Brownout^CC: Cascaded Control for Bounding the Response Times of Cloud Applications

摘要

Cloud computing has emerged as an inexpensive and powerful computing paradigm, to the point that now even applications with hard deadlines are executed in the cloud. It may happen, due to unexpected events, that an application becomes popular and receives a lot of attention and client requests in a short period of time. Provisioning computing capacity for such applications is quite a difficult task, because content popularity cannot be easily predicted. One of the main problems in case content has to be served with a hard deadline is to ensure that this deadline is respected, even in the presence of popularity spikes. To this end, partial computation and graceful degradation were exploited, originating the brownout framework. Applications would degrade the user experience in the presence of load variations, to guarantee that deadlines are met. Two different control paradigms were applied to brownout: discrete-time control of optional content percentage over a period and event-based queue management. The first one had reasonable performance providing formal guarantees about the solution. The second one was able to improve the performance and keep the response time at the setpoint better, but suffered from the drawback of not providing formally-grounded mathematical guarantees. In this work we combine the best of both worlds, providing a cascaded controller for brownout, based on a more precise model of the cloud application with respect to the original design. The Brownoutn^CCncontroller achieves performance comparable with the event-based version, without sacrificing formal guarantees.