On Thu, Aug 16, 2018 at 8:38 AM, Jeegn Chen <jeegnchen@xxxxxxxxx> wrote: > It is great to know anther way to turn the performance up. > > Instead of OSD, we verify the optimization on RGW side and remarkable > performance improvement do be observed. > > However, it is still be puzzle to us why memory log can impact the > performance so much. Thus our team spend some time to investigate it > further. Since memory log level reduction is the trigger, the first > suspect comes to us is the memory allocation mechanism. A more > detailed digest on the shows: Hello, Thanks for your testing! When taking a look on an OSD at my machine, I observe significant overhead coming just from crafting of the messenger's debugs (operator<< and related). getnameinfo() is substantial (but definitely not the sole one!) contributor. Also MOSDOp::print() (and several other things) played a role. I'm using async messenger. We could try to fight the overhead by e.g. introducing a cache in entity_addr_t. However, the simplest solution would be also the most performant one IMHO. The question is about acceptability of lowering the default. Regards, Radek
