Collaborative Research: CNS Core: MEDIUM: MASON: Memory at Scale on Networks

University Of California-Santa Cruz, Santa Cruz CA

Investigators

Abstract

Many of the essential services on which Americans depend---including finance, education, and health---require processing large amounts of data that far outstrip the capabilities of a single computer. Historically, to build such services, system designers employ what is called a ``scale-out, shared-nothing architecture,'' in which data is spread across an ensemble of computers, each of which processes the data locally when it can, and coordinates its actions by exchanging messages with other computers. However, as increasingly data-hungry applications require these systems to process ever-larger inputs in shorter amounts of time, these shared-nothing designs have begun to show their weaknesses: the interfaces they provide are too difficult to evolve as applications change; they assume that computers have roughly equivalent capabilities, which is less and less true as processors become more diverse; and they place undue burden on the application developer to decide exactly where data should be placed and when it should move. The traditional alternative to the shared-nothing architecture is "shared-everything", in which all computers accessed the same shared memory resource. However, in the past, shared-everything designs were dismissed as infeasible. We argue that several emerging technological trends, including increased network speeds, new types of persistent memory, and the heterogenization of computing make the conditions ripe to revisit the alternative shared-everything approach. This project will develop a tightly coupled co-design of novel operating system and network, leveraging these trends, to realize a practical shared-everything architecture. If the project succeeds, it will result in a radically different, and dramatically simpler, programming model for building distributed systems. The proposed system, named MaSON(Memory at Scale On Networks), will explore how to provide reliable, low-latency access to large pools of remote, disaggregated memory from an increasingly distributed and heterogeneous computing environment. In particular, it will explore fundamental questions about: (i) the design of memory protocols over network fabrics, (ii) the appropriate role of computation-enhanced networks, and (iii) providing fault-tolerance and robustness for persistent memory. Programming models for disaggregated, persistent memory will also be considered. The work will result in a network/operating system co-design around a shared notion of long-lived data references, supporting remote access to disaggregated memory. The project's output is broadly applicable to industry, government and academia dealing with large datasets. 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.