Simulation Goals



next up previous
Next: Simulation Environment Up: Performance Results Previous: Performance Results

Simulation Goals

This section describes simulations results of the NUMAchine prototype. There are three main reasons for developing a simulator for NUMAchine: to estimate the performance of the prototype machine, to locate any weaknesses in the NUMAchine architecture and/or the particular parameters of the prototype, and to investigate a number of tradeoffs. In this document, the only simulations shown are those that provide indicators of NUMAchine's performance; however, the simulator is being used on a continuing basis to investigate architectural parameters, to improve the NUMAchine architecture. A complete behavioral simulation of a "virtual" machine has been implemented in software, using state-of-the-art simulation techniques. The simulation environment described in the next section is one step removed from a circuit-level simulation; all behaviour affecting timing and functionality was modelled in detail for each system component that could act independently.

For design verification our primary concerns were the efficiencies of: the rings, the Network Cache, and the cache coherence protocol. For the rings, the obvious question is whether network contention causes serious performance degradation. For the Network Cache, we are interested in its effectiveness at reducing network traffic. Finally, as mentioned in Section 2.3, the coherence protocol was designed optimistically, assuming that certain cases would not happen frequently; thus we wish to determine the actual frequency of these cases, in order to assess whether the protocol will perform as efficiently as hoped.

Beyond the above goals, there are many other questions on enhanced functionality that can easily be asked in a simulation. For example, the benefits of prefetching, broadcasting and weaker consistency models are all of interest. The answers to these questions (and others) are currently under investigation, but for brevity will not be reported here.



Stephen D. Brown
Wed Jun 28 18:34:27 EDT 1995