Managing the performance of parallel and distributed computing systems is today a serious challenge, because of their complexity and the diversity of new applications.
Digital audio and video data, with strict delay and jitter constraints, are moved and manipulated in a computing, storage and communication medium where more traditional batch, transaction and query processing applications are already in operation and expected to achieve preset levels of throughput or average response time. Enterprise wide systems will soon surface where, for example, resources are shared among applications running high volume, low response time inventory control transactions with strong CPU time demands, applications managing videoconferences with tight delay and massive bandwidth requirements, engineering visualization tools with similar requirements, background decision support queries with soft completion deadlines and large memory needs, or process control tasks with hard real time constraints and demands for immediate use of CPU, memory or bandwidth. In chapter 4 we formalize these and other user-oriented performance objectives.
Even when, under normal conditions, there is enough capacity for all applications to operate, load upsurges or system reconfiguration may generate sudden overloading that has to be managed decisively and expeditiously. In addition, the probability of system overloading increases as applications and their load increase with time. Difficult tradeoff decisions have to be made on how to allocate resources to applications and how to protect important applications from secondary ones.
Managing the performance of these systems through direct operator adjustment of system parameters (e.g. buffer pool sizes, dispatching priorities, multiprogramming level, routing policies, data and bandwidth allocation policies, etc.) is impractical. Not only does the system administrator have to tune each resource (processors, memory, bandwidth, data to storage mappings, etc.) separately and then in conjunction with the other system resources, but also has to consider several layers of tuning, and their interaction. The administrator has to tune both the base operating system on each processor and then each individual software server such as mail servers, database managers, transaction managers, etc., which typically have internal resource allocation policies (priorities, buffer pool management, etc.).
In addition, the system administrator has an interest in keeping all the computational,