Hi, Nathan Zimmer found that once we get over 10+ cpus, the scalability of SPECjbb falls over due to the contention on the global 'epmutex', which is taken in on EPOLL_CTL_ADD and EPOLL_CTL_DEL operations. Patch #1 removes the 'epmutex' lock completely from the EPOLL_CTL_DEL path by using rcu to guard against any concurrent traversals. Patch #2 remove the 'epmutex' lock from EPOLL_CTL_ADD operations for simple topologies. IE when adding a link from an epoll file descriptor to a wakeup source, where the epoll file descriptor is not nested. Patch #3, restores the size of 'struct epitem' to <= 128 bytes, which was broken in Patch #1. Its a bit hacky, so I decided to break it out as a separate patch for review purposes. Performance of SPECjbb improves considerably. From Nathan Zimmer's testing. Thread: http://marc.info/?l=linux-kernel&m=137908766013329&w=2 " On the 16 socket run the performance went from 35k jOPS to 125k jOPS. In addition the benchmark when from scaling well on 10 sockets to scaling well on just over 40 sockets. I should also note there system responsiveness of various commands was quite improved when under the full load of the benchmark. ... Currently the benchmark stops scaling at around 40-44 sockets but it seems like I found a second unrelated bottleneck. " Jason Baron (3): epoll: optimize EPOLL_CTL_DEL using rcu epoll: Do not take global 'epmutex' for simple topologies epoll: restore 'struct epitem' size fs/eventpoll.c | 150 ++++++++++++++++++++++++++++++++++++++++----------------- 1 file changed, 105 insertions(+), 45 deletions(-) -- 1.8.2 -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
