Accelerating Support Count for Association Rule Mining on GPUs

In this work, we present a highly parallel work-efficient algorithm for performing support count on a GPU. We develop a compressed data layout scheme that enables high off-chip memory bandwidth utilization. Our data layout results in low overhead parallel coordination while reducing the memory requirements of support count. We evaluate our algorithm through extensive experimentation both on synthetically generated and real data. We achieve maximum throughput of 50 billion evaluations per second for our parallel two phase algorithm, while outperforming that of non work-efficient implementations on a multi-core CPU and a GPU by almost 40×. Resolving bank conflicts results in reduction of the execution time per iteration of our algorithm up to 6%. Employing additional optimizations such as loop unrolling leads to improvement in execution time up to 18%.