RASSP Token-based Performance Modeling Application Note

A highly abstract performance model could resolve the time consumed by cluster of processors to perform major system functions like search, FFT or FIR. A less abstract performance model could describe the time required to perform detailed tasks such as a single CPU memory access. In the context of Lockheed Martin Advanced Technology Laboratories (ATL) RASSP system design, the typical abstraction level of a token-based performance model is at the multiprocessor network level; sometimes called a network architecture performance model.

Token-based performance modeling is defined in the RASSP Taxonomy as a performance model of a system's architecture that represents data transfers abstractly as a set of simple symbols called tokens. Neither the actual application data nor the transforms on it are described other than that required to control the sequence of events in time. The application-data is not modeled, while only the control-information is modeled. For example: answer = 5 is not modeled, but control node = search is modeled.

Typically, the token-based performance model resolves the time for a multiprocessor networked system to perform major system functions. It keeps track of the usage of resources such as: memory buffer space, communication linkages, and processor units. The structure of the network is described down to the network node level. The network nodes include processor elements, network switches, shared memories, and I/O units. The internal structure of the network nodes is not described in a Token Based Performance Model.

Digital systems are growing ever more complex with improving technology and integration densities. In particular, multiple processor elements are harnessed to extend a system's processing power beyond that of single processor technology. Conventional design methods such as physical prototyping or gate level simulation, become prohibitively costly and time consuming for such systems.

Unlike the testing of an individual IC design which typically requires on the order of thousands of clock-cycles or a fraction of a second of simulated time, digital system simulation typically requires the simulation of significant portions of an application algorithm spanning several seconds or minutes of simulated time. The simulation of multiple cooperating Processing Elements (PEs), each executing many millions of instruction cycles per second (MIPS), over such time spans represents the execution of an extraordinarily large number of events. Therefore, simulation of a multi-processor system at or below a chip-behavior level becomes impractical due to the large memory and simulation run-time that would be needed.

However, full-system simulations are required to validate the overall system concept; to jointly optimize the hardware and software; and to investigate the interactions between cooperating subsystems prior to embarking on time-consuming and detailed designs that could unknowingly be misguided. Fortunately, the number of events can be reduced by adroitly using abstractions without jeopardizing the accuracy or validity of the model. Therefore, selecting an appropriate modeling abstraction level is crucial for timely design.

Simulations must execute and return results from simulation runs in roughly an hour or less to permit designers to rapidly explore, optimize, and verify system design solutions. Such a turn-around time allows several design iterations per day and ensures convergence on a valid system design in a matter of days instead of months. Therefore, more abstract prototyping methods are needed for verifying all functional aspects of a complete integrated system early in the design process.

Token-based performance modeling is an abstract form of system modeling that addresses these challenges. The simulations help identify bottlenecks, validate early system expectations or coarse timing requirements, and highlight initial design options.