In this paper, the authors discuss issues in scheduling (that is, match making) of job streams with end-to-end service-level agreements (SLAs) with agreed-upon quality of service (QoS) and allocation of resources in a distributed cloud environment such as Amazon EC2. Here, an SLA gives earliest start time, an execution time, and a deadline, and a failure is characterized by a missed deadline. Using the mathematical programming model MapReduce, the authors formulate this mapping (match making plus scheduling) as a constraint programming (CP) problem that minimizes failures. The solutions are implemented on Apache Hadoop, a Java-based implementation of MapReduce. The optimization issues are handled using the IBM CPLEX optimization programming language (OPL). The mapping part of the paradigm is called the MRCP-RM algorithm. The resource management part of the paradigm is given via another algorithm called HCP-RM. In HCP-RM, a master node schedules submitted jobs by spawning threads on various nodes. However, the architecture for components of MapReduce--namely input reader, map function, partition function, compare function, reduce function, and output writer--and the communications mode between these functions have not been clearly stated in this paper.

To verify their propositions, the authors set up metrics in terms of job arrival rate, task execution time, earliest start time, number of resources, and failures. The authors describe their prototype experimental setup using Hadoop 1.2.1 and the simulation setup. Based on their experiments, the authors conclude that reduced failure rates were evident.

This well-written paper would interest researchers dealing with distributed processing, clusters, and job scheduling.