[ On Mon, December 21, 1998 at 11:50:50 (+1030), Ian Dall wrote: ]
> Subject: Re: userid partitioned swap spaces. 
>
> I suggest non-root processes merely become blocked if they need swap
> above the high water mark. Processes which don't need swap continue to
> run and if they exit and free up space, then the blocked processes
> become runnable again. Of course, it is possible, even likely that the
> system could wedge with all processes are either "permanant" daemons
> or blocked waiting for swap. That requires either human intervention
> or an "intelligent" process to handle the situation.

Indeed such a solution is just begging for un-avoidable and effectively
un-detectable deadlock.

I also think this "intelligent" (nothing I've proposed to date requires
much, if any, "intelligence") process should either be an integral part
of the VM subsystem (i.e. in-kernel), or at minimum be a kernel thread.

I really don't like the idea of having something outside the kernel try
to fix a problem such as this -- it introduces far too many variables,
something that's definitely not a good idea when your goal is
robustness.

I don't like kernel bloat any more than anyone else, but I don't think
any of the algorithms necessary, beyond the actual VM accunting that
seems to be necessary anyway, are very complex or difficult to
implement, and as such I think it's fair to request that they be
implemented inside the kernel where they can do the most good and where
they'll cause the least trouble under stress.

> Systems with no users at all such as file servers, gateways etc I
> would have thought would be quite predicatable regarding how much swap
> they use and could be fairly easilly configured appropriately.

Well yes that's true in *some* cases, but even when it is true a robust
system must deal with abnormal conditions, as well as normal ones.

-- 
							Greg A. Woods

+1 416 218-0098      VE3TCP      <gwoods@acm.org>      <robohack!woods>
Planix, Inc. <woods@planix.com>; Secrets of the Weird <woods@weird.com>