DU4.0's mmap() call

From: Sam Nelson <Sam.Nelson_at_cs.stir.ac.uk>
Date: Wed, 16 Dec 1998 13:08:37 +0000 (GMT)

I recently ported a Perl-database/WWW application from its original home on
Solaris 2.6 to a DU4.0b box, the idea being that the DU box in question is
substantially more powerful and will hence do the job much more quickly.
The port itself was troublefree. Recent Perl, recent Berkeley DB, recent
Apache, a couple of add-ons---took about a day, including interruptions.

Query performance on the database in question has indeed improved, and the new
system looks a lot snappier than its predecessor. However, the database reload
(which happens from scratch two or three times a week) is far slower. On
Solaris 2.6, the database reload took 15/8 mins real/CPU time, typically. On
DU4.0b, with a 500MHz CPU rather than 300MHz, the reload takes 46/10 mins
real/CPU time.

The only idea I've come up with so far is the notion that the implementation
of data access of FS objects mapped with mmap() must be substantially inferior
on DU as opposed to Solaris (Berkeley DB gets at its database files, where it
can, with mmap). Can anyone confirm or deny this hypothesis, and, assuming
it's correct, will any particular patch/patch-cluster/update help matters?

Database reloads, of course, can be done unattended overnight, so a
performance problem like this isn't mission-critical. Even so, we use Perl
DB stuff extensively around here, and there could well be other apps for which
a move to or implementation on DU4.0 would be a bad move.

Thanks for any hints,

Sam.
--
Sam Nelson, Comp Sci, Stirling U, FK9 4LA, Scotland         ,->0->M
Email: sam_at_cs.stir.ac.uk    Phone: +44 7050 165499     I->3-+->2->R=->-+->4->O
Office: +44 1786 467443      Fax: +44 1786 464551           `->1->S=->-'
URL: http://www.cs.stir.ac.uk/~sam  NS80799622/38m   R$+_at_$+  $_at_smtp$#$2$:$1_at_$2
Received on Wed Dec 16 1998 - 13:09:31 NZDT

This archive was generated by hypermail 2.4.0 : Wed Nov 08 2023 - 11:53:38 NZDT