Memory-Centric Architectures for the Cloud and HPC
Authors: Dr. Sudhakar Yalamanchili (Georgia Institute of Technology)
Abstract: Computing systems are at an inflection point where fundamental changes in technology and applications are poorly matched to memory systems of modern computing-system designs. At all levels of the system, memory is becoming the engine of performance growth leading to new memory hierarchies, near data computing architectures, and new data centric programming models and system management frameworks.
The goal of this BoF is to bring together application developers, system software developers, memory system designers, and system architects to create community that can identify major trends and technical challenges, and catalyze collaborative opportunities between industry, academia, and government to address them.
Long Description: Computing systems are an inflection point where fundamental changes in technology and applications are poorly matched to memory system of modern computing-system designs. Technologically, the energy costs of data movement are exceeding the energy costs of computation. Many systems are experiencing declining ratios of compute bandwidth to memory bandwidths due to use of accelerators and core scaling while growth in compute and memory capacity is challenged due to limited growth in available power. Concurrently, emerging and future applications are characterized by sparse data structures, irregular data access patterns, poor temporal and spatial locality, streaming behaviors, low operation density, and high degrees of time-varying nested parallelism. Data management dominates the energy cost of execution, dominates the execution time of an application, and dominates the complexity of application software, while advances non-volatile storage technology are revolutionizing memory hierarchies and management and processing of persistent data. Thus, at all levels of the system, memory is becoming the engine of performance growth, and unless we re-architect systems to reflect this fact, growth in performance and consequent economic benefits will stall.
Advances are being made independently on multiple fronts – new storage device technologies, new memory hierarchies, and new computing hierarchies (e.g., near data computation). Fundamental changes in systems architectures will require coordinated effort across all layers of the software and hardware stacks. The BoF can provide a means for the communities to begin interacting towards a high level shared vision of memory-centric system architectures.
The goal of this BoF is to create a community of researchers and practitioners who can i) share experiences and insights, ii) identify major trends and technical challenges, iii) inspire the community to contribute to an open source ecosystem of shared software, e.g., performance analysis tools and applications benchmarks, iv) catalyze advanced efforts within their organizations, and iv) open collaborative opportunities between industry academia, and government.
Conference Presentation: pdf
Birds of a Feather Index