EXTRACTION AND PREDICTION OF SYSTEM PROPERTIES USING VARIABLE-N-GRAM MODELING AND COMPRESSIVE HASHING

In modern computer systems, memory accesses and power management are the two major performance limiting factors. Accesses to main memory are very slow when compared to operations within a processor chip. Hardware write buffers, caches, out-of-order execution, and prefetch logic, are commonly used to reduce the time spent waiting for main memory accesses. Compiler loop interchange and data layout transformations also can help. Unfortunately, large data structures often have access patterns for which none of the standard approaches are useful. Using smaller data structures can significantly improve performance by allowing the data to reside in higher levels of the memory hierarchy. This dissertation proposes using lossy data compression technology called ’Compressive Hashing’ to create “surrogates”, that can augment original large data structures to yield faster typical data access. One way to optimize system performance for power consumption is to provide a predictive control of system-level energy use. This dissertation creates a novel instruction-level cost model called the variable-n-gram model, which is closely related to N-Gram analysis commonly used in computational linguistics. This model does not require direct knowledge of complex architectural details, and is capable of determining performance relationships between instructions from an execution trace. Experimental measurements are used to derive a context-sensitive model for performance of each type of instruction in the context of an N-instruction sequence. Dynamic runtime power prediction mechanisms often suffer from high overhead costs. To reduce the overhead, this dissertation encodes the static instruction-level predictions into a data structure and uses compressive hashing to provide on-demand runtime access to those predictions. Genetic programming is used to evolve compressive hash functions and performance analysis of applications shows that, runtime access overhead can be reduced by a factor of ~3x-9x