Summary of Key Findings
This section summarizes key z/VM 6.2 performance items and contains links that take the reader to more detailed information about each one.
Further, our performance improvements article gives information about other performance enhancements in z/VM 6.2.
To compare performance of z/VM 6.2 to previous releases, IBM ran a variety of workloads on the two systems. For the base case, IBM used z/VM 6.1 plus all Control Program (CP) PTFs available as of September 8, 2011. For the comparison case, IBM used z/VM 6.2 at the "code freeze" level of October 3, 2011.
Regression measurements comparing these two z/VM levels showed nearly identical results for most workloads. Variation was generally less than 5%.
Because of several improvements brought either by z/VM 6.2 or by recent PTFs rolled into z/VM 6.2, some customers might see performance improvements. Customers whose partitions have too many logical PUs for the work might see benefit, likely because of improvements in the z/VM spin lock manager. Storage-constrained systems with high pressure on storage below 2 GB might see benefit from z/VM's improved strategies for using below-2-GB storage only when it's really needed. Workloads with high ratios of busy virtual PUs to logical PUs might see smoother, less erratic operation because of repairs made in the z/VM scheduler. For more discussion of these, see our improvements article.
Key Performance Improvements
z/VM 6.2 contains the following enhancements that offer performance improvements compared to previous z/VM releases:
Memory Scaling Improvements: z/VM 6.2 contains several improvements to its memory management algorithms. First, z/VM now avoids using below-2-GB memory for pageable purposes when doing so would expose the system to long linear searches. Further, z/VM now does a better job of coalescing adjacent free memory; this makes it easier to find contiguous free frames when they're needed, such as for segment tables. Also, z/VM now uses better serialization techniques when releasing pages of an address space; this helps improve parallelism and reduce memory management delays imposed on guests. Last, z/VM now offers the system administrator a means to turn off the guest memory frame reorder process, thereby letting the administrator decrease system overhead or reduce guests stalls if these phenomena have become problematic.
ISFC Improvements: Preparing ISFC to be the transport mechanism for live guest relocation meant greatly increasing its data carrying capacity. While the improvements were meant for supporting relocations, the changes also help APPC/VM traffic.
Since the time of z/VM 5.4, IBM has also shipped several good z/VM performance improvements as PTFs. For more information on those, refer to our improvements discussion.
Other Functional Enhancements
z/VM 6.2 offers the ability to relocate a running guest from one z/VM system to another. This function, called live guest relocation or LGR, makes it possible to do load balancing among a set of z/VM partitions bound together into a Single System Image or SSI. LGR also makes it possible to take down a z/VM system without workload disruption by first evacuating critical workload to a nearby z/VM partition.
IBM measured the SSI and LGR functions in two different ways. First, IBM ran a number of experiments to explore the notion of splitting a workload among the members of an SSI. In our resource distribution article we discuss the findings of those experiments. Second, IBM ran many measurements to evaluate the performance characteristics of relocating guests among systems. These measurements paid special attention to factors such as the level of memory constraint on source and target systems and the capacity of the ISFC link connecting the two systems. In our guest relocation article we discuss the performance characteristics of LGR.
Though it first appeared as a PTF for z/VM 5.4 and z/VM 6.1, the new CPU Measurement Facility host counters support deserves mention as a key z/VM improvement. z/VM performance analysts are generally familiar with the idea that counters and accumulators can be used to record the performance experience of a running computer. The CPU Measurement Facility host counters bring that counting and accruing scheme to bear on the notion of watching the internal performance experience of the CPU hardware itself. Rather than counting software-initiated activities such as I/Os, memory allocations, and page faults, the CPU MF counters, a System z hardware facility, accrue knowledge about phenomena internal to the CPU, such as instructions run, clock cycles used to run them, and memory cache misses incurred in the fetching of opcodes and operands. The new z/VM support periodically harvests the CPU MF counters from the CPU hardware and logs out the counter values in a new Monitor sample record. Though Performance Toolkit offers no reduction of the counter values, customers can send their MONWRITE files to IBM for analysis of the counter records. IBM can in turn use the aggregated customer data to understand machine performance at a very low level and to guide customers as they consider machine changes or upgrades. For more information about this new support, refer to our performance considerations article.