Optimal Dynamic Resource Provisioning to Meet QoS Requirements in Cloud
Open AccessAs cloud applications are becoming increasingly deadline-sensitive, meeting desired QoS requirements is more critical, especially in shared and heterogeneous clusters. It has been shown that a few slow tasks, called stragglers, could significantly adversely impact job execution times. Moreover, poor scheduling of data analytics applications can lead to inefficient resource usage, and eventually hurt system performance. One way to mitigate stragglers is by launching extra attempts (clones) for each task upon job submission. In addition, ideally, in homogeneous clusters where all machines have the same configuration, all tasks will finish processing at the same time. However, in real systems, homogeneity is not guaranteed as servers are made of different generation and differ in terms of capacities of the processor, speed, memory, storage and networking bandwidth. It has been shown that the divergence and uncertainty in task processing times resulting from machine heterogeneity and execution dynamics could lead to significant performance degradation of up to 75% due to the existence of residual workload and stragglers.In the first part of this dissertation, we consider dynamic cloning to meet application deadlines in cloud. The performance of parallel cloud processing frameworks when processing data are often negatively affected by stragglers, which make them unsuitable to latency-sensitive jobs that require job completion time guarantees. We propose Shed, an optimization framework that leverages dynamic cloning to jointly maximize jobs’ Probability of Completion before Deadline (PoCD) by fully utilizing the available resources. Our work includes a novel online scheduler that dynamically recomputes and reallocates resources during a job’s execution for PoCD maximization. The results show that Shed is able to leverage cloud resources and maximize the percentage of jobs that meet their deadlines – up to 100% in our experiments compared to typically around 60% and 40% for another cloning approach called Dolly, and Hadoop with speculation enabled, respectively.The second part of this dissertation proposes Shed+, which is an enhanced version of Shed. Shed+ differs from Shed in the sense that it does more fine-grained allocation of resources to jobs. Shed+ utilizes dynamic speculation that aims to maximize the jobs’ PoCD by making full use of available resources. Notably, our work encompasses a new online scheduler that dynamically recomputes and reallocates resources during the course of a job’s execution. According to our findings, Shed+ successfully leverages cloud resources and maximizes the percentage of jobs meeting their deadlines. In our experiments, we have seen this percentage for heavy load going up to 98% for Shed+ as opposed to nearly 68%, 40%, 35% and 37% for Shed, Dolly, Hopper and Hadoop with speculation enabled, respectively.The last part of this dissertation presents Forseti, a novel processing scheme that is able to reshape data chunk size processed by heterogeneous machines in a dynamic fashion, and as a result, mitigates residual workload and stragglers to achieve significant improvement in performance. We note that Forseti does not require any a priori knowledge of the machine configuration nor job statistics. Instead, it infers such information on the fly and adjusts data chunk sizes at runtime, making the solution robust even in environments with high volatility. In its implementation, Forseti exploits “JVM reuse” feature to avoid task start-up and initialization cost associated with launching new tasks. We prototype Forseti on a real-world cluster and evaluate its performance using several realistic benchmarks. The results show that Forseti outperforms a number of baselines, including default Hadoop by up to 68% and SkewTune by up to 50%.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.