Deepak Rajan

PREPARE: Predictive Performance Anomaly Prevention for Virtualized Cloud Systems

Virtualized cloud systems are prone to performance anomalies due to various reasons such as resource contentions, software bugs, and hardware failures. In this paper, we present a novel Predictive Performance Anomaly Prevention (PREPARE) system that provides automatic performance anomaly prevention for virtualized cloud computing infrastructures. PREPARE integrates online anomaly prediction, learning-based cause inference, and predictive prevention actuation to minimize the performance anomaly penalty without human intervention. We have implemented PREPARE on top of the Xen platform and tested it on the NCSUs Virtual Computing Lab using a commercial data stream processing system (IBM System S) and an online auction benchmark (RUBiS). The experimental results show that PREPARE can effectively prevent performance anomalies while imposing low overhead to the cloud infrastructure.

On splittable and unsplittable flow capacitated network design arc–set polyhedra

Abstract.We study the polyhedra of splittable and unsplittable single arc–set relaxations of multicommodity flow capacitated network design problems. We investigate the optimization problems over these sets and the separation and lifting problems of valid inequalities for them. In particular, we give a linear–time separation algorithm for the residual capacity inequalities [19] and show that the separation problem of c–strong inequalities [7] is ??–hard, but can be solved over the subspace of fractional variables only. We introduce two classes of inequalities for the unsplittable flow problems. We present a summary of computational experiments with a branch-and-cut algorithm for multicommodity flow capacitated network design problems to test the effectiveness of the results presented here empirically.

SODA: an optimizing scheduler for large-scale stream-based distributed computer systems

This paper describes the SODA scheduler for System S, a highly scalable distributed stream processing system. Unlike traditional batch applications, streaming applications are open-ended. The system cannot typically delay the processing of the data. The scheduler must be able to shift resource allocation dynamically in response to changes to resource availability, job arrivals and departures, incoming data rates and so on. The design assumptions of System S, in particular, pose additional scheduling challenges. SODA must deal with a highly complex optimization problem, which must be solved in real-time while maintaining scalability. SODA relies on a careful problem decomposition, and intelligent use of both heuristic and exact algorithms. We describe the design and functionality of SODA, outline the mathematical components, and describe experiments to show the performance of the scheduler.

FLEX: a slot allocation scheduling optimizer for MapReduce workloads

Originally, MapReduce implementations such as Hadoop employed First In First Out (fifo) scheduling, but such simple schemes cause job starvation. The Hadoop Fair Scheduler (hfs) is a slot-based MapReduce scheme designed to ensure a degree of fairness among the jobs, by guaranteeing each job at least some minimum number of allocated slots. Our prime contribution in this paper is a different, flexible scheduling allocation scheme, known as flex. Our goal is to optimize any of a variety of standard scheduling theory metrics (response time, stretch, makespan and Service Level Agreements (slas), among others) while ensuring the same minimum job slot guarantees as in hfs, and maximum job slot guarantees as well. The flex allocation scheduler can be regarded as an add-on module that works synergistically with hfs. We describe the mathematical basis for flex, and compare it with fifo and hfs in a variety of experiments.

A directed cycle-based column-and-cut generation method for capacitated survivable network design

A network is said to be survivable if it has sufficient capacity for rerouting all of its flow under the failure of any one of its edges. Here, we present a polyhedral approach for designing survivable networks. We describe a mixed-integer programming model, in which sufficient slack is explicitly introduced on the directed cycles of the network while flow routing decisions are made. In case of a failure, flow is rerouted along the slacks reserved on directed cycles. We give strong valid inequalities that use the survivability requirements. We present a computational study with a column-and-cut generation algorithm for designing capacitated survivable networks.

COLA: optimizing stream processing applications via graph partitioning

In this paper, we describe an optimization scheme for fusing compile-time operators into reasonably-sized run-time software units called processing elements (PEs). Such PEs are the basic deployable units in System S, a highly scalable distributed stream processing middleware system. Finding a high quality fusion significantly benefits the performance of streaming jobs. In order to maximize throughput, our solution approach attempts to minimize the processing cost associated with inter-PE stream traffic while simultaneously balancing load across the processing hosts. Our algorithm computes a hierarchical partitioning of the operator graph based on a minimum-ratio cut subroutine. We also incorporate several fusion constraints in order to support real-world System S jobs. We experimentally compare our algorithm with several other reasonable alternative schemes, highlighting the effectiveness of our approach.

A polyhedral study of production ramping

We give strong formulations of ramping constraints—used to model the maximum change in production level for a generator or machine from one time period to the next—and production limits. For the two-period case, we give a complete description of the convex hull of the feasible solutions. The two-period inequalities can be readily used to strengthen ramping formulations without the need for separation. For the general case, we define exponential classes of multi-period variable upper bound and multi-period ramping inequalities, and give conditions under which these inequalities define facets of ramping polyhedra. Finally, we present exact polynomial separation algorithms for the inequalities and report computational experiments on using them in a branch-and-cut algorithm to solve unit commitment problems in power generation.

Survivable Network Design: Routing of Flows and Slacks

We present a new mixed—integer programming model and a column generation method for the survivable design of telecommunication networks. In contrast to other failure scenario models, the new model has almost the same number of constraints as the regular network design problem, which makes it effective for large instances. Even though the complexity of pricing the exponentially many variables of the model is NP—hard, in our computational experiments, we are able to produce capacity—efficient survivable networks with dense graphs up to 70 nodes.

Temperature-Aware Scheduling: When is System-Throttling Good Enough?

In computing centers, power-aware operating systems ensure that processor temperatures do not exceed a threshold by utilizing system-throttling. In this technique, the system load (or alternatively, the clock speed) is scaled when the temperature hits this threshold. At other times, the system operates at maximum load. In this paper, we show that such simple system-throttling rules are in fact the best one can achieve under certain assumptions. We show that maintaining a constant operating speed (and thus temperature) always does more work than operating in alternating periods of cooling and heating. As a result, for certain settings and for a reasonable temperature model, we prove that system-throttling is the most effective temperature-aware scheduling. Naturally, these assumptions do not always hold; we also discuss the scenario when some of our assumptions are relaxed, and argue why one needs more complex scheduling algorithms in this case.

On the optimization of schedules for MapReduce workloads in the presence of shared scans

We consider MapReduce clusters designed to support multiple concurrent jobs, concentrating on environments in which the number of distinct datasets is modest relative to the number of jobs. In such scenarios, many individual datasets are likely to be scanned concurrently by multiple Map phase jobs. As has been noticed previously, this scenario provides an opportunity for Map phase jobs to cooperate, sharing the scans of these datasets, and thus reducing the costs of such scans. Our paper has three main contributions over previous work. First, we present a novel and highly general method for sharing scans and thus amortizing their costs. This concept, which we call cyclic piggybacking, has a number of advantages over the more traditional batching scheme described in the literature. Second, we notice that the various subjobs generated in this manner can be assumed in an optimal schedule to respect a natural chain precedence ordering. Third, we describe a significant but natural generalization of the recently introduced FLEX scheduler for optimizing schedules within the context of this cyclic piggybacking paradigm, which can be tailored to a variety of cost metrics. Such cost metrics include average response time, average stretch, and any minimax-type metric—a total of 11 separate and standard metrics in all. Moreover, most of this carries over in the more general case of overlapping rather than identical datasets as well, employing what we will call semi-shared scans. In such scenarios, chain precedence is replaced by arbitrary precedence, but we can still handle 8 of the original 11 metrics. The overall approach, including both cyclic piggybacking and the FLEX scheduling generalization, is called CIRCUMFLEX. We describe some practical implementation strategies. And we evaluate the performance of CIRCUMFLEX via a variety of simulation and real benchmark experiments.