The advantage of dynamic tuning in distributed asymmetric systems

An analytical performance modeling approach is presented for load sharing policies that schedule jobs based on a global system state. The main contributions of this work are: the derivation of analytical bounds to compute the average delay in large asymmetric systems under dynamic load sharing control, when job routing is based on the global system state; the experimental derivation of parameters which affect the implementation of load sharing policies at the user level; the demonstration that load sharing is effective even in the presence of high overheads; and the demonstration that in the asymmetric environment under study, carefully tuned algorithms for load sharing provide a significant improvement in performance over simpler algorithms