On 9/14/19 3:23 PM, Daniel Xu wrote: > Some workloads can require far more than 4K oustanding entries. For > example memcached can have ~300K sockets over ~40 cores. Bumping the max > to 32K seems to work pretty well. Should clarify that this is for poll entries, not actual pending IO. But makes sense, thanks, applied. We've got rlimit in place for mem usage, so there's really no reason to put anything else in place before bumping this (arbitrary) limit. -- Jens Axboe
