[1587] in Release_7.7_team
IRIX 6.5 install and on-line documentation
daemon@ATHENA.MIT.EDU (Robert A Basch)
Fri Jan 15 12:27:29 1999
To: release-team@MIT.EDU
Date: Fri, 15 Jan 1999 12:27:22 EST
From: Robert A Basch <rbasch@MIT.EDU>
I've done some analysis to see if there's a feasible way to reduce the
IRIX 6.5 install time. On an O2, the full install is now taking a little
over two hours (about 2:10), as compared to about 1:45 for IRIX 6.3. On
an R5000 Indy, it's over 3 hours, compared to about 2:30 for 6.2. In
other words, install times that were already unacceptably long stand
to be even longer. (The update will presumably take longer than the
install again, since it has the extra step of having to remove
existing packages).
As in the current releases, a big chunk of time is spent in "exit
commands", invoked by the installation tool after package files have been
installed in the target, to do various post-install configuration tasks.
This seems to be worse now, probably because the install is split into
more passes. At any rate, I examined the time spent executing these
commands, and found that there are some big wins to be had, if we're
willing to live with the downsides.
1) The single biggest culprit is due to the installation of the
on-line books packages, as exit commands are invoked to build a
searchable index to these books. When I removed the books packages
from the install, 45 minutes was taken off the O2 install time.
To make the books accessible, we would minimally have to symlink a
couple directories to the os volume; if we wanted to allow private
machine owners to install books packages, we'd have to symlink the
individual files. I think it is pretty doubtful that not installing
these books packages could present package conflict problems for
future updates, but it would still violate the desired goal of
having an accurate and complete inst database.
2) By skipping the installation of man page and release note packages,
we could save another 15 minutes or so. The issues here are similar.
[In deciding about these documentation packages, we should also
consider that the 6.5 infosearch program seems to require an http
server. The tools for accessing the individual components, e.g.
man and insight, are still filesystem-based.]
3) Another 15 minutes or so is wasted by doing repeated make's in
/usr/lib/filetype, to build the IRIX file type database. This is an
expensive process, which is made worse here because many of the files
are linked to AFS. Unfortunately, there is no good way to avoid
doing all these make's. If we're willing to consider a bad way to
avoid it, I've found that we can rename the Makefile after the first
pass installs it, then restore it at the end of the install and invoke
the 'make' ourselves.
By doing all of the above, the current install time on the O2 goes down
to just under an hour; on the R5000 Indy, it's a little over an hour.
Even the R4600 Indy installed in 1:30.
We can discuss these issues at the next meeting; in particular, we
should focus on how to deal with web-based documentation issues in
general (whether we should maintain our own doc web server, or rely
on the vendor server, etc.). My understanding is that there are
similar issues for Sun machines.
Bob
