Workload Placement with Bounded Slowdown in Disaggregated Datacenters

Disaggregated Data Center (DDC) is a modern datacenter architecture that decouples hardware resources from monolithic servers into pools of resources that can be dynamically composed to match diverse workload requirements. While disaggregation improves resource utilization, it could negatively impact workload slowdown due to the latency of accessing disaggregated resources over the datacenter network. To this end, we consider CPU and memory disaggregation and conduct measurements to experimentally profile several popular datacenter workloads in order to characterize the impact of disaggregation on workload execution slowdown. We then develop a workload placement algorithm, called Iterative Rounding-based Placement (IRoP), that given a set of workloads, determines where to place each workload (i.e., on which CPU) and how much local and remote memory allocate to it. The key insight in designing IRoP is that the impact of remote memory latency on slowdown can be substantially masked by assigning workloads to higher-performing CPUs, albeit at the cost of higher energy consumption. As such, IRoP aims to find a workload placement that minimizes the DDC energy consumption while respecting a bounded slowdown for each workload. We provide extensive simulation results to demonstrate the flexibility of IRoP in providing a wide range of trade-offs between energy consumption and workload slowdown. We also compare IRoP with several existing baselines. Our results indicate that IRoP can reduce energy consumption and slowdown in the considered scenarios by up to 8% and 12%, respectively.