First-Class Operating System Management of Computational Accelerators

Our research investigates operating system mechanisms to manage hardware accelerators and high-bandwidth devices in a safe, fair, and protected manner while maintaining high performance. Computational accelerators, including general-purpose graphical processing units (GP-GPUs) and engines for encryption, compression, media transcoding, pattern matching, parsing, etc., are increasingly common in modern computer systems. For the data center, extremely fast devices, including system area networks and massively parallel storage, are also increasingly common. For the sake of safety and fairness, such accelerators must be managed by the operating system, but for the sake of performance, they must be accessible either directly from user-level applications or in streamlined OS code, with little or no coordination across cores. In some cases (GP-GPUs in particular), the conflict between fairness and performance is exacerbated by the opacity of proprietary library/driver/hardware interfaces. Our research seeks a balanced solution to these conflicting goals through:

  1. an operating system resource management architecture that allows direct user-level access in the common case, but intercedes in the existing accelerator access path when necessary to delay and re-order requests;
  2. a scheduling discipline that minimizes re-apportionment of resources across cores, to avoid synchronization and cache misses in the common case;
  3. a tool chain that uncovers hidden interface semantics required for resource management, together with a characterization of the information needed from vendors in the future;
  4. an integrated management and scheduling strategy across the full set of computational resources in a given system.

Publications:

Open-Source Software:

Project Members:

Michael L. Scott (PI), University of Rochester
Kai Shen (co-PI), University of Rochester
Mohammad Hedayati-Goudarzi, University of Rochester
Joesph Izraelevitz, University of Rochester

Alumni:

Konstantinos Menychtas, Google
Hammurabi Mendes, Davidson College

Support:

This project is supported by the National Science Foundation under Grant No. 1319417. Link to the award abstract.

Note:

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.