> On Tue, 30 Jul 2002 09:37:09 -0400, Bill Bitner <[EMAIL PROTECTED]>
> wrote:
>
> >It depends. With Linux it depends even more.
>
> The ongoing work on NUMA support will eventually benefit the S/390 -
> z/VM Linux/390 environment. If multiple virtual machines need to work
> cooperatively, it is exactly like some NUMA machine where certain
> storage is under local control, some storage is shared (shared
> seqments) but need locking, and communication between machines has
> some special requirements.
There are some issues w/r/t potential gains in multi-threading:
some workloads aren't all that capable of being subdivided
amongst many CPUs.
Some time ago (over 2 years, IIRC) I sent off a note to a fellow
describing the three key attributes of a mainframe; It was
put up on mainframes.com and I tripped over a reference to it
on another site, and I wondered who that John Campbell was.
I was surprised to find something I wrote being used as canon.
And I wrote the damned thing with minimal depth. Sheesh.
Anyway, a mainframe excels at single-thread performance, mostly
because the workload (sort/merge, database row insertion, key
balancing, etc) all have a requirement for single-thread
processing (yes, you _can_ subdivide a sortation workload across
multiple CPs but you'll be dependant upon ONE such CP to handle
the merge of the multiple subsort results). B-tree key balancing
is another items that is annoying since the B-tree itself will
usually be inaccessible to other processes looking to access the
related material.
So some workloads aren't easily subdivided. This is where the
phrase "it depends" comes in.
You can, using the Beowulf architecture, build up an awesome
aggregate of processing power as long as the workload is easily
subdivided- but, as soon as you fall afoul of an algorithm that
doesn't like subdivision, you're back to "the weakest link"- and
I'm not talking about rude remarks on television, either.
One attribute of mainframes is the paranoia implicit in their
RAS (Reliability, Availability, Serviceability) requirements.
I missed this aspect in my original write-up. Appendix "A" of
the "Linux for the S/390" redbook talks about this subject in
detail. A cluster of PC-type hardware (be it Intel-, AMD, Sparc-
or PPC-based) cannot easily compete with a mainframe when RAS
is important, even when single-thread throughput isn't a limiting
factor.
No matter which platform you choose, though, scaling, throughput,
capacity, all of these adjectives are dependant upon the workload.
So it's not a hardware/OS/hypervisor question, it's a workload
question. If you understand the workload you can then begin to
think about how it'll fit within various cabinets...
"What's on the mainframe, then?"
"Looks like a penguin"
"I don't mean what's on the cabinet, I mean what program!?"
"Let me log in... ...looks like a penguin"
--
John R. Campbell Speaker to Machines [EMAIL PROTECTED]
- As a SysAdmin, yes, I CAN read your e-mail, but I DON'T get that bored!
Disclaimer: All opinions expressed above are those of John R. Campbell
alone and are seriously unlikely to reflect the opinions of
his employer(s) or lackeys thereof. Anyone who says
differently is itching for a fight!
