I think others have covered most of the factors on which it
depends. One area that seemed to beg for a little more discussion
was MDC in central storage vs. expanded storage. Two things to bear in
min are:
1. CP uses access register mode to address guest memory for MDC in
   many cases
2. the architecture limits xstore operations to page boundaries
So if guest I/O buffers are page aligned, movement can be directly
from xstore to the guest pages. Otherwise, the data is moved into
cstore and then into the guest pages. A lot of CMS applications
and workloads have buffers page aligned; less so with guest
environments. Because of this processing, it is important to have
some cstore, otherwise CP spins its wheels using cstore for
non-aligned buffers and almost immediately freeing it up.

The answer also depends on what release. Prior to 5.2.0,
environments that were constrained on memory below 2GB would
often be suggested that one disable the xtore MDC to create
a greater chance of access register mode being used.

I tend not to just look at the hit ratio, but at the
I/Os avoided. 50% hit ratio on 1000 I/Os per second is
better than 90% hit ratio on 100 I/Os per second.

Bill Bitner - VM Performance Evaluation - IBM Endicott - 607-429-3286

Reply via email to