MPLEX: In-Situ Big Data Processing with Compute-Storage Multiplexing

摘要

Cloud-based services are increasingly popular for big data analytics due to the flexibility, scalability, and cost-effectiveness of provisioning elastic resources on-demand. However, data analytics-as-a-service suffers from the overheads of data movement between compute and storage clusters, due to their decoupled architecture in existing cloud infrastructure. In this work, we propose a novel approach of in-situ big data processing on cloud storage by dynamically offloading data-intensive jobs from compute cluster to storage cluster, and improve job throughput. However, it is challenging to achieve this goal since introducing additional workload on the storage cluster can significantly impact interactive web requests that fetch cloud storage data, with strict SLA (service-level agreement) for tail latency. In this work, we present MPLEX, a system that augments data analytics-as-a-service by efficiently multiplexing compute and storage cluster to improve job throughput without violating the SLA of cloud storage service in terms of tail response time. It applies an SLA-aware opportunistic job scheduling technique supported by a machine learning based prediction model to exploit the dynamic workload conditions in the compute, and storage cluster. Performance evaluations on an OpenStack Swift cluster, and an OpenStack based virtual cluster of Hadoop VMs built atop NSFCloud's Chameleon testbed show that MPLEX improves the Hadoop job throughput by up to 1.7X, while maintaining the SLA for cloud storage service requests.