Next: Summary statistics
Up: Related Work
Previous: Related Work
Observation tools and techniques provide timing measurements of
parallel system behavior. Depending on their purpose, these
measurements may be grouped by their location (i.e., source code or
processor), operation (e.g., send, receive, barrier), or data structure
accessed. These measurements can be performed at various levels of
abstraction as follows:
- Summary statistics
- provide high-level performance
characteristics, such as running time or the number of messages
exchanged.
- Execution events
- capture detailed performance data about
individual operations, such as the time to send a single message or
complete a synchronization operation.
- Abstract events
- express performance behavior at a level
higher than execution events. Abstract events are determined by
grouping many execution events and analyzing their meaning.
Examples of abstract events are load imbalance and
insufficient parallelism.
In the sections that follow we discuss each of these approaches in more
detail.
Wagner Meira
Wed Feb 28 15:37:47 EST 1996