Re: Single Stack for all generators

On Wed, May 11, 2016 at 4:54 PM, Dmitry Stogov <dmitry@zend.com> wrote: > Hi, > > > Nikita, please review the patch > > https://gist.github.com/dstogov/06116f1610f45f8152ine3a9927c6c243ac > > It's the next attempt to use the single stack for all generators. > > Now I don't see any problems or BC breaks. > > > In case "yield" is used as an expression in context of function call. e.g. > var_dump(yield), engine backups and then restores active call frame chain. > > > Bob tested the patch on application that intensively uses generators, and > reported: > > 15% less L2 cache loads > > 20% less cycle stalls due to memory subsystem > > 3% more CPU_CLOCK_UNHALTED due to additional freeze/restore. > > > I'm going to commit this patch tomorrow. > The patch looks fine to me. From my own testing with a "Hello World" Aerys server, after multiple benchmark runs it previously started to mmap+munmap two new MM segments on each request, leading to a decline from 50k rps to 5k. With the patch this no longer occurs and the rps is stable at > 50k rps. (Though that may be more luck than anything.) Nikita