Subject: Re: RFC: migration to a fully dynamically linked system
To: None <email@example.com>
From: R. C. Dowdeswell <firstname.lastname@example.org>
Date: 12/30/2001 18:52:43
On this thread a few people have made some assertions about the
performance impact of shared object libraries, but we haven't
actually started discussing numbers. I thought that to start the
discussion, we should consider /bin/sh using lmbench on a couple
of machines both statically and dynamically linked.
I have run the experiment on two systems, my lapdog which is a Sony
Vaio PCG-Z505LS and an AlphaStation 200 4/233. A bit more information
about the machines:
Model: Sony Vaio PCG-Z505LS
CPU: Pentium III Mobile 750MHz
OS: NetBSD 1.5ZA
Model: AlphaStation 200 4/233
CPU: Alpha 21064A 233MHz
OS: NetBSD 1.4.2_ALPHA
Results of `/usr/pkg/bin/lmbench/lat_proc shell' a few times:
Lapdog: 4134.0000 5979.0000
Workstation: 37961.0667 59417.8889
So, for preliminary results, it would seem that there is a
potentially significant performance impact to this change
that should be examined.
1. Please note that I am not expressing either a desire for
a statically linked or a dynamically linked root partition.
2. I built the dynamically linked shell simply by:
$ cd src/bin/sh
$ vi ../Makefile.inc
$ make cleandir
$ make depend
$ make all
# make install
This is not necessarily indicative of the actual performance
of any of the proposed ideas.
3. The shell was built from a cvs checkout from a basically
random time [and different time] on both machines.
== Roland Dowdeswell http://www.Imrryr.ORG/~elric/ ==
== The Unofficial NetBSD Web Pages http://www.Imrryr.ORG/NetBSD/ ==
== The NetBSD Project http://www.NetBSD.ORG/ ==
== Ponte, Inc. http://www.ponte.com/ ==