Simulation for disaggregated architecture

Abstract

The increasingly more data-intensive workloads imposed on cloud computing are calling for fundamental changes in the computer architecture level. The current machine-oriented architectures, where hardware resources, such as main processors, accelerators, and memory, are rigidly tied to a machine, do not allow flexible resource management. Disaggregated architectures propose breaking the tight bonding of the compute and storage components inside a machine. Each component becomes a node on a high- bandwidth fabric interconnect network and is managed in resource pools of its type. The disaggregated architecture cloud can be flexibly scaled in a fine-grained manner to react to the demanded workload. However, the architectures are often inaccessible to researchers as they employ specialized hardware or require a large-scale investment to set up a testing environment. This dissertation presents the design and implementation of a simulator for disaggregated architectures called DisaggSim. We generalize existing disaggregated architectures and design an extensible and detailed simulation model. Also, we implement the simulator on top of gem5, a broadly-used system simulator in the research community. Through our evaluation, we demonstrate that DisaggSim can be used to evaluate various performance characteristics of a given disaggregated architecture specification. We publicly release the simulator to facilitate rapid prototyping and evaluation of research on disaggregated architectures.