CSR: Small: Delegation Systems for Efficient and Safe Multi-Core Programming

University Of Illinois At Chicago, Chicago IL

Investigators

Abstract

This project helps software on highly parallel (multi-core) computers access shared data structures more efficiently. The conventional approach is based on individual cores taking turns to access the data structure. However, every time the data structure changes hands there is a considerable slowdown, while other cores waste valuable time awaiting their turn. Delegation is an alternative approach. Here, a single core (server) is in charge of the data structure, and performs work as requested by other cores. Focusing on one job makes the server more efficient, often by far outweighing the costs of delegation, given a high-performance delegation design. The project is divided into three main tasks. First, delegation should theoretically be 10x faster than it is today, using current hardware. The project investigates what currently holds back delegation performance, and reaches for the fundamental performance limits of delegation. Second, delegation currently relies on some degree of hand-tuning of parameters to achieve peak throughput. The project investigates mechanisms to allow delegation to automatically adjust to current conditions, improving ease of use. Finally, delegation today is fundamentally client-server oriented. The project investigates whether a fully distributed model can achieve further gains, both theoretically and in practice. The aim of the project is to make delegation both extremely high-performing, and easy to use. If the project is successful, and if delegation as a result sees adoption among software developers, the result will be programs running considerably faster on the same hardware. This, in turn, reduces the amount of computing hardware required to perform the same amount of work, thereby reducing both equipment cost and electricity consumption. Software artifacts produced in the course of this project will be made publicly available on the research lab's github page: http://github.com/bitslab, where they will remain well beyond the project period. 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.