II-New: A Computational Infrastructure for Scalable Transactional Memory Abstractions in Managed Languages

Purdue University, West Lafayette IN

Investigators

Abstract

This proposal supports the acquisition of an Azul Vega 3 compute appliance for research on developing an open-source general-purpose transactional infrastructure for high-level managed languages. This appliance is equipped with hundreds of processor cores, substantial memory, and bandwidth availability. The cores compute in a flat SMP configuration, simplifying programmability and management. The appliance has a number of novel features including hardware support for concurrent garbage collection, and optimistic lock elision that provide critical building blocks for efficient transactional memory (TM) support. The project's research agenda is to explore new approaches to scalable parallel program construction that exploits underlying hardware primitives that mask or mitigate critical bottlenecks that have thus far hampered effective scalable software transactional memory systems. The project will use the novel features of the Azul appliance to (i) explore new language abstractions with special focus on scalability of transactional applications; (ii) understand how primitive support for garbage collection and transactions influence algorithm and application development built using these abstractions; (iii) examine the impact of these mechanisms on compiler and runtime design; and (iv) develop a large number of benchmarks and workloads to better understand TM performance characteristics. This project is part of a larger NSF-funded effort to catalyze TM research, which in turn is a key technique to improving software development for general-purpose parallel computers. Widespread adoption of new scalable concurrency abstractions such as transactional memory is an essential component in the technology roadmap of major industry players, and the focus of much attention in academia.