way you're going to be able to be fair to all processes, so some of them will have to be swapped out and made to hibernate until things even out; Linux currently doesn't do that (well, last time I looked, anyway). However, one place where ML could, IMO, have a significant impact, is on the page-replacement policy. Any method that lets us make more accurate choices about what pages to read in from disk and what pages to throw away, will be a win, because it will postpone the thrash point. For example, if you have an algorithm that, over many runs of a program, makes weighted associations between the program inode and text and data pages that it uses, you may be able to make very accurate page-replacement choices *without* having to let pages filter all the way through the page cache. In other words, you may be able to implement a policy much closer to "Not Going To Be Used" than the current "Not Recently Used". That could significantly improve VM performance, and may have a positive influence on filesystem performance as well. Of course, that's all speculation. But might be worth a look. Cheers, -- Joe "I should like to close this book by sticking out any part of my neck which is not yet exposed, and making a few predictions about how the problem of quantum gravity will in the end be solved." --- Physicist Lee Smolin, "Three Roads to Quantum Gravity" -- Kernelnewbies: Help each other learn about the Linux kernel. Archive: http://mail.nl.linux.org/kernelnewbies/ IRC Channel: irc.openprojects.net / #kernelnewbies Web Page: http://www.kernelnewbies.org/
