Memory-management changes for CXL

Posted May 12, 2023 15:18 UTC (Fri) by joseph.h.garvin (guest, #64486)
Parent article: Memory-management changes for CXL

I don't understand the point of MMAP_EXMEM if it's just a single bitflag. If there are potentially an arbitrary number of NUMA nodes with wildly different access latency, it doesn't really narrow down where you want to allocate from. Seems like you either need to be able to pass an explicit node ID with every allocation or you need to change the policy right before every allocation to get that effect?

When I think about high performance apps I assume they are going to be doing their own management of anticipating what will be hot and cold and determining what tier of memory to use because they will have information the kernel doesn't.

Memory-management changes for CXL

Posted May 23, 2023 0:50 UTC (Tue) by balbir_singh (subscriber, #34142) [Link]

I would encourage Kim to look at https://lwn.net/Articles/713035/ and https://lwn.net/Articles/720380/. We've tried this in the past, but were held back due to HMM at that time. The idea of using a new MMAP is a bit too open, using madvise() and mempolicy() as alternatives might be better suited.