That's really interesting, as I brought one of our services at my old job over onto HHVM and gained 4 times more performance, I'm really curious what workload you were throwing at it, and I know that the HHVM people would love to take a look at it; performance regressions from PHP to HHVM have been basically nil for me over the past 12 months.
Re:Coming Soon in HHVM
by andrewchambers on Wednesday February 18, 2015 @12:39AM
(#9067176)
Did you run enough requests to allow the jit to work? some jits require thousands of runs (in this case requests) to warm up.
I would hope ab -c 5 -n 1000 would be enough. I've used -c 5 because the VM wasn't exactly beefy, but that's still quite a few requests.
Re:Coming Soon in HHVM
by fredemmott on Wednesday February 18, 2015 @01:22AM
(#9067267)
do you have any profiling information to share? if you could post it as a github issue, we'd be very interested. This sounds unusual.
Re:Coming Soon in HHVM
by fredemmott on Wednesday February 18, 2015 @04:57PM
(#9071232)
also, if this is an open source workload, any chance you could submit it to https://github.com/hhvm/oss-performance ? We generally optimize based on those benchmarks, and Facebook itself.
However, HHVM was generally between 2 and 5 times slower than vanilla PHP at rendering pages. In micro benchmarks it was way ahead, but running the whole application was a very different matter.
I decided to wait until the thing matures a bit more before trying again and then hopefully with more time to investigate the causes of this huge difference in performance.