Sent out today as a request for comments (RFC) by a Linux kernel engineer employed by IBM is a machine learning library for the Linux kernel. The intent is on plugging in running ML models to the Linux kernel that could be used for system performance optimizations and various other purposes.
Viacheslav Dubeyko who is employed by IBM as a Linux kernel developer sent out an RFC patch series today that aims to introduce a machine learning library for the Linux kernel.
It’s wild that one of the problems is that the kernel doesn’t use floating point numbers.
It isn’t really so strange that they aren’t used, if you think about it. Floating point is subject to fuzziness in the last several digits, and you can’t guarantee that a given value is going to round the same way when you’re dealing with multiple arches (or even multiple versions of what’s nominally the same arch, since optimizations change over time). Undefined behaviour is nasty. Floating point is useful for many things, but I’d keep it out of a cross-platform system kernel unless I liked hard-to-diagnose bugs.
@nyan @Gobbel2000 This, totally. But then if you are going to throw unconstrained machine learning in there you are going to run into hard-to-diagnose bugs.