Towards dynamic reconfigurable load-balancing for hybrid desktop platforms

s the Pus using the OpenCL API as the platform independent programming model. It has the proposal to extend OpenCL with a module that schedule and balance the workload over the CPU and GPU for the specific case study in a high level. Starting with an initial scheduling configuration just when the application starts, an online profiler monitors and stores tasks' execution times and platform conditions. Since the tasks are nondeterministic, during application execution, a reconfigurable dynamic scheduling is performed considering changes on runtime conditions. Figure above depicts the approach; and bellow, the tow-phase scheduling approach. First Assignment: the first guess faces a multidimensional scheduling problem with NP-hard complexity and become more complex when dealing with more than two PUs and several tasks. To optimize, this assignment is based on heuristics taking into account the performance benchmark described on the right. The first scheduling is performed in a static...

• Type: Journal
• Year: 2010
• Where: IPPS
• Authors: Alécio Pedro Delazari Binotto, Carlos Eduardo Pereira, Dieter W. Fellner
• Keywords: Distributed And Parallel Computing, Initial Scheduling Configuration, IPPS 2010, Reconfigurable Dynamic Scheduling, Runtime Conditions