Bug tracking in the linux development enviroment
SMP, highmem and NUMA are all becoming increasingly common, on servers and high-end workstations. This paper examines some of the VM issues encountered as we push the limits of what Linux can do on larger systems, and what can be done to address them. The target audience is anyone
interested in the stability and performance of Linux servers that have more than 1 CPU (or hyperthreaded), or more than 1GB of RAM. It is pitched at anyone familiar with kernel internals, not just VM hackers.
Subject areas covered will include virtual address space exhaustion (how to avoid masses of shared global data - either by locality, temporary mapping windows, or elimiation of bloat), the creation and use of a per-process kernel space (uarea / UKVA), and efficient cache usage (TLB and data/instruction cache).
Also covered will be the performance and space consumption problems of the pte-chain based rmap mechanism, an object based approach, and the benefits of a hybrid scheme using both mechanisms. Finally, the mechanisms used for efficient NUMA support (eg local allocation and replication).
