Benefits of Configuring More Memory in the IBM z/OS Software Stack

Readers' comments

Readers' comments (8) 

lockedThis discussion is now locked


Posted by Hari Shanmugadhasan on 6 April 2017 at 14:50

PGFIX(NO) for every MDM BP is an odd choice for a Redbook about exploiting more memory. See Fig 2-10 on page 17 for the baseline with no mentioned of any PGFIX change for the subsequent test.

Posted by Hari Shanmugadhasan on 6 April 2017 at 14:51

Using FRAMSIZE(4K) frame size for every MDM BP is an odd choice for a Redbook about exploiting more memory, but is forced by the odd PGFIX choice.

Posted by Hari Shanmugadhasan on 6 April 2017 at 14:53

The GBP CASTOUT strategy is odd, "particularly because most of the MDM workload is read only" (see p16). Plus, the GBPCACHE ALL usage with GBP sized to hold all the MDM pages means the pages are "retained" without using "GBP castout thresholds were intentionally set high so that pages are retained in the GBP" on page 17.

Posted by Hari Shanmugadhasan on 6 April 2017 at 14:54

If GBPCHKPT was increased beyond the already high baseline of 10 minutes, which is already substantially higher than the 4 minute default, it would seem to be designed to avoid DASD write activity during the test measurement at the expense of realism and availability.

Posted by Hari Shanmugadhasan on 6 April 2017 at 14:56

The page population strategy on page 17 seems almost guaranteed to leave the local BPs almost full of initially useless pages while forcing the high cost of page registration and deregistration for every page, a cost which appears to be left out of the test results.

Posted by Hari Shanmugadhasan on 6 April 2017 at 14:59

It would be helpful to have a figure corresponding to the baseline Fig 2-10 that highlights the changes in the configuration for the subsequent test.

Posted by Hari Shanmugadhasan on 6 April 2017 at 15:02

It would be helpful if the read/write ratio of the MDM workload was described, so one could see why it made sense to have GBPs holding everything, instead of larger local BPs to avoid synch I/O with relatively small GBPs to hold only changed pages, especially when "most of the MDM workload is read only."

Posted by Hari Shanmugadhasan on 6 April 2017 at 15:06

Why was PGSTEAL(NONE) used for all MDM BPs for both members? It appears to about double initial activity that is not captured by the subsequent test.


Profile

Publish Date
21 January 2017

Last Update
18 February 2017


Rating: Not yet rated


Author(s)

ISBN-10
0738455962

IBM Form Number
REDP-5238-01