Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!uunet!tut.cis.ohio-state.edu!att!dptg!ulysses!andante!alice!andrew
From: andrew@alice.UUCP (Andrew Hume)
Newsgroups: comp.sys.sgi
Subject: anomoly in cross controller i/o rates
Keywords: file system i/o
Message-ID: <10639@alice.UUCP>
Date: 29 Mar 90 14:52:14 GMT
Organization: AT&T Bell Laboratories, Murray Hill NJ
Lines: 28


we bought a second smd disk controller for our 4D/240 and
(of course) tried to measure how much better things were.
the times are given below. the question is, why does
the i/o speed go DOWN as the buffer size increases for
the case of inter-controller i/o?

	andrew hume (andrew@research.att.com)

---------------------------------------------------------

timing tests for i/o through the file system (efs) on a 4D/240
with 2 Xylogics 754 controllers and Sabre disks. these show clock seconds for
copying a 40MB file 1) on a single disk, 2) between two disks on the
same controller, and 3) between disks on different controllers.

		c1d1->c1d1	c1d1->c1d3	c1d1->c2d0
cp		100		73		26
dd bs=4k	98		73		26
dd bs=8k	98		74		25
dd bs=16k	97		73		25
dd bs=64k	66		54		28
dd bs=125k	58		51		29
dd bs=500k	51		48		31
dd bs=2000k	47		44		40

times seem accurate to +- 1s. buffer pool flushed between trials.
(average of 3 single-user runs.)