A Year Later, Speculative Page Fault Code Revised For Possible Performance Benefits


It’s been nearly one year already since the previous patch series working on speculative page faults for the Linux kernel were sent out for review. Fortunately, IBM’s Laurent Dufour has once again updated these patches against the latest code and sent them out for the newest round of discussions.

The simple summary is the set of 31 kernel patches can potentially improve concurrency for highly threaded processes. The improvement comes by handling user-space page faults without holding the mmap semaphore and in turn eliminating some waits within the page fault handler.

When using a “popular in memory multi-threaded database product”, IBM found the Linux performance with earlier revisions of these patches to be up by as much as 30% better in transactions per second. They are still testing these new “v12” patches but are hoping for a similar outcome.

More details via this patch message. Assuming the performance benefits pan out, hopefully it won’t be another year before seeing the next round of revisions or finding the code mainlined within the Linux kernel.