On Sat, Feb 08, 2025 at 05:40:41PM +0000, Lorenzo Stoakes wrote: > > > > We tried THP around 2012 and rejected it. The latency tail became a lot > > longer and fatter. Various things have changed that might make THP less > > bad today, but I am not aware of anyone reevaluating it. > > A _lot_ has changed. Try it again :) > > > > > I think the problem with THP was the mmap_sem. Given a heavily threaded > > process, the mmap_sem tends to be the one dominant lock in the kernel. > > A lot of work has been done on reducing mmap_sem contention. Again, worth > another shot ;) Probably worth it, agreed. But unlikely to happen in the near term for various reasons. One of the biggest improvements was actually userspace. Glibc malloc is very eager to call mprotect for reasons that imo never made any sense. Replacing glibc malloc with any sane allocator will significantly reduce the number of mmap_sem taking system calls. Another hard-to-avoid problem is all the variations of "ps" that take the mmap_sem for every process in the system. Jörn -- You can't do much carpentry with your bare hands and you can't do much thinking with your bare brain. -- unknown
