Accomplishments against goals:
============================
Researched AF_UNIX stat issue for Miguel.
Reviewed SRB.
Discovered that the global BMH lock is a significant performance block.
Continued to debug and tune turnstile locks.
Next Goals:
=========
Run scs to svn conversions to test new svn.
Check on the restarted 17.2 conversions in vostest.
Start other releases if it's going well.
19.3:
Add option for setting receive window to sftp server.
swd-1583
wtm-992
See if latest Windows 10 works with Teraterm Kanji
Add test case to regression tests.
Finish cleaning up the regression tests.
vos-6455
Aquarius
Help Pat and Otto with performance issues:
Implement Linear spin locks.
Test Kernel.
The optimization where we don't wake up all read
lockers if the last locker is still queueing may
be counter-productive in that we never get
back into the simple read locking mode.
Test with this change and start doing stress
runs.
Run Otto's disk performance tests on engineering kernel
as a test.
Incorporate turnstile lock changes:
as_dump_cache doesn't display some list headers correctly.
Performance Testing.
Investigate large process count, and spin tuning, first.
version 3 q_generation fixed (partition 5)
version 4 also aligned queue spin lock (partition 8)
Left 25usec running overnight on m131.
Left 25usec running overnight on m140
Peak performance in sort is keyed off 'Disk
Allocators',
which is hurting because the queueing comes before
the spin. Look into whether it can be a spin lock
(see efi>stile>make_bmh_lock_spi).
Split into multiple locks
Working on this in separate bmh workspace.
New lock calls at end of disk_allocators.pl1.
Populate speeds up and slows down (with matching DSU changes
Sort shows this behavior, too, but less pronounced.
CPU utilization and ops drops way off as # procs
reaches # cores.
Could be spin check for running proc?
Once reaches # cores, uses about 13.5 CPU secs in
lui no matter how many procs.