Spearmints: Hardware Support for Performance Measurements in Distributed Systems

摘要

Event-triggered measurements are a common approach to obtain the performancecharacteristics of distributed systems. They are based on the collection, chronological ordering and analysis of events occurring locally in the system under test. As it is crucial for the observed system to suffer as little interference as possible from the measurement, in most approaches the events are analyed by a measurement system running on a separate hardware platform. So the transfer of events to the measurement system becomes the most important factor affecting the observed system and accuracy of the measurement. The authors investigated these critical aspects and developed a set of sensors that provide a facility to generate descriptions of signalized events and to transfer them to a measurement system. In the system under test only minimal overhead is necessary to signalize an event to one of the sensors. The sensor generates an event record containing the signalized data and a highly accurate time stamp which is derived from a global time base of high resolution. The event records are collected and transferred to the measurement system without any further impact on the observed system. They are used to establish a global, chronologically ordered list of events, which is analyzed to get the performance data of interest. (Copyright (c) GMD 1992.)