NetNews Usenet Archive 1992 #31/NN_1992_31.iso/spool/comp/os/os2/advocacy/10984

Internet Message Format | 1992-12-27 | 4.4 KB

Path: sparky!uunet!gatech!concert!ais.com!bruce
From: [e-mail] (Bruce C. Wright)
Newsgroups: comp.os.os2.advocacy
Subject: Re: OS/2 bigot meets NT....
Message-ID: <[e-mail]>
Date: 27 Dec 92 19:06:44 GMT
References: <[e-mail]> <[e-mail]> <[e-mail]> <[e-mail]>
Organization: Applied Information Systems, Chapel Hill, NC
Lines: 72

In article <[e-mail]>, [e-mail] (Marc Fraioli) writes:
> In article <[e-mail]> [e-mail] (Bruce C. Wright) writes:
>>In article <[e-mail]>, [e-mail] (Joseph C. Pallante) writes:
>>> I have a question....
>>>
>>> All this debate is over:  Many, Many crashes because a PC has only
>>> 8 megs of RAM.
>>>
>>> My question:  Why does NT crash (or OS/2 for that matter) because of lack
>>> of enough RAM?  I would expect it to be slow because the OS would have
>>> to manipulate the memory, do some swapping, etc...   But, if it follows
>>> all the rules it should, theoreticaly, it should not crash.  It should
>>> just take longer to do its job, due to the overhead of running on
>>> a machine with little memory.
>>
>>Even on a virtual memory system, the system still needs a certain amount
>>of real memory in order to run itself, map I/O buffers, and keep track of
>>virtual memory, etc.  If the system is sufficiently overcommited on memory,
>>it may not be able to do all these things with available memory and in
>>fact may encounter situations where every process within the system is
>>waiting for memory currently in use by another process to be freed.
>>
> I thought that for this reason, the kernel of the OS is not made to be
> swappable, and runs at the highest possible priority, a priority that no
> other process can have.  This way, the system should always be able to
> recover.

I don't particularly like the term `swappable' when applied to an OS
kernel -- to me it implies that the entire kernel might be moved out
to disk, which makes no sense.  `Pageable' makes more sense, at least
when talking about some parts of the kernel.  (I've usually seen the
term `swap' to mean that the entire memory-resident portion of the
process or object gets written to page or swap files, and `page' to
mean that only certain sections [pages] get written to the pagefile).

But this is something of a nit.

Where I think there's a flaw in your reasoning is in assuming that
the `OS kernel' == `OS as a whole'.  Large operating systems are often
broken up internally into a number of pieces;  the file system code,
for example, really doesn't have to be permanently resident in its
entirety.  You just need enough of the device-level code to be resident
for the system to load the higher-level file system code.  GUIs are
also candidtates for being moved out to disk in low memory situations.
Even some portions of the kernel itself might be pageable -- it depends
on what contexts in which you could stand to incur a page fault and call
the disk driver, which is going to be dependent on the OS internal
architecture.

> I just remember IBM making a big deal out of the fact
> that it's AIX 3.1 (I think that's the right version #) kernel was
> actually pageable-- the implication being that most aren't.

Probably many aren't, but it's not uncommon.  That sounds like it was
probably typical marketing hype.  Note that there has to be at least a
minimal part of the kernel that isn't pageable -- the part that handles
page faults (!) and that handles low-level reads and writes to the disk
(to be able to do something with the page faults when they occur).

Now I should say that I don't _know_ that this is what is happening
to those people who have reported system lockups/crashes under low
memory conditions -- it's also possible that there's a bug somewhere.
But I think that as far as it goes it's an accurate description of
what _can_ happen in such systems.  One thing you have to keep in mind
is that when a large system like that `runs in 4MB', it doesn't mean
that there are 4MB of virtual memory used by the OS and its system-
level processes, but that there are 4MB of real memory required to
properly deal with the much larger virtual address space in use by
the OS in all its pieces.  Running it in a lot less can cause all
sorts of strange behavior, depending on what runs out first.

Bruce C. Wright