System Sizing Guidelines for Integrity Virtual Machines Deployment -- Hardware Consolidation with Integrity Virtual Machines

7
Calculating Platform Utilization
In many cases, candidate workloads for consolidation will move from other, probably older,
servers. This immediately creates a challenge in addressing how to accommodate the difference in
hardware performance. The problem of comparing different hardware platforms for purposes of
workload performance expectation is not new and certainly not unique to Integrity VM. You must
consider issues such as server performance capability and application performance metrics when
consolidating workloads onto a new system.
Expected Platform Utilization
Generally speaking, one needs a „multiplier‟ to convert utilization on the “source” hardware
resource to an expected utilization on the “destination” hardware. For example:
old_utilization  multiplier = new_utilization
The multiplier is the fraction of the new hardware resource required to deliver one unit of the old
hardware‟s performance. Typically, this will be a number less than 1 because the target hardware
platform is an Integrity Server and the workloads are running on older hardware.
Every workload has a performance (and capacity) footprint that is unique. Therefore, you should
use a different multiplier for each workload you are planning to consolidate. The best multiplier is
the ratio of the workload‟s performance on the old hardware to that on the new hardware
configuration. If you do not have a multiplier based on the actual workload‟s performance, you
must use another approach to identify a multiplier.
For CPU utilization, the processor clock speed may be an adequate measure. This is especially
effective when the processors are in the same architecture family, but it is of limited value when
comparing processors from different architecture families (for example, PA-RISC and Itanium) due to
the various CPU architecture features that impact performance, such as local cache size, cache line
size, data latency, pipelining, number of functional units, and so on.
Standard benchmarks are sometimes good indicators of a given application‟s performance. They
also provide the benefit of being available for various server configurations – across multiple
vendors. SPECint® and TPC-C® are examples of such benchmarks.
Do not use peak performance specifications when deriving multipliers for network or mass storage
utilization. If possible, use realizable bandwidth or latency performance (whichever is appropriate
for the workload) of the I/O hardware to predict utilization.