CPU Configuration Guidelines for vPars

6
3 Establishing a Baseline Configuration
The key to achieving consistent system performance is matching the system resources (CPU, memory,
and I/O) to its dynamic workloads. Over configuring a system is one way of achieving this goal but it
comes at the cost of under-utilized resources when peak workloads are absent. One method of
utilizing CPU’s more effectively is to migrate them to where the workload on the system requires them.
The dynamic CPU migration functionality of vPars is well suited for this type of environment.
This section will detail the steps required to establish the CPU configuration of a vPar to achieve the
performance objectives for steady state and peak workload conditions. The performance objective
should be based on some form of quantitative measurement such as: response time, throughput, or
utilization. Response time and throughput are good metrics for determining application performance
because they can be measured at the application level. CPU utilization is another useful metric and
tools like the UNIX utility “top” and GlancePlus can be used to display CPU utilization. The method
that will be described here can be performed on a vPars enabled system or on a standalone system in
preparation for consolidating onto vPars.
To start with, the system to baseline should have enough CPU’s, memory and I/O to sustain the peak
workload conditions and achieve the desired performance objective. In this example, a 2 cell hard
partition with 8 CPU’s is being used to determine what the ratio of bound to floating CPU’s should be.
The system is running HP-UX 11i v2 mainly for the processor sets (psets) functionality that is integrated
into this release (For HP-UX 11i v1, the processor sets product can be downloaded and installed). The
performance objective is to bring CPU utilization to 60% for steady state and peak workload
conditions. The utilization rate chosen here is used just as an example; higher utilization rates are
achievable but factors such as future workload growth should be factored in.
For this example, a steady state workload consisting of file system and disk intensive components is
applied to the system and CPU utilization is captured with top. A real-time processor set is created to
isolate excess CPU’s from processing the workload. When a CPU is moved from the default processor
set to the real-time processor set, the CPU will be unloaded of all its applications and I/O interrupt
processing responsibilities. The applications and I/O interrupt processing responsibilities will be
distributed amongst the remaining CPU’s in the default processor set. Since no applications have been
assigned to the real time processor set, CPU’s in this processor set will remain idle. This is the desired
attribute of real-time processor sets that is useful for establishing the baseline CPU configuration.
After each run, move one or more CPU’s to the real-time processor set and re-launch the workload
and capture the CPU utilization numbers. When the CPU utilization rate meets the 60% objective, the
number of bound processors necessary to satisfy the steady state workload has been established. If
the 60% utilization objective is exceeded, move a CPU from the real time processor set back to the
default processor set.
Now increase the workload by 60% and capture the CPU utilization numbers. Move CPU’s from the
real-time processor set into the default processor set until the utilization rate of CPU’s in the default
processor set drops to 60%. The number of CPU’s in the default processor set when the 60%
utilization rate is reached is the maximum number of CPU’s that should be configured for the vPar.
The difference between the maximum CPU’s and the minimum CPU’s is the number of floating CPU’s
that should be configured.
3.1 Determining Minimum CPU Configuration
The objective of this section is to experimentally determine the minimum number of CPU’s that a
specific workload at steady state will need in order to satisfy the 60% CPU utilization objective. The
tests will be performed on an 8 CPU hard partition with the caveat that all of the memory of the hard