Today's mainframeanything to new 553

All sessions to the machines in question are via ssh and a pretty paranoid vpn based on OpenBSD boxes. This may explain the "mux feel"; a very slight delay in echo of the first characer in a session and then perfect response again. Tops20 is notorious for that "mux feel", even if it is very slight.

However, the OpenBSD boxes give a "RT-os" feel (like I know from QNX and variuos other realtime stuff); the characters are there on the screen before you can perceive pressing the key. Even over the very same ssh-vpn stuff.

What is new in Linux 2.6 is that I can do an accounting run; this is a process that currently uses 1 h 28 minutes cpu in the mysql server, and around 6 minutes in the application code. Yes, we have a policy to keep all significant stuff in SQL. Since the box has multiple CPUs and multiple stripe-mirror sets I farm out several processes to handle unrelated queries; like running a process for every day of accounting records.

In Linux 2.4 a process per cpu (2 on a 2-way, 4 on a 4-way) begins to be felt; the system feels like a swapping Primos. Not too bad, but definate impact; and a random keypress delay. The load goes to somewhat above 6 on a 4-way machine with 4 processes.

With Linux 2.6 I notice nothing before I make the disks thrash; this seems to happen around 12 processes on a 4-way system. I settled on 8 processes dispatched on the production system; this executes the whole thing in 16 minutes walltime; and emacs over X (over ssh+paranoid VPN) gives good response. Not all code in this run can be multi-threaded. There is at least a 100:1 query:update ratio, so database caches etc. can run without much locking problems.

From my gut feeling; FreeBSD would give a definate loaded-feel, and OpenBSD wouldn't handle that many cpus. Solaris might do it; but not with such a good "feel"; I don't know about AIX. IRIX would probably do very nicely. I would love to see a similar example from the windows side of things.

-- mrr