Date: Mon, 1 Jul 2002 20:14:46 -0700 (PDT) From: Matthew Dillon <dillon@apollo.backplane.com> To: Julian Elischer <julian@elischer.org>, "Long, Scott" <Scott_Long@adaptec.com>, FreeBSD current users <current@FreeBSD.ORG> Subject: -current results (was something funny with soft updates?) Message-ID: <200207020314.g623Eke5038019@apollo.backplane.com> References: <Pine.BSF.4.21.0206281233500.75410-100000@InterJet.elischer.org>
next in thread | previous in thread | raw e-mail | index | archive | help
SMP builds are still producing panics every 2-4 buildworlds after the
KSE commit, I'm still trying to track that down. But I was able to
complete the softupdates/non-softupdates test with a UP build of
-current:
with softupdates (UP BUILD, CURRENT):
3122.30 real 2360.70 user 532.54 sys
3083.17 real 2361.14 user 529.53 sys
3085.05 real 2361.59 user 529.32 sys
without softupdates (UP BUILD, CURRENT):
3361.70 real 2365.23 user 535.50 sys
3451.55 real 2368.22 user 537.26 sys
3454.85 real 2369.42 user 536.69 sys
^^^^^
~350 second dif note user times
for real.
Included below are the original tests that were done under stable... the
overall 'real' times are NOT COMPARATIVE since the original tests were
done with an SMP build, but the user times should be, and that is where
I believe the major difference is occuring. My guess is that the new
GCC in -current is eating a massive amount of extra cpu. It is eating
over 2300 cpu seconds under -current and only 1400 under -stable while
system time remains roughly comparable (remember that the interrupts
are not charged to processes under -current).
The difference in real time softupdate vs non-softupdates between
current and stable is around 350 seconds under current, and 889
seconds under stable. This is fairly comparable when we consider that
a good portion of the extra user time eaten in -current is absorbing
the stall delays for processes undergoing I/O that softupdates mostly
fixes.
My conclusion is that softupdates is working fine and (A) the new GCC
is a whole lot less efficient then the old GCC and (B) user times are
masking gains (due to high parallelism) that would otherwise be
realized with softupdates.
: (original tests under -stable)
:test1# cat x1 (SMP BUILD, STABLE, WITH SOFTUPDATES)
: 1497.09 real 1397.98 user 612.06 sys
: 1500.12 real 1399.33 user 609.79 sys
: 1494.82 real 1398.30 user 612.46 sys
:test1# cat x2 (SMP BUILD, STABLE, WITHOUT SOFTUPDATES)
: 2449.14 real 1401.34 user 625.54 sys
: 2389.75 real 1400.38 user 629.86 sys
: 2358.82 real 1403.26 user 624.93 sys
:
( ~889 second difference in real time)
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200207020314.g623Eke5038019>