On Choosing a Load-Balancing Algorithm for Parallel Systems with Temporal Constraints

A key point in parallel systems design is the way clients requests are forwarded and distributed among the servers, trying to obtain the maximum throughput from them or, in other words, the load-balancing policy. Although it is a largely studied theme, with well accepted solutions, the inclusion of temporal constraints, also denoted as deadlines in this work, to the requests brings new complexities to the load-balancing problem: how to distribute the tasks and minimize the miss rate. The experiments describe along this paper attests that the workload variability plays a crucial role in this problem, pointing the big requests as the most critical elements. Our results also shows that even dynamic load-balancing algorithms are not able to reach an acceptable miss rate, since they handle both short tasks and big tasks the same way. Hence, we propose a new load-balancing algorithm, called ORBITA, which has a request identification and classification mechanism and an admission control module as well, restricting the number of big tasks within the system. This algorithm outperforms its competitors, which means that it has a bigger rate of tasks that end within the deadline, specially when the system is under high load. A prototype was also built in order to check the correctness of the simulation phase. The experiments were run against a benchmark tool, TPC-C, and all the results confirmed the previous assumptions, leading to the conclusion that it is a good practice to understand the system's workload in order to minimize the miss rate.