<span class="var-sub_title">Advanced Architecture Testbeds: A Catalyst for Co-design Collaborations</span> SC18 Proceedings

The International Conference for High Performance Computing, Networking, Storage, and Analysis

Advanced Architecture Testbeds: A Catalyst for Co-design Collaborations


Authors: Kevin Barker (Pacific Northwest National Laboratory), Jeffrey Young (Georgia Institute of Technology), Alice Koniges (University of Hawaii), Jeffrey Vetter (Oak Ridge National Laboratory), James Laros (Sandia National Laboratories)

Abstract: Our supercomputing community is in a renaissance of computer and system architectures. The explosion in architectural diversity is supported by the use of testbeds to analyze application performance and energy efficiency on real system software stacks. Reasons for active participation in this BoF session include alignment of efforts, increased coverage of diverse architectures, sharing of lessons learned with advanced architectures, support for establishing and useing common proxy application and benchmarks. This BoF will bring together advanced architecture testbed efforts: at CENATE – PNNL, HAAPS – SNL, Rogues Gallery – GA-Tech, HPCMP – testbeds at several sites, and ExCL at ORNL.

Long Description: The supercomputing community is in the midst of a period of unprecedented architectural innovation. The explosion in architectural diversity leads to a number of challenges, including the challenge of understanding the potential performance impact of new architectural technologies on workloads of interest. To address these challenges, a number of architectural testbed efforts have been established, including CENATE at Pacific Northwest National Laboratory, HAAPS at Sandia National Laboratory, Rogues Gallery at Georgia Tech, HPCMP at several testbed sites, and ExCL and Oak Ridge National Laboratory. This BoF brings together researchers and practitioners involved in these and other efforts to ensure alignment of efforts, increased coverage of diverse architectures, sharing of lessons learned in advanced architectural exploration, and support for efforts to establish and use common proxy applications and benchmarks.

Our BoF aims to cover a number of topics, foremost among them the dimension of architectural diversity. As architectural designers grapple with increased demands on performance and energy efficiency, evermore complex processor, memory, and network designs result. In the processing space, compute elements such as multi-core and many-core CPUs, GPUs, FPGAs, and other advanced accelerators are increasingly finding their way into HPC system installations. Similarly, memory technologies including conventional DRAM, stack memory (e.g., HBM and HMC), non-volatile memory and storage-class memory, and near-memory logic are leading to heterogeneity in the memory architecture and placing additional burdens on applications seeking to make most effective use of limited memory resources. In the interconnection network space, electrical packet-switched networks, optical circuit-switched networks, and advanced NIC and switch designs are leading to new network topologies and performance characteristics that are having an impact on application communication behavior. BoF participants will have an opportunity to describe the testbed systems that are currently in place as well as proposed future testbeds targeting emerging technologies. In this way, the community will gain a deeper understanding of the current gaps in testbed installations.

The BoF will also provide the opportunity for practitioners to identify currently used proxy applications and benchmarks, as well as the current gaps that need to be addressed with additional development. Proxy applications are typically provided with source code to allow for architecture-centric implementations and optimizations. However, what is required to complete the definition of a proxy application is a driver problem to ensure the computation, memory access, and communication patterns are representative of realistic workloads. Suitable definitions of proxy applications and driver problems allow for cross-comparison of testbed measurements and will facilitate interaction among the community.

Hardware/software co-design efforts are supported by architectural modeling and simulation tools, but the development and evaluation of software stacks to support novel architectural technologies require the availability of testbed systems. Such access is critical to ensure the timely availability of new architectures and technologies. Key challenges in this area include the need to work with immature software stacks, as well as non-disclosure agreements that must be in place to support early access to pre-production hardware. This BoF will provide an opportunity for participants to share lessons learned in these areas.





Back to Birds of a Feather Archive Listing