Date: Mon, 18 Feb 2002 12:40:59 -0800 (PST) From: Matthew Dillon <dillon@apollo.backplane.com> To: Mike Smith <msmith@hub.freebsd.org> Cc: John Polstra <jdp@polstra.com>, hackers@FreeBSD.ORG Subject: ACPI timecounter tests original fast version vs masked version Message-ID: <200202182040.g1IKex936264@apollo.backplane.com> References: <20020204171736.A13719@hub.freebsd.org>
next in thread | previous in thread | raw e-mail | index | archive | help
Mike, did you get my email clarifying what the masking code does?
gettimeofday performance:
1 process / fast acpi timer 514000 calls/sec
1 process / dillon acpi timer 366284 calls/sec
(previously posted patch set)
1 process / dillon acpi timer 395400 calls/sec
with the mask <<= 2
instead of <<= 1
(one less bit of resolution)
2 processes with Giant completely removed from the path (userret and
gettimeofday): 417000, 291000, and 316000.
I don't have the performance for the original _slow code, sorry, but
I believe it is similar to the the second and third sample due to
it reading the timer three times at a minimum.
In anycase, I think this is reasonable especially if we explicitly
use the fast code for those chipsets known to be good. I would like
to commit it. It can handle *ANY* sort of ripple or fast-carry breakage,
really any type of breakage since it looks for two identical samples
after masking rather then using an inequality.
-Matt
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200202182040.g1IKex936264>