z/VM HiperDispatch Workloads
To evaluate z/VM HiperDispatch IBM constructed a tiled workload built upon Virtual Storage Exerciser (VIRSTOEX).
The VIRSTOEX virtual machines vary in:
- Number of virtual CPUs,
- The percent-busy at which they run their virtual CPUs,
- T/V ratio,
- Virtual storage size,
- Number of storage pages they instantiate, and
- Number of storage locations they touch on each instantiated page.
The VIRSTOEX virtual machines are organized into groups called tiles. A tile consists of an assortment of VIRSTOEX virtual machines of specific configuration. To ramp up a workload, the number of tiles is increased.
A LIGHT tile consists of the following assortment of VIRSTOEX virtual machines:
- One virtual 1-way with size 4 GB, trying to run each virtual CPU 15% busy, instantiating 17000 pages (66 MB), touching one doubleword on each page.
- One virtual 2-way with size 16 GB, trying to run each virtual CPU 33% busy, instantiating 33000 pages (132 MB), touching one doubleword on each page.
A HEAVY tile consists of the following assortment of VIRSTOEX virtual machines:
- One virtual 1-way with size 4 GB, trying to run each virtual CPU 15% busy, instantiating 17000 pages (66 MB), touching one doubleword on each page.
- One virtual 4-way with size 16 GB, trying to run each virtual CPU 31% busy, instantiating 33000 pages (132 MB), touching one doubleword on each page.
- One virtual 8-way with size 64 GB, trying to run each virtual CPU 50% busy, instantiating 131000 pages (512 MB), touching one doubleword on each page.
In a given run, all VIRSTOEX machines are set either to run with T/V as low as possible (1.00) or to use frequently issued Diag x'0C' calls so as to elevate its T/V.
A given run consists either of a number of LIGHT tiles or a number of HEAVY tiles, all virtual machines running at either low T/V or high T/V.
A given run is done in a dedicated N-way partition of a given number of logical CPUs, with enough real storage that the workload is guaranteed never to page.
Spectra of workloads are created by varying shape of a tile, number of tiles, machine type, and N-way level of the partition.
Transaction rate for this workload is taken to be the rate at which its virtual servers collectively stride through memory touching pages. Actual stride rates are typically divided by some large denominator to keep ETRs tenable.
This workload does no virtual I/O and does not page. Thus its only constraint is the performance of CPU or memory, and studying ETR is therefore sufficient for studying its performance.