Re: Config understandings

[John Klos <john%sixgirls.org@localhost>]

> > So you put the tuna WITH the mayo...
> 
> Sort of. Now, that I slept enough for two nights in a row, I can even remember
> to explain why: Normal assembler-language bignum  routines for m68k like to 
> use
> the 32x32->64bit multiplication of m680[234]0, which was removed on the M68060
> and causes a kernel trap. The compiler avoids it when -m68060 or -m68020-60
> is given, but of course it has no idea that it should use different assembler
> files (or none at all). Similar stuff happens when you use some floating point
> routines that are written with the 6888[12] FPU in mind; they are 
> trap-emulated
> on 68040 and 68060, and using a specialized libm helps a lot here in some
> cases.

Yes... avoiding those costly traps is what it's all about.

> > I rebuilt my RSAREF, OpenSSL, and OpenSSH, and startup of sshd dropped
> > from about 60 seconds of CPU time to about 15 (comparatively, the same
> > config on a 40 mhz 68040 Mac takes about 31 seconds).
>
> I didn't look at RSAREF, OpenSSL and OpenSSH myself, but apparently somebody 
> else did.
> 
> It was a factor-of-10-improvement in startup time to pgp2 and sshd, when
> I made the change.

I should've been more clear: to build all of the keys anew for the first
time took surprisingly long when I first installed OpenSSH (more than 10
minutes). However, after all of the keys are in place, generation of the
768 bit RSA key, which used to take more than 60 seconds, now only takes
about 15.

I really don't want all of my users to end up with "host fingerprint has
changed!" messages, so I won't be able to time the regeneration of all of
the keys, so I'll just say I assume it's faster.
 
> > In the last two weeks (up until I restarted ssh), the main sshd daemon had
> > accumulated over 130 minutes of CPU time.
> 
> Yesyes. the m68020-40 ssh-1.2.26 sshd needs about 2 minutes of startup time
> on a 50 MHz M68060, and another two minutes each time the session startup
> key is recalculated.

If the ssh and OpenSSH code was similar enough, this would sound about
right for the slower code.

> > Which leads me to another question: no matter how long the machine is
> > idle, the load average never drops below 1.08 (yes, I'm running the
> > distributed.net client).
> 
> Are you running xload or some similar applications? They chew up some 
> cpu time for themselves as well as in the X server, which is not noticable
> in todays standard Pentium or PowerPC machines, but noticable on <= 50 MHz
> m68k.

I have nothing that rates anywhere in top running (I don't even run X on
my server; it's colocated); I do run lots of stuff that could, and does,
raise the load average (almost 50 users, 125 megs of web traffic a day,
POP/IMAP, and so on), but the 1.08 load average is the same on that 40 mhz
68040 Mac I have here running absolutely nothing except the dnet client.

I guess it's no big deal, but it's probably a little difference in the way
that load averages are calculated. One day I'll look at the code and
figure out how they are calculated, and perhaps I'll even find out why
system time doesn't always measure properly.

Thanks,
John Klos