Date: Mon, 4 Jun 2001 12:48:16 -0400 (EDT) From: Andrew Gallatin <gallatin@cs.duke.edu> To: freebsd-alpha@freebsd.org Subject: interrupt latency Message-ID: <15131.48080.899655.349197@grasshopper.cs.duke.edu>
next in thread | raw e-mail | index | archive | help
OK, so we do suck relative to Tru64
Here are some timings taken using the interperf tool and a PCI Pamette
(PCI Pamette V1R1, Firmware 1.10, Serial Number 200040, Configuration:
4010E 4010E 4010E 4010E).
Intrperf - Interrupt Performance Measurements
The application programs lca0 to issue an interrupt via the Ring0 line.
It records the time at which the interrupt was issued, cleared
by the Pamette driver (i.e., kernel code) and cleared by
the user code. All these times are recorded in PCI-cycles, so there
is no need for platform specific cycle counters. It works under DIGITAL
Unix on Alpha and under Windows NT on Alpha and x86.
(and on FreeBSD/alpha and FreeBSD/i386)
(To get the user timings, the driver sends a sigio to a user process.)
These were taken on a 500 MHz Miata (GL) with both Tru64 and 4.3:
issue -> system issue -> user system -> user
=================================|======================|======================
Tru64 4.0D:
min (us) 3.5 39.7 34.2
max (us) 12.7 67.6 61.4
avg (us) 5.4 42.7 37.3
med (us) 5.5 40.8 35.3
FreeBSD 4.3-RELEASE
min (us) 4.8 53.7 48.2
max (us) 11.2 300.0 293.4
avg (us) 6.4 59.3 52.9
med (us) 7.7 58.4 50.8
I'll try to do some -current timings later. They're *really* going to
suck, because you can't call psignal() from a fast interrupt handler in
-current.
Drew
To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-alpha" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?15131.48080.899655.349197>