Fixes bug 1157261.
Since the majority of the work keystone does is cryptographic
calculations and filtering database records, keystone is CPU-bound.
Given that a keystone-all process has only one thread (i.e.,
eventlet's thread), keystone-all's throughput is limited to the
throughput of a single CPU core. To increase keystone-all's
throughput, we need to increase its CPU parallelism, which entails
running more keystone-all processes.
This patch adds two configuration options, public_workers=N and
admin_workers=N, that determine the number of keystone-all processes
that handle requests for keystone's public and admin WSGI applications
respectively.
Note that simply running keystone-all multiple times won't work
because care has to be taken for all of the worker processes to be
using the same socket (i.e., listen() then fork()).
DocImpact
Change-Id: If74f13bc2898e880649ee809967f5b5859b793c6
Co-Authored-By: Stuart McLaren <stuart.mclaren@hp.com>
3580c2af1b