Black Holes , Sacri cial Lambs , and a Robust Approach toTransaction

We consider a transaction processing system with multiple servers (e.g. computer systems) in which transactions may fail at the diierent servers independently (due to the heterogeneity), and only a subset of the total transactions are aaected. Due to this nature, a system is faulty for some transaction classes and good for others. Because failed transactions have a lower processing time, algorithms such as JSQ (join shortest queue) will route a disproportionate number of transactions to the failure prone systems , causing an artiicially high transaction failure rate. We call this the black hole phenomena. In this paper, we present two algorithms for transaction routing to detect and control the black hole problem. The rst algorithm routes transactions to systems based on server queue lengths and perceived failure rate. It also ensures that error prone systems are not completely ignored; occasional transactions (called sacri-cial lambs) probe systems to detect repair. Detailed simulation results are presented using realistic system and workload models. Our results show that the new algorithm reduces the number of transaction failures per class compared to the well known JSQ algorithm while still performing dynamic load balancing. The second algorithm is based on a non-linear optimization problem which is presented under simpliied system and workload models. This algorithm studies the analytical basis of the black hole problem to provide insight and comparisons. The numerical results indicate that the second algorithm is better in controlling the failure rates than the rst.