Path: utzoo!attcan!uunet!ibmarc!drake
From: drake@ibmarc.uucp (Sam Drake/99999999)
Newsgroups: comp.sys.ibm.pc.rt
Subject: Re: AIX - How is it? (Summary - long)
Message-ID: <972@ks.UUCP>
Date: 30 Jul 89 08:16:58 GMT
References: <557@limbic.UUCP> <29176@beta.lanl.gov>
Sender: news@ibmarc.UUCP
Reply-To: drake@ibmarc.UUCP (Sam Drake)
Distribution: comp
Organization: IBM Almaden Research Center, San Jose
Lines: 49

In article <29176@beta.lanl.gov> dd@beta.lanl.gov (Dan Davison) writes:

>The only thing you need to know about AIX is that they swap the
>*kernel*.  Can you say geologic time to run things?  I knew you
>could.

As an IBMer (but not an IBM spokesman), perhaps I shouldn't respond
to this, but here goes anyway.

The implied statements here are that (1) AIX/RT is slow, and that (2)
it's slow because the kernel is swapped.  I don't believe that either
one of these statements is especially accurate.

I won't use this append for sales purposes, but will simply comment that 
AIX/RT systems have been benchmarked in the UNIX magazines, and the numbers
look fine to me...read 'em yourself and judge.

It is certainly true that much of the AIX/RT kernel is pagable.  The 
lower level VRM interface, on which the AIX system resides, is permanently 
resident and never paged.  Claiming that this leads to a performance 
problem is not accurate, in my opinion.

Why should all pages of the kernel be permanently resident in storage?
If a particular page of the kernel is the Least Recently Used page in the
system, why leave it in storage?  To leave it in, though it's not been
referenced for hours, means that system performance will suffer, since
pages that have been recently used are paged out to disk instead of this
more deserving page.  Is it really important that the text of the 
kernel panic handler be kept around, hogging up real memory, when in reality
it will be referenced far less than once per MONTH?  I don't think so.
Granted, if a kernel panic occurs, one more disk I/O will be required
before the system abend dump can be taken, but I think that's a fair tradeoff
for having that page available to user processes for all those thousands 
of seconds.

If main memory is so badly over-committed that useful pieces of the kernel are
being paged, then the situation is of course not as obvious.  But even in
the case of such over-commitment the principle of paging the Least Recently
Used pages first is a winner.  If it's been 20 seconds since the serial port
device driver has been used, get it out of there!  Use the memory for running
more vital things (like the user's application).

Many if not most operating systems have a resident portion and a pageable
portion.  All do it because LRU beats non-pageable kernels.  A non pageable
kernel is certainly easier to write, but it's not a performance boon ...
quite the opposite.

Cheers,  ...Sam Drake / IBM Almaden Research Center
Sam Drake / IBM Almaden Research Center