MIDAS: An Execution-Driven Simulator for Active Storage Architectures

Many applications today are highly data intensive and have stringent performance requirements. In order to meet the performance demands of these applications, we need to optimize both the processing and I/O subsystems. One promising approach to optimize performance is to use “Active Storage” systems, where we use disk drive controllers and storage array controllers as offload processors for the data intensive parts of an application and exploit DataLevel Parallelism (DLP) across the ensemble of processing components. From the architecture viewpoint, the design space of such active storage systems is large. There are a number of choices for the microarchitecture of the processors at the disk, storage array controller, and host, the organization of the electro-mechanical parts of the disk drive, and the characteristics of the interconnection network between these components. Since these design choices can have a significant impact on performance, we need a simulator that can give us detailed insights into the behavior of these systems. This paper presents the Modeling Infrastructure for Dynamic Active Storage (MIDAS). MIDAS is an accurate execution-driven simulator that captures both the processing and I/O behavior of active storage systems. We describe the design of MIDAS, providing details about the various simulation models and how they interact. We then present three case studies that illustrate how MIDAS can be used to study architectural design tradeoffs in active