Performance Improvements
In Summary of Key Findings this report gives capsule summaries of the performance notables in z/VM 6.2. The reader can refer to the key findings chapter or to the individual enhancements' chapters for more information on these major items.
z/VM 6.2's Single System Image function contains changes to the management of Minidisk Cache (MDC) for minidisks shared across systems. Prior to z/VM 6.2, the system administrator had to ensure that when a real volume was shared, MDC was set off for the whole real volume. This helped assure data integrity and up-to-date visibility of changes, but it shut off the performance advantages of running with MDC. In z/VM 6.2, the members of the SSI cooperate to turn on and off MDC according to whether the members' links taken together would permit MDC to be on or rather would require it to be off. For example, if all members' minidisk links are all read-links, MDC can be enabled on all members without danger. But if one member write-links a minidisk on the shared volume, the other members must turn off MDC for that minidisk for the duration of the one member's write-link. The members of the SSI negotiate these MDC transitions automatically. Customers using MDC will welcome this improvement.
z/VM 6.2 also contains a significant number of changes or improvements that were first shipped as service to z/VM 5.4 or z/VM 6.1. Because IBM refreshes this report only at z/VM release boundaries, we've not yet had a chance to describe the many improvements shipped in these PTFs.
Spin Lock Remediation: In VM64927 IBM significantly reworked the Control Program's spin lock manager. Instead of blindly using Diag x'44' to yield its PR/SM time slice when a spin was protracting, CP now determines which other logical PU holds the sought spin lock, uses SIGP Sense-Running-Status to determine whether said logical PU is already executing, and then issues Diag x'9C' to yield to that specific other logical PU if said logical PU is not running. These changes decrease z/VM's tendency to induce PR/SM overhead and can result in significantly decreased CPU utilization in extreme cases.
SHARE ABSOLUTE LIMITHARD: In VM64721 IBM repaired the z/VM scheduler so that hard-limiting an absolute share user's CPU consumption works correctly. Customers using absolute CPU shares with hard-limiting will want to apply this fix.
VSWITCH Failover Behavior: In VM64850 IBM repaired performance problems incurred on VSWITCHes when failover happens. During failover processing CP failed to provide the VSWITCH with sufficient QDIO buffers. This in turn limited data rate, increased CPU utilization, and opened the possibility for packet loss. The PTF repairs the problem.
Erratic System Performance: In VM64887 IBM repaired performance problems encountered on systems having a very high ratio of virtual CPUs to logical CPUs where the majority of the virtual CPUs were runnable a large fraction of the time. A condition called PLDV overflow was occasionally not being sensed and so runnable virtual CPUs were occasionally not being run when they should have been run. Customers running with high ratios of runnable virtual CPUs to logical CPUs should notice smoother operation.
VARY ON / VARY OFF Processor: In VM64767 and VM64876 IBM repaired CP VARY PROCESSOR command problems which could occasionally cause system hangs or abends.
MCW002 During QDIO Block Processing: In VM64527 IBM repaired problems with the management of memory used to hold FCP Operation Blocks, commonly known as FOB blocks. Freed (released) FOB blocks could pile up on per-logical-PU free-storage queues and never be released to the system-wide free queues. This imbalance could eventually cause a system abend.
Master-only Work on SYSTEMMP: In VM64756 IBM repaired a situation which could cause non-master logical processors no longer to service work stacked on the SYSTEMMP VMDBK. The work-stacking logic for SYSTEMMP work was changed so that non-master logical PUs would not quit checking for SYSTEMMP work once a piece of master-only work got stacked onto SYSTEMMP. In extreme cases the defect could have caused abends.
Short-Term Memory Leak: In VM64633 IBM repaired a memory leak encountered when BACKING=ANY free storage was needed but no storage above 2 GB was available. In this situation CP would unconditionally extend the backed-below-2-GB chain instead of checking whether there was free storage already available on that chain. In other words, this was basically a leak of memory below 2 GB.
Performance of CMS FORMAT: In VM64602 and VM64603 IBM shipped performance improvements for CMS FORMAT. Together these changes allow CMS to format a whole track with a single I/O.
Erasing Large SFS Files: In VM64513 IBM improved the performance of erasing large SFS files. A control block list search was eliminated, thus decreasing CPU utilization and elapsed time for the erasure.
Several Memory Management Improvements: In VM64774 IBM introduced the SET REORDER command. In VM64795 and VM65032 IBM improved the coalesce function for adjacent free memory. In VM64715 IBM improved the serialization technique used during Diag x'10' page-release processing. Our storage management article describes the aggregate effect of these improvements.
IBM continually improves z/VM in response to customer-reported and IBM-reported defects or suggestions. In z/VM 6.2, the following small improvements or repairs are notable:
- CP frame table scans now exclude entries marked permanently nonpageable.