On 07/06/2018 06:28 PM, Al Viro wrote: > On Fri, Jul 06, 2018 at 03:32:45PM -0400, Waiman Long wrote: > >> With a 4.18 based kernel, the positive & negative dentries lookup rates >> (lookups per second) after initial boot on a 2-socket 24-core 48-thread >> 64GB memory system with and without the patch were as follows: ` >> >> Metric w/o patch neg_dentry_pc=0 neg_dentry_pc=1 >> ------ --------- --------------- --------------- >> Positive dentry lookup 584299 586749 582670 >> Negative dentry lookup 1422204 1439994 1438440 >> Negative dentry creation 643535 652194 641841 >> >> For the lookup rate, there isn't any signifcant difference with or >> without the patch or with a zero or non-zero value of neg_dentry_pc. > Sigh... What I *still* don't see (after all the iterations of the patchset) > is any performance data on workloads that would be likely to feel the impact. > Anything that seriously hits INCLUDE_PATH, for starters... I wrote a simple microbenchmark that does a lot of open() system calls to create positive or negative dentries. I was not seeing any noticeable performance difference as long as not too many negative dentries were created. Please enlighten me on how kind of performance benchmark that you would like me to run. Thanks, Longman
