Load balancing for massively-parallel soft-real-time systems
Global load balancing, if practical, would allow the effective use of massively-parallel ensemble architectures for large soft-real-problems. The challenge is to replace quick global communications, which is impractical in a massively-parallel system, with statistical techniques. In this vein, the author proposes a novel approach to decentralized load balancing based on statistical time-series analysis. Each site estimates the system-wide average load using information about past loads of individual sites and attempts to equal that average. This estimation process is practical because the soft-real-time systems of interest naturally exhibit loads that are periodic, in a statistical sense akin to seasonality in econometrics. It is shown how this load-characterization technique can be the foundation for a load-balancing system in an architecture employing cut-through routing and an efficient multicast protocol.
- Research Organization:
- Stanford Univ., CA (USA). Dept. of Computer Science
- OSTI ID:
- 6391821
- Report Number(s):
- AD-A-200912/4/XAB; STAN-CS-88-1222
- Country of Publication:
- United States
- Language:
- English
Similar Records
Data Locality Enhancement of Dynamic Simulations for Exascale Computing (Final Report)
Scalable Molecular Dynamics