Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watmath!clyde!rutgers!ames!cit-vax!mangler
From: mangler@cit-vax.UUCP
Newsgroups: comp.sys.misc,comp.periphs,comp.unix.wizards
Subject: Re: magtape, and what's hot
Message-ID: <2716@cit-vax.Caltech.Edu>
Date: Sun, 17-May-87 01:22:07 EDT
Article-I.D.: cit-vax.2716
Posted: Sun May 17 01:22:07 1987
Date-Received: Sun, 17-May-87 08:59:59 EDT
References: <878@wjvax.wjvax.UUCP> <242@uwslh.UUCP> <6567@mimsy.UUCP> <161@auvax.UUCP>
Distribution: na
Organization: California Institute of Technology
Lines: 32
Xref: utgpu comp.sys.misc:567 comp.periphs:355 comp.unix.wizards:2158
Summary: getting more throughput from your eagles

In article <161@auvax.UUCP>, rwa@auvax.UUCP (Ross Alexander) writes:
> ...how did you get your eagles to go so fast?

There are several factors affecting raw I/O throughput with Eagles.

First off, Emulex controllers are a millisecond faster than SI 9900's.
Assuming you have sector & index on the "B" cable, the Emulex will
always knows which sector is coming up; the SI 9900 has to find out
by selecting one drive and looking at the headers that are flying by.
(Like the Xylogics 450).  Notice the non-functional RMLA register.

Unfortunately, on controllers that do implement RMLA, the 4.2 and
4.3 BSD drivers throw away disk revolutions by incorrectly deciding
whether the requested sector is "coming up soon".  The 4.3 version
can be made correct by reversing the sense of the test, and swapping
all the sdist and maxdist numbers.  This makes about a 25% difference.

Surprisingly, tunefs makes little difference in dump throughput,
because of contention from 4.3 BSD dump's multiple processes.
In a disk-limited situation, the context switch rate will soar to
300 per second, as processes fight over one raw I/O buffer header.
An easy change to hpread(), allowing any of its raw I/O buffers to
be used with any of its disks, improves dump throughput by 25%.

Raw I/O has less per-kilobyte overhead in 4.3 than in 4.2 BSD.

Increasing dump's tape blocksize has the side-effect of improving
disk throughput, by making the read buffer larger.  This means
fewer disk reads broken across tape blocks, and fewer context
switches; a throughput improvement of about 15%.

Don Speck   speck@vlsi.caltech.edu  {seismo,rutgers,ames}!cit-vax!speck