Micro-architectural Analysis of OLAP: Limitations and Opportunities

Understanding micro-architectural behavior is profound in efficiently using hardware resources. Recent work has shown that, despite being aggressively optimized for modern hardware, in-memory online transaction processing (OLTP) systems severely underutilize their core micro-architecture resources [25]. Online analytical processing (OLAP) workloads, on the other hand, exhibit a completely different computing pattern. OLAP workloads are read-only, bandwidth-intensive and include various data access patterns including both sequential and random data accesses. In addition, with the rise of column-stores, they run on high performance engines that are tightly optimized for the efficient use of modern hardware. Hence, the micro-architectural behavior of modern OLAP systems remains unclear. This work presents the micro-architectural analysis of a breadth of OLAP systems. We examine CPU cycles and memory bandwidth utilization. The results show that, unlike the traditional, commercial OLTP systems, traditional, commercial OLAP systems do not suffer from instruction cache misses. Nevertheless, they suffer from their large instruction footprint resulting in slow response times. High performance OLAP engines execute tight instruction streams; however, they spend 25 to 82% of the CPU cycles on stalls regardless of the workload being sequential- or random-access-heavy. In addition, high performance OLAP engines underutilize the multi-core CPU or memory bandwidth resources due to their disproportional compute and memory demands. Hence, analytical processing engines should carefully assign their compute and memory resources for efficient multi-core micro-architectural utilization.