Well, it all depends on what you want to do, the requirements you have
and how much money you have <G>.

For a particular application, cache is always better closer to the
application then farther away.  i.e. you create an array within your
program and load the data there, and you can access it much faster then
on any disk.

But, you waste memory when the application isn't in use.  Real memory
costs real money (and if the memory is paged out, the speed of accessing
the array is greatly dimished).

In Linux, it will use any unused memory for cache.  That cache can be
used for any/all programs that is being used.  Sounds great.  And that
is great on a dedicated platform.  After all, on a dedicated platform,
there is no other use for the memory.

But, since we are talking about MDC which means VM, it usually isn't a
dedicated platform.

If you are running low on memory and can't afford the next bumb in
memory (yet), you tune back the z/Linux caching, and let MDC handle more
of the load.  MDC is a shared cache, that is shared across all dasd
where it isn't explicitly turned off.  If one Linux image becomes very
active, it will load MDC with it's data.  But when the image becomes
more idle, the cache will be used for more active images.  More bang for
the GB.

Yep, not as good as dedicated cache within a z/Linux image, but still
can ramp up good.

But then you have to consider how fast your Disk Subsystem is.  I
thought it was fast with MP3000 Internal dasd, after all, no channels.
But now we have a DS6800 Ficon-2 attached.  I can do a CMS format of a
3390-3 in under 2 minutes.  It use to take 20 minutes on the MP3000 and
45 minutes on escon attached Ramacs.

Perhaps I don't need cache anymore<G>.  Well, of course I do.  But I
can easily justify a lower cache hit rate and still have great response
from my applications.

The performance docs on the Shark said that a properly configured Shark
can substain 30,000 I/Os per second.  The number on the DS6800 state
300,000 I/Os per seconds.  I take that with a grain of salt as I have
configured the unit for our needs, not outright performance.  But even
if I only got 25% of the performance, that is still 75,000 I/Os per
seconds.

NOBODY NEEDS DATA THAT FAST!  (just kidding)

At some point, eventually, we might make the DS6800 work, not very
hard, at providing our data needs.  Cache, whether MDC, Linux cache,
Oracle SGA, all eliminate the need for real I/O at a cost of real
memory.

Perhaps you have an Oracle database, that needs great response time.
That might be a larger SGA (use up most of the available memory for
cache, here) and the hell with all other work loads.

Perhaps you needs everyone, most every one to get acceptable response.
Then MDC may be the way.  Perhaps turning off MDC for your test systems.
 (I usually don't do that as I consider programers to me be just another
department of "users" that we need to keep productive.)

IMHO, the hidden question in your question is... Do we need to rethink
the use of MDC?  And the answer normally is... Of course!  As
hardware/software and applications change, we should, always,
eventually, take the time to rethink the proper usage of our resources.

But then, who gets paid for that? <G>.

Tom Duerbusch
THD Consulting


>>> [EMAIL PROTECTED] 09/05/05 3:01 AM >>>
Johnny Tan wrote:
> By default, z/VM minidisk cache is enable. As I know, z/VM minidisk
cache is using z/VM real memory and expanded storage. Would anyone
recommend to use minidisk cache on z/VM  where the main purpose of z/VM
here is to support Linux guests?
>
> Today's DASD technology has fast write, which means that writing to
DASD will no longer write direct to the disk, however, it will write
direct to the cache (which is memory) of the DASD controller. In view of
this fast write feature, is minidisk cache still relevant? If it is
still relevant, does the z/VM minidisk cache work well to predict open
system I/O characteristics in view that I/O pattern of open system is
random (not sequential) as compared to MVS ??
Given that Linux has good disk caching built-in, mdc does only make
sense for shared volumes.
Deactivating mdc for those volumes that are local to either of your
guests saves double caching.
--

Carsten Otte
IBM Linux technology center
ARCH=s390

----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [EMAIL PROTECTED] with the message: INFO LINUX-390
or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390

----------------------------------------------------------------------
For LINUX-390 subscribe / signoff / archive access instructions,
send email to [EMAIL PROTECTED] with the message: INFO LINUX-390 or visit
http://www.marist.edu/htbin/wlvindex?LINUX-390

Reply via email to