[02:49:15] building and deploying PHP from source would allow PGO/FDO [02:51:14] I guess that's not incompatible with apt [02:55:56] the profiles would be pretty stable over time because the hot execution paths don't change [10:38:18] ori: https://research.facebook.com/publications/bolt-a-practical-binary-optimizer-for-data-centers-and-beyond/ [10:39:24] looks like a ~6.8% improvement is claimed for HHVM despite it not impacting the dynamically compiled bytecode [10:40:35] we'd need to measure first though IMO whether "[PHP] spends more time in statically compiled code than in the dynamically generated code" is also true in our environment [10:41:00] and right now we for sure don't have frame pointers for all dependencies and I'm not sure whether we have debug symbols [10:51:36] we could also consider a more aggressive -march given WMF's hardware lifecycle, but any benefits from that would also hinge on the above [14:17:12] what is the hardware lifecycle? what is the hw config of appservers now? [15:05:12] ori: s/app servers/wikikube workers/ [15:05:16] (I don't know) [15:28:48] according to SRE, it's 5 years [15:30:25] so targeting a higher x86-64 microarchitecture level could be feasible [15:30:50] but I'd be hesitant to do anything without data to justify it, and that data itself would need some upfront investment to collect [15:35:23] and most of it would probably be bogus until debian enables frame pointers, so probably until 2040 or so [18:44:50] updating the -march is probably sub 1% improvement at best [18:45:02] from experience