In order to satisfy increasing demands for Cloud services, modern computing systems are often massive in scale, typically consisting of hundreds to thousands of heterogeneous machine nodes. Parallel computing frameworks such as MapReduce are widely deployed over such cluster infrastructure to provide reliable yet prompt services to customers. However, complex characteristics of Cloud workloads, including multi-dimensional resource requirements and highly changeable system environments, e.g. dynamic node performance, are introducing new challenges to service providers in terms of both customer experience and system efficiency. One primary challenge is the straggler problem, whereby a small subset of the parallelized tasks take abnormally longer execution time in comparison with the siblings, leading to extended job response and potential late-timing failure.
The state-of-the-art approach to straggler mitigation is speculative execution. Although it has been deployed in several real-world systems with a variety of implementation optimizations, the analysis from this thesis has shown that speculative execution is often inefficient. According to various production tracelogs of data centers, the failure rate of speculative execution could be as high as 71%. Straggler mitigation is a complicated problem in its own nature: 1) stragglers may lead to different consequences to parallel job execution, possibly with different degrees of severity, 2) whether a task should be regarded as a straggler is highly subjective, depending upon different application and system conditions, 3) the efficiency of speculative execution would be improved if dynamic node performance could be modelled and predicted appropriately, and 4) there are other types of stragglers, e.g. those caused by data skews, that are beyond the capability of speculative execution.
This thesis starts with a quantitative and rigorous analysis of issues with stragglers, including their root-causes and impacts, the execution environment running them, and the limitations to their mitigation. Scientific principles of straggler mitigation are investigated and new algorithms are developed. An intelligent system for straggler mitigation is then designed and developed, being compatible with the majority of current parallel computing frameworks. Combined with historical data analysis and online adaptation, the system is capable of mitigating stragglers intelligently, dynamically judging a task as a straggler and handling it, avoiding current weak nodes, and dealing with data skew, a special type of straggler, with a dedicated method. Comprehensive analysis and evaluation of the system show that it is able to reduce job response time by up to 55%, as compared with the speculator used in the default YARN system, while the optimal improvement a speculative-based method may achieve is around 66% in theory. The system also achieves a much higher success rate of speculation than other production systems, up to 89%.