Load Balancing Physics Routines

This paper presents three algorithms for load balancing physics routines with dynamic load imbalance. Results are presented for the two most computationally demanding load imbalanced physics routines (short wave radiation and convection) in the UKMO's forecast and climate models. Results show between 30% and 40% performance improvement in these routines running on the UKMO's Cray T3E. 1 Introduction The physics phase of forecast and climate models is particularly amenable to large scale parallelism as, typically, each grid column can be independently computed. However, the existence of load imbalance in this phase is well known [6, 8] and has been widely acknowledged as being one of the major factors in loss of performance, particularly when using large numbers of processors, for example see [5, 9, 14, 15]. Interestingly, the same load imbalance problem is also observed in data assimilation [9, 10]. A domain decomposition approach is usually employed during the physics phase of paral..