Research Statement -muntasir Raihan Rahman

My research goal is to build adaptive big data and cloud systems that can meet a spectrum of user requirements expressed using service level agreements (SLA) and service level objectives (SLO). During my PhD, I have worked on several angles of tracking and enforcing SLA/SLO guarantees in cloud systems, including in Mapreduce clusters, and NoSQL key-value storage systems. My future research goal is to incorporate adaptivity and SLA/SLOs as first class citizens for a wider spectrum of cloud systems including distributed graph processing and machine learning systems, and distributed transactional systems. Today cloud users have to fine-tune and manually manage every distributed cloud system they deploy, ranging from distributed storage systems to distributed computation systems. There are few ways of allowing users to specify their requirements and having the system adapt automatically to meet these, no matter what the workload and environmental behavior. My research aims at incorporating user requirements, specified as SLAs or SLOs into cloud systems. In a storage system like a NoSQL/key-value store, these SLAs/SLOs might specify conflicting latency and consistency requirements. In a computation system like Hadoop, they might entail job priorities and deadlines. In my work, I have designed adaptive query routers and schedulers that change the system behavior to meet these SLAs/SLOs at all times. My PhD research has evolved around incorporating adaptivity and SLA/SLOs into Big Data storage, computation and recovery systems. In my PCAP project, we replaced the traditional CAP theorem [4] binary choice between consistency (C) and availability (A) under partitions (P) with probabilistic tradeoffs which characterize the unachievable envelope among C, A, and P for a distributed storage system using probabilistic parameters [19]. We then use this result to design and implement adaptive variants of NoSQL databases that (1) perform close to the optimal envelope, and (2) adaptively satisfy probabilistic consistency and latency SLAs under arbitrary network variations. Our Natjam system [5] supports SLOs in terms of arbitrary job priorities and deadlines for resource-constrained Mapreduce computation clusters. To enable flexible tradeoff between the cost of data loss and checkpointing overhead, I proposed a practical scalable software-defined mechanism for taking crash-consistent snapshots of a group of virtual machines [18]. Finally in my current project I am incorporating adaptivity into graph computation systems to meet deadline and resource SLOs. Besides my main dissertation research I have also worked on: (1) multiple frameworks for formal modeling and verification of distributed systems [11, 12, 13, 14], (2) new distributed storage abstractions for social sensing [2],(3) resource allocation algorithms for network virtualization and software-defined networking [6, 7, 15, 16], and (4) metrics for quantifying consistency for eventually consistent storage systems [9, 10, 17]. Next I will elaborate more on my main PhD research projects and future research interests.