CARMA: Contention-aware Auction-based Resource Management in Architecture

TL;DR

This paper introduces CARMA, a game-theoretic, auction-based resource management system for CMPs that dynamically allocates resources by considering contention, improving scalability and performance over traditional static methods.

Contribution

It proposes a novel contention-aware auction mechanism for resource management in CMPs, enabling distributed, scalable, and adaptive resource allocation based on application bidding strategies.

Findings

Outperforms static and traditional resource allocation methods.

Scalable and effective in cache and processor congestion scenarios.

Enables applications to learn optimal resource strategies through repeated interactions.

Abstract

As the number of resources on chip multiprocessors (CMPs) increases, the complexity of how to best allocate these resources increases drastically. Because the higher number of applications makes the interaction and impacts of various memory levels more complex. Also, the selection of the objective function to define what \enquote{best} means for all applications is challenging. Memory-level parallelism (MLP) aware replacement algorithms in CMPs try to maximize the overall system performance or equalize each application's performance degradation due to sharing. However, depending on the selected \enquote{performance} metric, these algorithms are not efficiently implemented, because these centralized approaches mostly need some further information regarding about applications' need. In this paper, we propose a contention-aware game-theoretic resource management approach (CARMA) using market auction mechanism to find an optimal strategy for each application in a resource competition game. The applications learn through repeated interactions to choose their action on choosing the shared resources. Specifically, we consider two cases: (i) cache competition game, and (ii) main processor and co-processor congestion game. We enforce costs for each resource and derive bidding strategy. Accurate evaluation of the proposed approach show that our distributed allocation is scalable and outperforms the static and traditional approaches.