Dynamic Analysis and Profiling of Multithreaded Systems - Advanced Operating Systems and Kernel Applications - page 162

and impact of different profiling techniques.
Our experimentation is based on measurements
taken from Microsoft Windows XP, running on
a dual-processor, hyper-threaded (Intel Xeon 2.8
GHz) system, executing a stress-test Web client/
server application. The measurements were taken
using both Windows performance counters and
the on-chip Intel performance counters. Table 3
shows the results.
The data listed in Table 3 is comprised primar-
ily of finer-grained metrics that occur at very high
frequencies in the lower levels of the system. Of
course, less frequent “application-level” events are
also of interest in understanding the behavior of
a system. For example, rare error conditions are
often of importance. The data in Table 3 shows
that the frequency (and therefore quantity) of
measurable events can vary signifi cantly by up
to nine orders of magnitude. Because the impact
of measurement is scaled proportionally, analysis
methodologies that work well for low-frequency
events may not do so for higher-frequency
events.

challenges of multithreaded System
Profiling

The remainder of this chapter focuses on the re-
alization and application of runtime profiling on
multithreaded sys tems. Profiling multithreaded
systems involves addressing the following key
challenges:

• Measurement of events at high frequen-

cies—Events of interest typically occur at
high frequency. The overhead and effect of
measurement on the system being measured
must be controlled carefully. Without careful
control of overhead, results become skewed
as the process of measurement directly alters
the system's behavior.

• Mapping across multilevel concepts—

Threads can be used at multiple levels of a
system. For example, threads can exist in the
operating system, virtual machine, middle-
ware, and in the application (lightweight
threads and fibers). Virtual machine and
application-layer threads can map to under-
lying operating system threads. Extracting

Table 3. Example metric ranges

Category Metric Range

Processor Clock Rate 2,793,000,000 Hz *

Micro-ops Queued 630,000,000 uops/second *

Instructions Per Second 344,000,000 instructions/second *

L2 Cache Reads 65,000,000 reads/second *

Thread Schedul-
ing

Number of Threads 500 total count

Context Switch Rate 800-170,000 switches/sec

Thread Queue Length 0-15 total count

Scheduling Quanta 20-120 ms

System Resources System Calls 400-240,000 calls/sec

Hardware Interrupts 300-1000 interrupts/sec

Synchronization Objects 400-2200 total count

* per logical processor