Thursday, February 4, Volen 101, 2:10-3:10 pm. (Refreshments at 2:00pm)
Distributed server systems are commonplace today. Examples include distributed web servers, distributed database servers, distributed batch processing servers for computationally intensive parallel or sequential jobs, etc. With every distributed system comes the question of what the Task Assignment Policy should be, i.e., what should be the rule for assigning tasks to hosts so as to maximize the performance of the system.
We focus on one particular simple architecture for a distributed server, common for batch processing servers, where task sizes are known, and server machines process the tasks in their queues in First-Come-First-Serve (FCFS) order. We compare various commonly used task assignment policies within this model, and introduce a new one: SITA-E. We use analysis and simulation to show that the question of which task assignment policy performs best is not obvious and depends on the task size distribution. Furthermore, choosing the right task assignment policy can improve performance by several orders of magnitude.
JOINT WORK WITH: Mark Crovella (BU) and Cristina Murta (Brazil)
Host: Liuba Shrira