Task Assignment in a Distributed System: Improving Performance by Unbalancing Load

We consider the problem of task assignment in a distributed system (such as a distributed Web server) in which task sizes are drawn from a heavy-tailed distribution. Many task assignment algorithms are based on the heuristic that balancing the load at the server hosts will result in optimal performance. We show this conventional wisdom is less true when the task size distribution is heavy-tailed (as is the case for Web file sizes). We introduce a new task assignment policy, called Size Interval Task Assignment with Variable Load (SITA-V). SITA-V purposely operates the server hosts at different loads, and directs smaller tasks to the lighter-loaded hosts. The result is that SITA-V provably decreases the mean task slowdown by significant factors (up to 1000 or more) where the more heavy-tailed the workload, the greater the improvement factor. We evaluate the tradeoff between improvement in slowdown and increase in waiting time in a system using SITA-V, and show conditions under which SITA-V represents a particularly appealing policy. We conclude with a discussion of the use of SITA-V in a distributed Web server, and show that it is attractive because it has a simple implementation which requires no communication from the server hosts back to the task router.

Pricing Algorithms For a Two-sided Internet Advertisement Market

The Google AdSense Program is a successful internet advertisement program where Google places contextual adverts on third-party websites and shares the resulting revenue with each publisher. Advertisers have budgets and bid on ad slots while publishers set reserve prices for the ad slots on their websites. Following previous modelling efforts, we model the program as a two-sided market with advertisers on one side and publishers on the other. We show a reduction from the Generalised Assignment Problem (GAP) to the problem of computing the revenue maximising allocation and pricing of publisher slots under a first-price auction. GAP is APX-hard but a (1-1/e) approximation is known. We compute truthful and revenue-maximizing prices and allocation of ad slots to advertisers under a second-price auction. The auctioneer's revenue is within (1-1/e) second-price optimal.

Typability and Type Checking in the Second-Order Λ-Calculus Are Equivalent and Undecidable (Preliminary Draft)

We consider the problems of typability[1] and type checking[2] in the Girard/Reynolds second-order polymorphic typed λ-calculus, for which we use the short name "System F" and which we use in the "Curry style" where types are assigned to pure λ -terms. These problems have been considered and proven to be decidable or undecidable for various restrictions and extensions of System F and other related systems, and lower-bound complexity results for System F have been achieved, but they have remained "embarrassing open problems"[3] for System F itself. We first prove that type checking in System F is undecidable by a reduction from semi-unification. We then prove typability in System F is undecidable by a reduction from type checking. Since the reverse reduction is already known, this implies the two problems are equivalent. The second reduction uses a novel method of constructing λ-terms such that in all type derivations, specific bound variables must always be assigned a specific type. Using this technique, we can require that specific subterms must be typable using a specific, fixed type assignment in order for the entire term to be typable at all. Any desired type assignment may be simulated. We develop this method, which we call "constants for free", for both the λK and λI calculi.

Dynamic Server Selection using Bandwidth Probing in Wide-Area Networks

Replication is a commonly proposed solution to problems of scale associated with distributed services. However, when a service is replicated, each client must be assigned a server. Prior work has generally assumed that assignment to be static. In contrast, we propose dynamic server selection, and show that it enables application-level congestion avoidance. To make dynamic server selection practical, we demonstrate the use of three tools. In addition to direct measurements of round-trip latency, we introduce and validate two new tools: bprobe, which estimates the maximum possible bandwidth along a given path; and cprobe, which estimates the current congestion along a path. Using these tools we demonstrate dynamic server selection and compare it to previous static approaches. We show that dynamic server selection consistently outperforms static policies by as much as 50%. Furthermore, we demonstrate the importance of each of our tools in performing dynamic server selection.

To Queue or Not to Queue: When Queueing is Better Than Timesharing in a Distributed System

We examine the question of whether to employ the first-come-first-served (FCFS) discipline or the processor-sharing (PS) discipline at the hosts in a distributed server system. We are interested in the case in which service times are drawn from a heavy-tailed distribution, and so have very high variability. Traditional wisdom when task sizes are highly variable would prefer the PS discipline, because it allows small tasks to avoid being delayed behind large tasks in a queue. However, we show that system performance can actually be significantly better under FCFS queueing, if each task is assigned to a host based on the task's size. By task assignment, we mean an algorithm that inspects incoming tasks and assigns them to hosts for service. The particular task assignment policy we propose is called SITA-E: Size Interval Task Assignment with Equal Load. Surprisingly, under SITA-E, FCFS queueing typically outperforms the PS discipline by a factor of about two, as measured by mean waiting time and mean slowdown (waiting time of task divided by its service time). We compare the FCFS/SITA-E policy to the processor-sharing case analytically; in addition we compare it to a number of other policies in simulation. We show that the benefits of SITA-E are present even in small-scale distributed systems (four or more hosts). Furthermore, SITA-E is a static policy that does not incorporate feedback knowledge of the state of the hosts, which allows for a simple and scalable implementation.

The Sensor Network Workbench: Towards Functional Specification, Verification and Deployment of Constrained Distributed Systems

As the commoditization of sensing, actuation and communication hardware increases, so does the potential for dynamically tasked sense and respond networked systems (i.e., Sensor Networks or SNs) to replace existing disjoint and inflexible special-purpose deployments (closed-circuit security video, anti-theft sensors, etc.). While various solutions have emerged to many individual SN-centric challenges (e.g., power management, communication protocols, role assignment), perhaps the largest remaining obstacle to widespread SN deployment is that those who wish to deploy, utilize, and maintain a programmable Sensor Network lack the programming and systems expertise to do so. The contributions of this thesis centers on the design, development and deployment of the SN Workbench (snBench). snBench embodies an accessible, modular programming platform coupled with a flexible and extensible run-time system that, together, support the entire life-cycle of distributed sensory services. As it is impossible to find a one-size-fits-all programming interface, this work advocates the use of tiered layers of abstraction that enable a variety of high-level, domain specific languages to be compiled to a common (thin-waist) tasking language; this common tasking language is statically verified and can be subsequently re-translated, if needed, for execution on a wide variety of hardware platforms. snBench provides: (1) a common sensory tasking language (Instruction Set Architecture) powerful enough to express complex SN services, yet simple enough to be executed by highly constrained resources with soft, real-time constraints, (2) a prototype high-level language (and corresponding compiler) to illustrate the utility of the common tasking language and the tiered programming approach in this domain, (3) an execution environment and a run-time support infrastructure that abstract a collection of heterogeneous resources into a single virtual Sensor Network, tasked via this common tasking language, and (4) novel formal methods (i.e., static analysis techniques) that verify safety properties and infer implicit resource constraints to facilitate resource allocation for new services. This thesis presents these components in detail, as well as two specific case-studies: the use of snBench to integrate physical and wireless network security, and the use of snBench as the foundation for semester-long student projects in a graduate-level Software Engineering course.

Collocation Games and Their Application to Distributed Resource Management

We introduce Collocation Games as the basis of a general framework for modeling, analyzing, and facilitating the interactions between the various stakeholders in distributed systems in general, and in cloud computing environments in particular. Cloud computing enables fixed-capacity (processing, communication, and storage) resources to be offered by infrastructure providers as commodities for sale at a fixed cost in an open marketplace to independent, rational parties (players) interested in setting up their own applications over the Internet. Virtualization technologies enable the partitioning of such fixed-capacity resources so as to allow each player to dynamically acquire appropriate fractions of the resources for unencumbered use. In such a paradigm, the resource management problem reduces to that of partitioning the entire set of applications (players) into subsets, each of which is assigned to fixed-capacity cloud resources. If the infrastructure and the various applications are under a single administrative domain, this partitioning reduces to an optimization problem whose objective is to minimize the overall deployment cost. In a marketplace, in which the infrastructure provider is interested in maximizing its own profit, and in which each player is interested in minimizing its own cost, it should be evident that a global optimization is precisely the wrong framework. Rather, in this paper we use a game-theoretic framework in which the assignment of players to fixed-capacity resources is the outcome of a strategic "Collocation Game". Although we show that determining the existence of an equilibrium for collocation games in general is NP-hard, we present a number of simplified, practically-motivated variants of the collocation game for which we establish convergence to a Nash Equilibrium, and for which we derive convergence and price of anarchy bounds. In addition to these analytical results, we present an experimental evaluation of implementations of some of these variants for cloud infrastructures consisting of a collection of multidimensional resources of homogeneous or heterogeneous capacities. Experimental results using trace-driven simulations and synthetically generated datasets corroborate our analytical results and also illustrate how collocation games offer a feasible distributed resource management alternative for autonomic/self-organizing systems, in which the adoption of a global optimization approach (centralized or distributed) would be neither practical nor justifiable.

Tracking a Large Number of Objects from Multiple Views

We propose a multi-object multi-camera framework for tracking large numbers of tightly-spaced objects that rapidly move in three dimensions. We formulate the problem of finding correspondences across multiple views as a multidimensional assignment problem and use a greedy randomized adaptive search procedure to solve this NP-hard problem efficiently. To account for occlusions, we relax the one-to-one constraint that one measurement corresponds to one object and iteratively solve the relaxed assignment problem. After correspondences are established, object trajectories are estimated by stereoscopic reconstruction using an epipolar-neighborhood search. We embedded our method into a tracker-to-tracker multi-view fusion system that not only obtains the three-dimensional trajectories of closely-moving objects but also accurately settles track uncertainties that could not be resolved from single views due to occlusion. We conducted experiments to validate our greedy assignment procedure and our technique to recover from occlusions. We successfully track hundreds of flying bats and provide an analysis of their group behavior based on 150 reconstructed 3D trajectories.

Design and Implementation of Statistical Rate Monotonic Scheduling in KURT Linux

Statistical Rate Monotonic Scheduling (SRMS) is a generalization of the classical RMS results of Liu and Layland [LL73] for periodic tasks with highly variable execution times and statistical QoS requirements. The main tenet of SRMS is that the variability in task resource requirements could be smoothed through aggregation to yield guaranteed QoS. This aggregation is done over time for a given task and across multiple tasks for a given period of time. Similar to RMS, SRMS has two components: a feasibility test and a scheduling algorithm. SRMS feasibility test ensures that it is possible for a given periodic task set to share a given resource without violating any of the statistical QoS constraints imposed on each task in the set. The SRMS scheduling algorithm consists of two parts: a job admission controller and a scheduler. The SRMS scheduler is a simple, preemptive, fixed-priority scheduler. The SRMS job admission controller manages the QoS delivered to the various tasks through admit/reject and priority assignment decisions. In particular, it ensures the important property of task isolation, whereby tasks do not infringe on each other. In this paper we present the design and implementation of SRMS within the KURT Linux Operating System [HSPN98, SPH 98, Sri98]. KURT Linux supports conventional tasks as well as real-time tasks. It provides a mechanism for transitioning from normal Linux scheduling to a mixed scheduling of conventional and real-time tasks, and to a focused mode where only real-time tasks are scheduled. We overview the technical issues that we had to overcome in order to integrate SRMS into KURT Linux and present the API we have developed for scheduling periodic real-time tasks using SRMS.

Fusion Artmap: A Neural Network Architecture for Multi-Channel Data Fusion and Classification

Fusion ARTMAP is a self-organizing neural network architecture for multi-channel, or multi-sensor, data fusion. Single-channel Fusion ARTMAP is functionally equivalent to Fuzzy ART during unsupervised learning and to Fuzzy ARTMAP during supervised learning. The network has a symmetric organization such that each channel can be dynamically configured to serve as either a data input or a teaching input to the system. An ART module forms a compressed recognition code within each channel. These codes, in turn, become inputs to a single ART system that organizes the global recognition code. When a predictive error occurs, a process called paraellel match tracking simultaneously raises vigilances in multiple ART modules until reset is triggered in one of them. Parallel match tracking hereby resets only that portion of the recognition code with the poorest match, or minimum predictive confidence. This internally controlled selective reset process is a type of credit assignment that creates a parsimoniously connected learned network. Fusion ARTMAP's multi-channel coding is illustrated by simulations of the Quadruped Mammal database.

Fusion ARTMAP: An Adaptive Fuzzy Network for Multi-Channel Classification

Fusion ARTMAP is a self-organizing neural network architecture for multi-channel, or multi-sensor, data fusion. Fusion ARTMAP generalizes the fuzzy ARTMAP architecture in order to adaptively classify multi-channel data. The network has a symmetric organization such that each channel can be dynamically configured to serve as either a data input or a teaching input to the system. An ART module forms a compressed recognition code within each channel. These codes, in turn, beco1ne inputs to a single ART system that organizes the global recognition code. When a predictive error occurs, a process called parallel match tracking simultaneously raises vigilances in multiple ART modules until reset is triggered in one of thmn. Parallel match tracking hereby resets only that portion of the recognition code with the poorest match, or minimum predictive confidence. This internally controlled selective reset process is a type of credit assignment that creates a parsimoniously connected learned network.