Parallel Execution Aware Data Structures

e observation that parallel processing inevitably runs into issues managing data movement is hardly new. e old joke “a supercomputer is an expensive device that transforms a compute-bound problem into an I/Obound problem” now has a new variant: “a manycore processor is a device that transforms a compute-bound problem into a memory bound problem.” As we enter the multicore and manycore era, data structures which make good use of memory resources become critically important. Fundamentally, parallel processing exerts more pressure on the memory subsystem than serial processing. Datasets in parallel programs tend to be large, multiple cores on each processor must contend for access to one offchip memory, and the memory subsystem is generally responsible for providing a coherent view of the system’s state. All these factors lead to memory subsystem utilization o en becoming the limiting factor in parallel computation, motivating the need for careful construction of data structures to support parallel execution. It is also important to note that instruction execution and data representation are duals of each other. Choosing a data representation o en leads to a particular algorithm or approach, and choosing a particular algorithm or parallel strategy o enhas implications on the kinds of data structures which are used. Since most pa erns for parallel processing focus on nding parallel instruction streams, we will speak in this pa ern as if the parallel execution strategy has been chosen, and the task is to nd a data structure which supports this execution strategy. In reality, the decision o en ows in the reverse direction: a particular data representation can suggest a parallel execution strategy. However, for clarity we assume that a parallel execution strategy has been xed, and our task is to nd a data structure that will allow that execution strategy to be utilized effectively. is task is a crucial one: failure to consider the implications of parallel processing on data structures usually leads to poor performance, and is a prime contributor to the failure of some parallel computations to scale productively to parallel processors, which ultimately reduces the usefulness of parallel processing. erefore, in addition to nding execution parallelism in our computations, we must also consider data structures which support this parallelism.