On Saturday, 10th July 1999, Matthew Dillon wrote:
> I'm trying to simulate your 486 setup. You must love pain! A make -j5
> buildworld on a 16MB-limited machine pages like hell (200-400 pageins/sec
> AND 200-400 pageouts/sec simultaniously, almost continuously).
Maximal pain, maximal gain! :-) The only reason I'm using a big, powerful
486 is that my 386 here died and there were none left to replace it. With
NFS src and obj, make world was taking over a week. No joking.
> Are you
> using any special sysctls or special kernel config options?
I have been using "sysctl -w vm.swap_async_max=2" for a while. It seems
to help throughput on this machine, and definitely helps interactive
performance. I suspect that a few extra I/O limiters, or some sort of
I/O rate quota system would help enforce fairness even on faster machines.
For example, we have a performance anomaly with squid on 3.2 that could
be over-eager pagedaemon behaviour flooding the I/O system.
> Also, try the latest -CURRENT and see if you can still get it stuck in
> objtrm. I haven't had any luck so far in my simulation. If you still
> get stuck in objtrm then try Alan's patch and see if that has an effect.
Maybe you should send me your latest patch, the atomic_* fixer and I'll give
it a whirl. It hasn't turned up in the cvs-cur CTM patches yet.
Stephen.
To Unsubscribe: send mail to [EMAIL PROTECTED]
with "unsubscribe freebsd-current" in the body of the message
