Computer Science Technical Report Experimental Evaluation of Blocking and Non-Blocking Multithreaded Code Execution

The objective of multithreaded execution models is masking the latency of inter processor communications and remote memory accesses in large-scale multiprocessors. Several such models combine aspects of data ow-like execution with the von Neumann model in an attempt to provide both e cient synchronization (as in the data ow model) and e cient exploitation of program locality (as in the von Neumann model). We refer to these models as data-driven multithreading models. One of the factors that distinguishes these models is the thread execution strategy: A thread can be either non-blocking or blocking. Another factor is the architectural support for dynamic synchronization: The locality present within and among threads can potentially be exploited by a proper storage hierarchy for synchronization store (operand storage). Two storage models have been proposed for data-driven multithreaded execution. One is frame based, in which all the threads belonging to a code-block share one storage segment called frame; the other is framelet based, in which each thread has its own storage segment, called framelet. This article experimentally compares two thread execution models and their related storage models. The rst is a blocking execution model that relies on a scheduler for the allocation of threads to processors and exploits inter thread locality within a code-block. It relies on the frame storage model and assumes a certain amount of compile time data distribution to minimize network accesses. The second is a non blocking execution model in which threads are dynamically scheduled based on data availability. It relies on the framelet storage model and makes no assumptions about the static allocation of data to processors. The experimental evaluation takes into account the impact of the storage hierarchy design on the performance of the two models.