174
Development Tools User’s Guide
Intel
®
IXP2400/IXP2800 Network Processors
Performance Monitoring Unit
The Performance Monitoring Unit (PMU) supports time and random based sampling for the
IXP2400 and IXP2800 network processors. The sampling modes describe the method by which
the supported performance counters are read (or sampled).
Time based sampling
(TBS) is the most understood and therefore, most common sampling
method in use today. TBS is also known as
periodic sampling
since sampling is triggered by a
periodic timer. TBS is supported in both point sampling mode and windowed mode. Point
sampling simply reads the performance counters on a periodic basis, while window mode enables
performance counters for a small window of time on a periodic basis.
Random based sampling
(RBS) represents an important enhancement over TBS window mode.
By randomizing the period between sampling windows, a statistical model of performance is
generated. This method generally exposes system phenomena that the period method can miss or
alias.
Both RBS and TBS are only supported using "high-speed" sampling (i.e. where the sampling is
driven by an interrupt on the network processor's Intel XScale
core). The so called "normal
sampling", where the sampling is driven by the workbench is not supported as it offer decreased
performance over "high-speed" sampling.
3.3.1
Time Based Sampling
Time based sampling (TBS) is the most common sampling method used for data collection. Two
forms of TBS exist and are both useful dependent upon workload and data type being collected. In
either case, the sampling (or reading of performance counters) is done on a fixed, periodic nature.
3.3.1.1
Point Sampling
Time based, point sampling is the simplest method of collecting performance counter data. On a
periodic basis, the performance counters are read, reset and re-enabled. The sampling period is
generally set to a large value (>100
µ
s) in order to not cause excess overhead on the system due to
sampling. Because of this, the data is generally average over the sampling period. As the period is
stretched out over time, peaks and valleys can be lost due to the averaging effect, and therefore, this
method is not recommended when instantaneous bandwidth analysis is desired. This method is
more suited to following transaction oriented data analysis.
Figure 55. TBS Point Sampling
B2739-01
F
F
F
F
F
F
Time
is Sample Point
is Fixed Delay
Sample, Reset,
Start Counters