Supporting fluctuating transactional workload

Abstract

This work deals with a fluctuating workload as in social applications where users interact each other in a temporary fashion. The data on which a user group focuses form a bundle and can cause a peak if the frequency of interactions as well as the number of users is high. To manage such a situation, one solution is to partition data and/or to move them to a more powerful machine while ensuring consistency and effectiveness. However, two problems may be raised such as how to partition data in a efficient way and how to determine which part of data to move in such a way that data are located on one single site. To achieve this goal, we track the bundles formation and their evolution and measure their related load for two reasons: (1) to be able to partition data based on how they are required by user interactions; and (2) to assess whether a machine is still able of executing transactions linked to a bundle with a bounded latency. The main gain of our approach is to minimize the number of machines used while maintaining low latency at a low cost.