1. A Load Balancing Framework for Adaptive and Asynchronous Applications Kevin Barker, Andrey Chernikov, Nikos Chrisochoides,Keshav Pingali ; IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 15, NO. 2, pp. 183-192 Feb. 2004 Presented by 張肇烜

2. Outline Introduction Software Framework and Programming Model Load Balancing Library Performance Evaluation Related Work Conclusions

3. Introduction Stop-and-repartition method: –Load-balancing is accomplished by dynamically repartitioning the data after the global synchronization phases. –Synchronization overhead can overwhelm the benefits of improved load balance.

4. Introduction (cont.) PREMA: –Parallel Runtime Environment for Multicomputer Applications. –Single-sided communication similar to that provided by Active Messages. –A framework which allows for the easy and efficient implementation of customized dynamic load balancing algorithms.

5. Introduction (cont.) Active Messages: –An efficient communication architecture for multiprocessors. –Efficient supports a variety of parallel program models, including message passing, shared memory and dataflow.

6. Software Framework and Programming Model Data Movement and Control Substrate (DMCS) : –DMCS provides a flexible and easy to understand application program interface for one-sided communication operations.

7. Software Framework and Programming Model (cont.) Mobile Object Layer (MOL): –A runtime substrate for parallel adaptive and irregular computations. –A mobile object may be referenced by any processor in the parallel system by using its associated mobile pointer.

8. Load Balancing Library PREMA’s ILB component library is built using the framework provided by DMCS and the MOL. Supports the rapid development of algorithms, allowing researchers to experiment without modification of existing application code.

9. Load Balancing Library (cont.) Load balancing methods: –Diffusion –Master-worker –Work-stealing –Multilist

10. Performance Evaluation

11. Performance Evaluation (cont.)

13. Related Work

14. Related Work (cont.) Zoltan: –Zonltan provides graph-based partitioning algorithms and several geometric load balancing algorithms. CHARM++: –CHARM++ is built on an underlying language which is a dialect of C++ and provides extensive dynamic load balancing strategies.

15. Conclusions We have demonstrated performance improvements of 15% over traditional stop and repartition methods, 30% over intrusive explicit load balancing methods, and 42% over no load balancing on configurations of 128 processors.