Workflow scheduling in computational grids: Opportunistic vs. planned

One of the prime-most issues that is faced by the parallel systems in the past and is inherited by the grid systems of today, is the mapping and the execution of jobs on the available machines. Various scheduling approaches that were suggested in classical parallel systems literature are adopted for the grid systems with appropriate modifications. Although these modifications made them suitable for execution in grid environment, these approaches failed to deliver on the performance front. Dynamic availability of resources, data dependence among the tasks and communication latencies are some of the factors that negatively affected the outcome of these scheduling algorithms. Various grid management systems implemented their custom tailored versions of scheduling algorithms for task execution in the grid. These approaches suffer from various shortcomings such as inefficient utilization of the execution queues and/or failure to cater for the dependencies among the tasks in a job. In this paper, we will compare the two prevalent scheduling models in the grid; opportunistic and planned. Opportunistic model is used in DAGMan for scheduling workflow applications, which in-turn is an integral component of Condor and Condor-G. Planned scheduling, on the other hand, is used by the grid management architectures such as GrADS and MARS.