Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!uwm.edu!cs.utexas.edu!sdd.hp.com!ucsd!ucbvax!hplabs!hplabsz!renglish
From: renglish@hplabsz.HPL.HP.COM (Bob English)
Newsgroups: comp.unix.aix
Subject: Re: Performance of AIX 1.2
Message-ID: <5864@hplabsz.HPL.HP.COM>
Date: 30 Aug 90 15:59:18 GMT
References: <1990Aug28.012042.21056@jdyx.UUCP>
Reply-To: renglish@hplabs.hp.com (Bob English)
Distribution: usa
Organization: Hewlett-Packard Laboratories
Lines: 19

In article <1990Aug28.012042.21056@jdyx.UUCP> shawn@jdyx.UUCP (Shawn Hayes) writes:
>The performance is being gauged by running the same benchmarks under OS/2 and
>AIX 1.2 on the same hardware.  Currently we are finding OS/2 to run about
>twice as fast as AIX 1.2.   Part of the problem is that for this test we
>want all writes to be posted to disk immediately so the test uses an open
>call with sync mode for each of the three files accessed.  

I'm not familiar with the details of OS/2, but AIX probably requires two
disk I/O's for every synchronous write to the file.  One of the writes
updates the data, and the other updates the inode to reflect the data
change.  Unless AIX 1.2 logs its inode updates to a separate spindle
than the data, this will cause two synchronous updates to OS/2's one,
and there's very little that can be done about it.

This, by the way, is one of the many reasons that dbms vendors build
their own filesystems and use raw I/O to access them.

--bob--
renglish@hplabs.hp.com	"I don't know how to speak for 90,000 people."