Evaluating Online Hot Instruction Sequence Profilers for Dynamically Reconfigurable Functional Units

Online profiling methodologies are studied for exploiting dynamic optimization. On a dynamic optimizable system with online profilers, it has to get accurate profile in early step of the program execution for effective execution. However, for getting more effective profile by online profiling, it has to satisfy “Rapidness” and “Accuracy”. They are conflicted requirements. Therefore, it has to choose trade-off point at implementation. We focused into online Hot Instruction Sequence (HIS) profiler to exploit reconfigurable functional units. To circumstantiate the effectiveness of online HIS profiling, we build some evaluation models for experimental evaluation. Our profiler models are SC/DM, SC/FA and JC/DM. These models have different policy of event counting and table lookup. Our event counting policies are simple-counting or jumble-counting. On the other hand, table lookup policies are direct-map or full-associative. In our experimental evaluation, SC/FA and JC/DM models scored higher accuracy than SC/DM. The JC/DM model is able to implement by lower cost for table lookup, but it scored high accuracy comparable to SC/FA. key words: online profiling, dynamically reconfigurable computing, hardware profiling