|
> While this may seem to be off topic for this list, I'm not so sure.
It is, but there are probably a decent amount of people interested in
this on the list, so I'll answer to the list. I'll be speaking from the
s/390 standpoint, as I'm more familiar with that; the zSeries 900 will be
similar, but on a 'next-gen' scale -- I expect 2-3 times the maximum
capacity, and it's 64 bit.
1. The s/390 is ONLY, at most, a 12-way machine in a single frame, 24-way
in a two-frame configuration. The CPU's are not super-powered;
they're normal CPU's, so imagine a normal 12-24 way, and you have a
good idea. It does have special crypto-processors built in, if you
can find a way to use them.
2. The s/390, however, has an obnoxiously fast bus -- 24GByte/s. Yes, I
did mean gigabytes. Also, I/O takes up almost no CPU time, as the
machines have sub-processors to take care of it.
3. The s/390 is a 31bit machine -- yes, 31. One bit defines whether the
code is 16 or 31 bit code. The z/900 is a 64bit machine.
Note that the s/390, afaik, suffers when attempting to access memory
over a certain amount, like any 31/32 bit machine would -- 2 gigs can
be addressed in a single clock cycle; greater than that takes longer
to process, since it requires more than 32 bits to address.
4. From top to bottom, the entire machine is redundant. There is no
single point of failure anywhere in the machine. According to IBM's
docs, the MTBF is 30 years. It calls IBM when it's broken, and they
come out and fix it. The refrigerator ad was no joke ;)
Of course, this doesn't protect you from power outages, but
interestingly enough, if I recall correctly, all RAM is either SRAM,
or battery backed -- the machine will come back up and continue right
where it left off when it lost power. No restarting instances or
apps required. No data lost.
There are five premises for the cost-savings:
1. You don't have to design a redundant system -- it's already built in.
2. One machine is easier to manage than n number servers.
3. One machine uses less facilities than n number servers.
4. A single machine, split many ways, can result in higher utilization.
5. Linux, Linux, Linux. All the free software you can shake a stick at.
On the flip-side, there are some constraints:
1. If you have 500 servers, all at 80% CPU usage, there's no way you're
going to cram them all onto the mainframe. Part of the premise is
that most servers sit at only a fraction of their maximum capacity.
2. The software must be architecture compatible.
3. Mainframe administrators and programmers are rare and expensive.
The ideal situation for an s/390 or z/Series is an application which is
not very CPU intensive, but is highly I/O intensive, that must _NEVER_ go
down. Could that be why many companies do databases on them? Think
airline ticketing systems, financial systems, inventory, etc :)
Realize, however, that your cost of entry is probably going to be well
over a million dollars, unless you want a crippled entry-level box.
You probably don't want to buy this server to run your web site. You
probably want to buy it to run your database.
That being said, if you happen to order more than you really need -- a
reasonably common phenomenon in IT shops -- you can now run Linux
instances with that extra capacity. :)
Thanks,
Kyle Sparger - Vice President, Technology
ksparger@xxxxxxxxxxxx - http://www.dialtone.com
Voice - (954) 581-0097 x 122
"Forget college, I'm going pro."
|
