SPX: Parallel Models and Algorithms for Emerging Memory Systems

Carnegie Mellon University, Pittsburgh PA

Investigators

Abstract

With the advent of highly-parallel many-core (computing) machines, memory has increasingly become a limiting factor in continued performance improvement and scalability, in terms of energy usage, component density, latency, bandwidth, and reliability. To help deal with these and other problems, the semiconductor industry has been developing new byte-addressable nonvolatile random access memory (NVRAM) technologies. These offer the promise of significantly lower energy needs and higher density than standard dynamic random access memory (DRAM), while not losing their state on power loss. However, in NVRAM technology, operations that write to memory are more costly in terms of throughput and energy than operations that read from memory. This project is developing and testing new abstractions for emerging large and extreme-scale computer systems based on NVRAM, and how to effectively leverage this asymmetry for better performance in large computing systems. The focus will be on combining theory and practice, and considering issues across multiple levels of abstraction, from the hardware itself, to high-level algorithms and programming models. The project will include an educational component that will teach students about the new technology and how to effectively use it. The project consists of three main components: (1) developing methodologies for systems combining volatile and nonvolatile memory that allow individual processors to fail while permitting the overall system to continue correctly, (2) developing efficient algorithms and caching policies for settings where writes are more expensive than reads, and (3) developing techniques to take advantage of the significant computing capability in each memory controller. In the first component, the project is studying how to automatically convert arbitrary concurrent programs into a setting where processors can fail so that the overhead for both running the converted program and recovering from a failure is low. In the second component, the project is developing general purpose techniques to reduce the numbers of writes compared to reads, or reduce the fraction of the memory that needs to be written to, and applying the techniques across a broad class of algorithms. The research team will both develop theory and experimentally measure the effectiveness of these techniques and algorithms. In the third component, the project is looking at how to use the memory controllers to reduce the cost of fault tolerance and allow for weaker memory models, with the purpose of scaling to large systems. A key intellectual challenge is to ensure that the models, techniques, and algorithms are simultaneously simple, elegant, and practical. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.