In Summary of Key Findings this report gives capsule summaries of the performance notables in z/VM 6.4.
z/VM 6.4 contains several small functional enhancements related to performance.
VM65733: Support for the new z Systems Vector Facility (SIMD).
VM65716: Support for the z13 Driver D27, including LPAR Group Absolute Capacity Capping. Also included were exploitation of Single Increment Assign Control to make dynamic memory upgrade more efficient and decreasing CP's real crypto adapter polling interval to benefit APVIRT users.
VM65680: Support for prorated core time. This in turn lets ILMT work correctly.
VM65583: Support for multi-VSWITCH link aggregation.
Service Since z/VM for z13
z/VM 6.4 also contains a number of performance repairs that were first shipped as service to earlier releases. Because IBM refreshes this report only occasionally, IBM has not yet had a chance to describe these PTFs.
VM64587: VDISK pages were not being stolen aggressively enough.
VM64770: The read-in of guest PGMBKs at guest logoff was inefficient.
VM64890: A bad loop counter caused excessive CPU consumption in Minidisk Cache (MDC).
VM64941: In some situations the guest change-bit could end up being wrongly set for an IBR page. This could in turn cause excess processing in the guest.
VM65097: PGMBK prefetch can slow the system when Diag x'10' requests are single-page.
VM65101: Invalid-but-resident (IBR) pages on the global aging list were being rewritten to DASD even though they had never been changed since having been read from DASD.
VM65189: An error in memory management was causing excessive and useless PGMBK reclaim work to be stacked on SYSTEMMP.
VM65199: Errors in the dispatcher were causing failures in how SYSTEMMP work was being handled. One failure was that once the master CPU started processing SYSTEMMP work, it would not process anything else until the SYSTEMMP queue was empty. Another failure was that some CPUs eligible to handle SYSTEMMP work were not being awakened when SYSTEMMP work needed to be done.
VM65420: Frames that should have been stolen from MDC were not being stolen.
VM65692: A defect in HCPARD was causing a certain interception control bit to remain on when it should not. This was causing excessive simulation overhead in CP.
VM65709: MDC processing was being done even though the MDIMDCP flag, which is supposed to inhibit MDC, was set. This useless processing caused waste of CPU time.
VM65748: Certain zHPF (High Performance FICON) features were unavailable to guests even though the hardware supported them.
VM65762: Under some circumstances CP might fail to deliver PCI thin interrupts to guests.
VM65794: MDC might fail to work for RDEVs whose device number is greater than or equal to x'8000'.
VM65801: Even though its OSA uplink port had reached its capacity, VSWITCH continued to redrive the uplink port, causing excessive CPU consumption.
VM65820: Because of a PE in VM65189, PGMBK reclaim could exit without releasing an important lock. This stopped all future PGMBK reclaim, which in turn made PGMBKs into a memory leak, which in turn consumed memory unnecessarily.
VM65824: The administrator had set Minidisk Cache (MDC) off for a given real volume, but if a user then logged on and said user had a DEVNO minidisk defined on that real volume, the system would turn on MDC for that volume.
VM65837: If a DASD recovery I/O gets queued at an inopportune moment, I/O to the DASD device can stall.
VM65845: If a hyperswap occurs at an inopportune moment, I/O to the swapped device can stall.
VM65869: Remove excessive LOGOFF delay for QDIO-exploitive guests.
VM65886: CCW fast-trans wrongly marked some minidisk I/Os as ineligible to be performed on HyperPAV aliases.
IBM continually improves z/VM in response to customer-reported or IBM-reported defects or suggestions. In z/VM 6.4 the following small improvements or repairs are notable:
Excessive SCSI retries: When IPLing the LPAR from a SCSI LUN, if one or more of the paths to the LUN were offline, the IPL would take a long time. It was found there were excessive retries of Exchange Config Data. The retry limit was decreased.
Lock hierarchy violation: A violation of a lock acquisition hierarchy slowed down processing in VSWITCH. The violation was repaired.
Memory leak in QUERY PROCESSORS: The handler for the CP QUERY PROCESSORS command contained a memory leak. If the leak were hit enough times the system could slow down. The leak was repaired.
Unnecessary VSWITCH real device redrives: Under some circumstances HCPIQRRD unnecessarily redrove an uplink or bridge port real device. The unnecessary redrives were removed.
Unnecessary emergency replenishment scans: Memory management would do unnecessary emergency replenishment scans in some situations. The scans could hang the system. The reason for the unnecessary scans was found and repaired.
Incorrect SMT dispatcher adjustments: In z/VM 6.3 the dispatcher behaved slightly differently for SMT than for non-SMT. This difference was in play for SMT-1 when it should not have been. The result of the error was that SMT-1 performance was worse than non-SMT performance. z/VM 6.4 repairs the system so that the difference is in play for only SMT-2. This change makes SMT-1 perform the same as non-SMT.
Unnecessary calls to MDC steal: Memory management was found to be making calls to steal MDC frames even when it was known MDC was using no frames. The unnecessary calls were removed.