Re: SMP scalability: 8 -> 32 CPUs

Larry McVoy (lm@bitmover.com)
Tue, 01 Dec 1998 10:34:41 -0800


: better find out how many cpu's still work. Several years ago
: i used an 8cpu one with mach and a 24 with dynix (ewww). they both rapidly ran
: [...]
: it worked well enough for me to hack with it and watch
: it slowly ooze through n-way compiles (my box was a herd of low end 386's).

This is a great example of a place where a cluster will handily beat an SMP.
And a cluster of SMP's will do even better.

I built the software required to do parallel remote make a long time ago.
I use this to do kernel builds on my cluster of 4 machines in my office;
it's been a while since I timed it, but I think I was almost keeping up
with Linus' 4 way Xeon, which was pretty good since he has 4x450Mhz and
I have 1x166, 1x233-K6, 1x233-MMX, and 1x300-K6.

I might be wrong about the performance, but there is an easy way for you
to run your own tests: go to

http://www.bitmover.com/bitcluster

and download the tarball and install it. Be warned of two things:

- I think this overwrites your vmstat; you might not want that
- It opens up a security whole the size of Brazil so don't install
it on machines outside your firewall.

What it does give you, however, is this

$ time on bitmover.com date
Tue Dec 1 10:31:25 PST 1998
0.01u 0.00s 0:00.05r

That's a remote command, complete with current working directory and
full environment preserved, done in 50 milliseconds. The same thing
locally is 15 milliseconds. So going remote is costing 35 milliseconds.
Not free, but pretty cheap considering it's all done at user level.

There are instructions in the tarball as to how to get parallel remote
make working, I believe. Ping me if they are missing or are incomplete.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/