Guidelines for Configuring Virtual Partitions on Cellular Platforms

7
It is worth pointing out that the objective of this focused study was not to determine the upper
bounds on the performance impacts for all classes of workloads in all possible virtual
partition and nPartition configurations.
We ran these two workloads and gathered metrics on different virtual partition
configurations each with the same amount of resources (one cell’s worth of: processors,
memory and associated I/O for the workload being run).
For (b) and (c) some artificially non-optimal configurations were intentionally chosen to
illustrate the influence of less than optimal resource localities on workload performance.
With those non-optimal configurations, negative performance impact on the workload was,
as expected, higher than normal.
While it may not be very obvious from the results of this focused study, it should be
emphasized that the performance penalties for non-optimal configurations could be even
higher on larger hardware configurations with member cells connected to different crossbars
The details of the experiments that follow in this paper provide the guidance to make sound
configuration choices that avoid non-optimal configurations and achieve good performance
results.
Details of Resource Layouts
The following section details the virtual partition configurations used for running the SAP SD
2-tier and SPECjbb benchmark workloads.
Layout 1: Baseline nPartition and Virtual Partition Runs
We started off the test runs by first collecting baseline data (to be used for comparisons later
on) by running the two benchmark workloads one after another on a single-cell nPartition
(configuration 1A). We then configured a single virtual partition on this single-cell nPartition
and assigned all the physical resources on the cell to this virtual partition (configuration 1B).
Each of these workloads was run in this virtual partition and results were used to compare
with the nPartition run to study the overheads introduced due to the vPars software stack.