Shared Memory Multiprocessor System Performance Simulating Shared

Shared Memory Multiprocessor System Performance

Simulating Shared Memory Multiprocessor System Performance

In simulating the performance of shared-memory microprocessors, the study Shared-Memory Multiprocessor Systems -- Hierarchical Task Queue (Serrazi, 2007) seeks to isolate the effects of parallel processing as it relates to memory and process resource allocations. The researcher who also wrote this analysis relied on a series of very precise simulations of multiprocessor performance that sought to define variables to quantify the allocation of memory and processor performance in centralized, distributed, and highly hierarchically-driven memory and application load performance across a standardized memory usage architecture (Serrazi, 2007). The researcher concludes that of the three approaches to testing shared-memory multiprocessor systems, hierarchically-based methodologies are the most effective in optimizing shared memory performance as they compensate for shared memory performance. The author and researcher concludes from this analysis that the hierarchical model is the best for equally balancing workloads across tasks queues and seeking an optimal performance level. This imbalance of tasks queues could easily be averted through the more use of more efficiently algorithms, yet the researcher reverts to a more hierarchically-based approach to allocating processing across the multiple tasks running in the simulation. A more robust approach to defining the overall optimization of tasks through the multi-processor simulation is needed for the simulations to be fully illustrative of overall performance. In addition the methodology takes a relatively simplistic approach to a very complex problem of measuring the relative performance of shared memory, processor and system-based task overhead. The rudimentary approaches to defining load balancing in the methodology need to be re-thought to take into account more of an iterative approach to testing load balancing and seeking to optimize this factors across system constraints. There is no constraint-based modeling in this simulation to the level that would be necessary to use these results for actual Research & Development (R&D) efforts and strategies (Unger, Bidulock, 1982). The simulation concludes that hierarchically-based approaches are best for managing shared memory and seeking to optimize shared memory queues. The simulation achieves its stated purpose yet has much room for improvement.

Improvements For This Study

The study has been designed with a methodology that virtually assures hierarchical aspects of testing will deliver the greatest level of performance for shared memory process configurations. The load balancing and task queues are designed to be optimally balanced across parallel or asynchronous queues, which further biases the results of the study. One of the major factors in the study being so optimized with a hierarchical design is that fact that optimizing across that structure is much more predicable from a memory allocation and processor performance standpoint.

Second, the methodology does not include enough randomization of data packet sizes, speeds, composition or the inclusion or exclusion of specific randomized packets. All of these factors will force a dual- and multi-processor configuration to exponentially slow down and even freeze if pushed to the limits. Yet the researcher does not do this; he doesn't get to the level of load on the system to force it to fail. This would be extremely valuable to know across each of the testing scenarios. Yet that is lacking from the study.

Third, the variation in cache levels, not randomized enough to stress the overall multi-processor performance, could use a much more rigorous level of analysis as well. The approach to randomization of cache levels is also nearly assuring the hierarchical aspects of the methodology will also be chosen. Again the researcher does not push the multi-processor configuration to the brink of failure and less than optimal results are generated as a result.