AT&T Labs Research - Joshi, Kaustubh

Today, multitier distributed systems power many online computing services that society takes for granted. These include everything from online shopping sites to communications services to online content providers. Often co-located in large data centers, these complex services can be very difficult to keep running smoothly. I am interested in the online management of such systems to improve dependability, security, performance, and power efficiency. My research straddles both theory and practice, ranging from analytical techniques to model and reason about system behaviors under conditions such as faults, attacks, and changing workloads, to runtime control systems that use the models to implement adaptive behaviors such as fault recovery and self-reconfiguration. I am especially interested in techniques that use data analysis or runtime measurements to construct models with very little human intervention, and for black-box systems for which very little design information is available. My work can be loosely categorized in the areas of dependable, adaptive, and autonomic systems, and is of special relevance to shared infrastructure approaches such as cloud computing.

Before coming to AT&T, I received my Ph.D. in Computer Science from the University of Illinois at Urbana-Champaign. My dissertation work investigated the use of Partially Observable Markov Decision Process (POMDP) models to perform online fault diagnosis and recovery in multitier systems. This research combined incomplete, sometimes inaccurate, and often conflicting reports from multiple monitoring systems to to make cost-benefit tradeoffs and choose recovery mechanisms that would bring a system back up with minimal disruption after a failure had occurred.

(c) ACM, 2011. This is the author's version of the work. It is posted here by permission of ACM for your personal use. Not for redistribution. The definitive version was published in ACM MobiHeld workshop 2011 , 2011-10-23.

{Mobile application developers pay little attention to the interactions between applications and the cellular net- work carrying their traffic. This results in wastage of de- vice energy and network signaling resources. We place part of the blame on mobile OSes: they do not expose adequate interfaces through which applications can in- teract with the network. We propose traffic backfilling, a technique in which delay-tolerant traffic is opportunis- tically transmitted by the OS using resources left over by the naturally occurring bursts caused by interactive traffic. Backfilling presents a simple interface with two classes of traffic, and grants the OS and network large flexibility to maximize the use of network resources and reduce device energy consumption. Using device traces and network data from a major US carrier, we demon- strate a large opportunity for traffic backfilling.}

(c) ACM, 2011. This is the author's version of the work. It is posted here by permission of ACM for your personal use. Not for redistribution. The definitive version was published in ACM SOSP Workshop on "Managing Large-Scale Systems via the Analysis of System Logs and the Appli, 2011-10-23.

{Chronics are recurrent problems that fly under the radar of operations teams because they do not perturb the system enough to set off alarms or violate service-level objectives. The discovery and diagnosis of never-before seen chronics poses new challenges as they are not detected by traditional threshold-based techniques, and many chronics can be present in a system at once, all starting and ending at different times. In this paper, we describe our experiences diagnosing chron- ics using server logs on a large telecommunications service. Our technique uses a scalable Bayesian distribution learner coupled with an information theoretic measure of distance (KL divergence), to identify the attributes that best distin- guish failed calls from successful calls. Our preliminary re- sults demonstrate the usefulness of our technique by provid- ing examples of actual instances where we helped operators discover and diagnose chronics.}

(c) ACM, 2011. This is the author's version of the work. It is posted here by permission of ACM for your personal use. Not for redistribution. The definitive version was published in ACM European Conference on Computer Systems - Eurosys 2011, 2011-04-10.

{We introduce cloud micro-elasticity, a new model for cloud Virtual Machine (VM) allocation and management. Currentcloud users over-provision long-lived VMs with large memory footprints, to better absorb load spikes and to conserve performance-sensitive caches. Instead, we achieve elasticity by swiftly cloning VMs into many transient, short-lived, fractional worker clones, to multiplex physical resources at a much finer granularity. The memory of micro-elastic clones is a logical replica of all parent VM state including caches, and its footprint is a fraction of the nominal maximum proportional to the workload. We enable micro-elasticity through a novel technique dubbed VM state coloring, which classifies VM memory into sets of semantically-related regions, and optimizes the propagation, allocation and deduplication of these regions. Using coloring we build Kaleidoscope and empirically demonstrate its ability to create micro-elastic cloned servers.We model the impact of microelasticity on a demand dataset from a hosting provider, and show that fine-grained multiplexing yields infrastructure reductions of 30% relative to state-of-the art techniques for managing elastic clouds.}

The definitive version was published in Proceedings of WOSN 2010, Usenix. , 2010-06-22

{Office and home environments are increasingly crowded with personal computers. Even though these computers see little use in the course of the day, they often remain powered, even when idle. Leaving idle PCs running is not only wasteful, but with rising energy costs it is in- creasingly more expensive. We propose partial migration of idle desktop sessions into the cloud to achieve energy- proportional computing. Partial migration only propa- gates the small footprint of state that will be needed dur- ing idle period execution, and returns the session to the PC when it is no longer idle. We show that this approach can reduce energy usage of an idle desktop by up to 50% over an hour and by up to 69% overnight. We also show that idle desktop sessions have small working sets, up to an order of magnitude smaller than their allocated mem- ory, enabling significant consolidation ratios.}

{Automatic system monitoring and recovery has the potential to provide effective, low-cost ways to improve dependability in distributed software systems. However, automating recovery is challenging in practice because accurate fault diagnosis is difficult given the common monitoring tools and techniques with low fault coverage, poor fault localization, detection delays, and false positives. In this paper, we present a holistic model-based approach that overcomes these challenges and enables automatic recovery in distributed systems. To do so, it uses theoretically sound techniques including Bayesian estimation and Markov decision theory to provide controllers that choose good, if not optimal, recovery actions according to a user-defined optimization criteria. By combining monitoring and recovery, the approach realizes benefits that could not have been obtained by using them in isolation. We experimentally validate our framework by fault injection on realistic e-commerce systems.}

Dynamic voltage and frequency scaling (DVFS) and virtual machine (VM) based server consolidation are techniques that hold promise for energy conservation, but can also have adverse impacts on system performance. For the responsiveness-sensitive multitier applications running in today’s data centers, queuing models should ideally be used to predict the impact of CPU scaling on response time, to allow appropriate runtime trade-offs between performance and energy use. In practice, however, such models are difﬁcult to construct and thus are often abandoned for ad-hoc solutions. In this paper, an alternative measurement-based approach that predicts the impacts without requiring detailed application knowledge is presented. The approach uses a new set of metrics, the CPU gradients, that can be automatically measured on a running system using lightweight and nonintrusive CPU perturbations. The practical feasibility of the approach is demonstrated using extensive experiments on multiple multitier applications, and it is shown that simple energy controllers can use gradient predictions to derive as much as 50% energy savings while still meeting response time constraints.

AT&T Labs Research — Leading Invention, Driving Innovation

site map