Adaptive and Scalable Predictive Page Policies for High Core-Count Server CPUs

Abstract

Increasing datacenter compute requirements has led to tremendous growth in the cadence of CPU cores on chip-multiprocessors. With large number of threads running on a single node, it is critical to achieve high memory bandwidth efficiency on large scale CMPs to support continued growth in the number CPU cores. In this paper, we present several mechanisms that improve the memory efficiency by improving the page hit rate for multi-core processors. In particular, we present memory page-policies that dynamically adapt to the runtime workload characteristics and use thread awareness to reduce contention between different memory address streams from the different threads. Unlike contemporary DRAM page policies such as static or timer-based, the proposed framework profiles the memory stream at runtime and uncovers opportunities to close or keep DRAM pages open, resulting in reduced page-conflicts and improved efficiencies. We implement the proposed policies in a cycle-accurate performance model simulating an 8-core processor. Our results show that the proposed adaptive page policies increase performance of high memory bandwidth workloads in SPECint2006 by up to 3%, and can attain 83% average performance relative to a “perfect” page prediction policy. We further show that the performance improvement from the techniques increases with the number of cores and with making the policies thread-aware in a many-core processor. The implementation cost of our techniques is extremely low, an area overhead of only 69 bits, making them extremely attractive for real-life products.