I think that the answer is probably a little of "it depends" (thank you
Bill Bitner). The "it depends" I think is related to your I/O rate and I/O
requirements. An example, and even tho it is very old, I think would still
apply, is that back in the late '80's I had some 3090's as well as some
3880-J21's. The 3090's had expanded storage and the J21's were a special
control unit with cache memory and 3350's for backing storage that were
intended for use only as paging devices. I ran some tests that involved
bringing up enough virtual machines running some code that would cause a
lot of paging. I don't remember the number of machines now but it was
enough to be able to saturate the paging system regardless of what the
paging configuration.
I ran the first test just paging to only spinning 3350's. The 3350's only
with no paging to the J21 cache or expanded storage would support "X" pages
per second, I think around 100. The next test, still without expanded
storage but paging to the J21 cache would support 10 "X" pages per
second. The last test, using expanded storage, J21 cache and spinning
3350's would support paging up to 100 "X" pages per second.
I don't remember the average I/O time for a 3350 even with chained page I/O
and I don't remember the channel configuration. I think that the channel
speed on a 3090 was 4.5 MB/sec which is considerable slower than ESCON or
FICON. I am also sure that memory transfer times was a lot slower on a
3090 than what you see today, but everything is likely to be relative.
So, I think that the answer is that you will be better off from a pure
speed standpoint if you use MDC rather than control unit cache, but you
could try with and without MDC and see if your I/O rate goes up with
MDC. If it does and if it is important to your application requirements
that you get as much I/O thru the system as you possibly can, you'd be
better off getting as much off your channel sub-system as you can. If your
I/O rate doesn't go up when using MDC, then you're really not gaining
anything in using it. No user, sitting at a terminal, is going to notice
if his/her I/O was satisfied at channel speed or memory speed.
Jim
At 03:40 AM 7/12/2006, you wrote:
On the, "Get close to the application" thread, remember that any I/O=20
request satisfied from MDC is likely to be satisfied (on modern boxes) in=
=20
less than a microsecond. Even cached, anything that involves a=20
conversation with a channel is going to experience a service time that is=
=20
several orders of magnitude higher.
It's always important to reflect on the cost and consider the best way of=
=20
investing ones limited currency (in this case, central / expanded=20
storage), but some MDC payback is usually a no-brainer.
I believe this to be especially true when dealing with synchronous I/O -=20=
and how much I/O nowadays is truly, "application asynchronous"? Not as=20=
much as in the, "batch sequential" days, that's for sure. (Not, of course=
,=20
that CMS has ever been strong in the, "I/O overlap" department from a=20
single user's perspective.)
It would be an unusual configuration that would benefit from the=20
reassignment of all MDC storage to, "other uses".
Jim Bohnsack
Cornell Univ.
(607) 255-1760
