Exploiting multi-core nodes in peer-to-peer grids

Highlights

• We manage multi-core nodes efficiently in a decentralized way in peer-to-peer grids.

• We schedule multi-threaded jobs in P2P grids effectively.

• We suggest a parameterized performance model for concurrent jobs in a multi-core.

• Our schemes perform competitively to centralized methods, confirmed by simulations.

• Large-scale simulations show that our schemes scale well.

Abstract

While the majority of CPUs now sold contain multiple computing cores, current grid computing systems either ignore the multiplicity of cores, or treat them as distinct, independent machines. The latter approach ignores the resource contention present between cores in a single CPU, while the former approach fails to take advantage of significant computing power. We provide a decentralized resource management framework for exploiting multi-core nodes to run multi-threaded applications in peer-to-peer grids. We present two new load-balancing schemes that explicitly account for the resource sharing and contention of multiple cores, and propose a parameterized performance prediction model that can represent a continuum of resource sharing among cores of a CPU. We use extensive simulation to confirm that our two algorithms match jobs with computing nodes efficiently, and balance load during the lifetime of the computing jobs.

Introduction

The majority of CPUs now sold contain multiple computing cores. However, even though multiple cores can often handle multiple tasks in parallel, overall performance with unoptimized software support can be poor. For example, Moore  [40] shows that multi-core processors cannot guarantee increased performance for supercomputing applications, and that some applications may even have worse performance than with single-core nodes. Given that it is not easy to utilize multi-core environments even on a single host, it is not surprising that there has been little research on integrating multi-core desktop machines into desktop grids. We address the problem of resource management for multi-core grids both to effectively utilize all available grid resources, including cores, and to run jobs that request multiple cores (presumably because they are multi-threaded).

Multi-core grids pose many challenges as follows. First, when a multi-core node runs jobs but all cores are not used, the number of free cores and amount of available shared resources should ideally be advertised to the grid system for matchmaking with future incoming jobs. Unfortunately, representing the dynamic aspects of node capability in current peer-to-peer grid systems is not straightforward. Second, running concurrent jobs on a multi-core node can result in contention for shared resources such as memory, cache, etc. There is currently no simple analytical model for this situation.

In this paper, we make the following contributions. First, we propose two new dynamic resource management schemes, Dual-CAN and the Balloon Model, which account for multi-core machines in a peer-to-peer (P2P) grid. The key idea behind both schemes is to use distinct logical peers to represent a single physical multi-core node. One logical peer represents the static and maximum node capability, and the other expresses the current available amount of resources. Second, we develop a scheme for efficiently mapping jobs to multi-core nodes within the new resource management frameworks. We extend our existing single-core approaches to aggregating information and representing global grid state, and we describe a “job-pushing” algorithm that efficiently balances load in both single-core and multi-core machines. Third, we suggest a new approach to modeling the performance of multi-core nodes, using a penalty factor to account for resource contention. We collected experimental data by running various benchmark tests, and interpolate those results to predict resource contention effects. We use the penalty factor to give less priority to nodes running memory-intensive jobs in our job scheduling algorithm to avoid performance degradation due to resource contention. Finally, we describe extensive simulation results that show that our new approaches outperform multi-core oblivious approaches.

This paper has unique contributions different from our previous work. Our initial work  [23], [24] gave a foundation of P2P-based desktop grid computing framework, and the advanced load balancing algorithm we proposed in  [22] improved job allocation performance. However, those approaches are limited by an assumption that nodes have only a single-core CPU; this paper overcomes this limitation to extent our resource management framework by exploiting multi-core CPUs for multi-threaded jobs. Most similar our previous work is  [26], which deals with job scheduling schemes across different types of computing elements (e.g. scheduling amongst General Purpose GPUs and CPUs). This paper focuses on multi-core CPUs which have homogeneous computing characteristics across cores.

The rest of the paper is organized as follows. Section  2 describes the Content-Addressable Network (CAN)  [43]-based peer-to-peer grid system within which we implement our new techniques. We present the distributed resource management scheme in Section  3. Section  4 presents the performance prediction model for concurrent jobs in multi-core nodes, and Section  5 shows our performance results. Section  6 addresses remained discussion. Section  7 summarizes related work, and we conclude in Section  8.