MGLRU Sees New Performance Optimizations For Linux

It's been a while since there have been any new advancements or performance optimizations to talk about for Multi-Gen LRU (MGLRU) that was upstreamed to the Linux kernel two years ago as a very exciting kernel innovation. But that's changing now with some fresh performance optimizations being worked on for the MGLRU code.

MGLRU is an alternative LRU (Least Recently Used) implementation to optimize page reclamation and improve performance when experiencing system memory pressure. MGLRU has shown some great performance results already when experiencing memory pressure and with an upcoming Linux kernel release it looks like it will be performing even better.

Yu Zhao at Google who has led the MGLRU effort has posted a set of new performance optimizations for the code. The new MGLRU code has reworked its aging feedback, optimized deactivation, reworked type selection, reworked the refault detection, and other changes. With one of the patches is now up to a 57% decrease in the number of refaults while running the Python TPC-C workload. The optimize deactivation patch there is reduced LRU locking contention.

The patch set of six has been tested against Android, FIO, Memcached, multiple virtual machines, and with MongoDB for helping to evaluate the MGLRU performance optimizations. Hopefully the code review will go well and these MGLRU updates will make it into the mainline kernel for an upcoming cycle.