Overview of z/VM HiperDispatch
In z/VM® 6.3, IBM introduced a new virtual server dispatching technology called HiperDispatch. The prime objective of z/VM HiperDispatch is to help virtual servers to get good performance from the z Systems™ memory subsystem.
z/VM HiperDispatch works toward this objective by managing the partition, and dispatching virtual CPUs in a way that takes into account the physical machine's organization, but especially its memory caches. Therefore, depending upon the type of workload, this z/VM dispatching method can help to achieve good performance on z Systems hardware.
- CPU cores are fabricated together on chips, perhaps six or eight cores to a chip, depending upon the model.
- Chips are assembled onto nodes, perhaps three to six chips per node, again, depending upon model.
- The nodes are then fitted into the machine's frame.
- Each core has its own L1 and L2.
- Each chip has its own L3.
- Each node has its own L4.
- Beyond L4 lies memory.
One way z/VM HiperDispatch tries to achieve its objective is by requesting that the PR/SM hypervisor provisions the LPAR in "vertical mode". A vertical mode partition has the property that the PR/SM hypervisor will repeatedly attempt to run the partition's logical CPUs on the same physical cores, and to run other partitions' logical CPUs elsewhere. In this way, the partition's workload benefits from having its memory references build up context in the caches. Therefore, the overall system behavior is more efficient.
Another way z/VM HiperDispatch tries to achieve its objectives is to repeatedly run the guests' virtual CPUs on the same logical CPUs. This strategy helps guests to experience the benefit of having their memory references build up context in the caches. This should also enable the individual workloads to run more efficiently.
http://www.vm.ibm.com/perf/tips/zvmhd.html