Mayur Deshpande

CREW: A Gossip-based Flash-Dissemination System

In this paper, we explore a new form of dissemination called Flash Dissemination that involves dissemination of fixed, rich information to a large number of recipients in as short a time as possible. Key characteristics of Flash Dissemination include unpredictability in its need, scalability to large number of recipients and autonomic performance in highly heterogenous and failureprone environments. Previous work either addresses large content delivery in heterogenous networks or fault-tolerant dissemination of (streaming) events. We investigate a peer-based approach using foundations from broadcast networks, gossip theory and random networks. In this paper, we propose CREW (Concurrent Random Expanding Walkers), a scalable, lightweight, and autonomic gossip-based protocol. CREW is also explicitly designed to maximize the speed of dissemination using adaptive and intelligent intra and inter node concurrency. We implemented CREW on top of a scalable middleware environment and compared it to optimized implementations of popular gossip and peer-based systems. Our experiments show that CREW outperforms both traditional gossip and current large content dissemination systems, across a wide range of comparative metrics, even though its design is counterintuitive from a systems perspective.

Design and implementation of a composable reflective middleware framework

With the evolution of the global information infrastructure, service providers will need to provide effective and adaptive resource management mechanisms that can serve more concurrent clients and deal with applications that exhibit quality-of-service (QoS) requirements. Flexible, scalable and customizable middleware can be used as an enabling technology for next-generation systems that adhere to the QoS requirements of applications that execute in highly dynamic distributed environments. To enable application-aware resource management, we are developing a customizable and composable middleware framework called CompOSE|Q (Composable Open Software Environment with QoS), based on a reflective meta-model. In this paper, we describe the architecture and runtime environment for CompOSE|Q and briefly assess the performance overhead of the additional flexibility. We also illustrate how flexible communication mechanisms can be supported efficiently in the CompOSE|Q framework.

Operating system performance in support of real-time middleware

Commercial-off-the-shelf (COTS) hardware and software is being evaluated and/or used in an increasing range of mission-critical distributed real-time and embedded (DRE) systems. Due to substantial R&D investment over the past decade, COTS middleware has matured to the point where it is no longer the dominant factor in the overhead, non-determinism, and priority inversion incurred by DRE systems. As a result, the focus has shifted to the COTS operating systems and networks, which are once again responsible for the majority of end-to-end latency and jitter. We compare and evaluate the suitability of popular COTS operating systems for real-time COTS middleware, such as Real-time CORBA. We examine real-time operating systems (VxWorks and QNX), general-purpose operating systems with real-time thread scheduling classes (Windows NT Windows 2K, and Linux), and a hybrid real-time/general-purpose operating system (Linux/RT). While holding the hardware and ORB constant, we vary these operating systems systematically to measure platform-specific variations in context switch overhead, throughput of the ORB in terms of two-way operations per-second and memory-footprint of the ORB libraries. We also measure how the latency and jitter of high-priority operations are affected as the number of low-priority operations increase.

The different dimensions of dynamicity

In a peer-to-peer (P2P) network, a fabric of overlay links helps peers discover and use other peers resources. This fabric, however, is highly dynamic and constantly changing. While different measures of dynamicity have been implicitly and explicitly proposed, there is lack of deeper understanding about the various aspects of dynamicity. We systematically evaluate and quantify different dimensions of dynamicity through controlled generation of different dynamic networks. We also introduce a new dimension of dynamicity, persistence, which quantifies stable nodes in a network. This measure could be quite useful in the design and testing of P2P protocols that exploit the presence of stable nodes. Also, quite coincidentally, and to our surprise, a certain type of dynamic network that we designed has node degree properties that resemble those observed in social networks.

Design and Performance of Asynchronous Method Handling for CORBA

This paper describes the design and performance of a new mechanism, called asynchronous method handling (AMH), that allows CORBA servers to process client requests asynchronously. AMH decouples the association of an incoming request from the run-time stack that received the request, without incurring the context-switching, synchronization, and data movement overhead of conventional CORBA multithreading models.This paper provides two contributions to the study of asynchrony for CORBA servers. First, it describes the design and implementation of AMH in The ACE ORB (TAO), a C++ CORBA ORB. The syntax and semantics of AMH are defined using the CORBA Interface Definition Language (IDL), the forces that guided the design of AMH are described, and the patterns and C++ idioms used to resolve these forces are presented. Second, we empirically compare a middle-tier server implemented using AMH against other CORBA server concurrency models and show the advantages of the AMH mechanism against the other models.

Towards Reliable Application Data Broadcast in Wireless Ad Hoc Networks

Application data broadcast in ad hoc networks is an important primitive that has received little systematic research - the main focus of prior research being on control data broadcast. In this paper, we show why control data broadcast and even multicast techniques are insufficient for reliable application data broadcast; in fact their reliability degrades sharply with increasing application data size. We discover the root cause of this to be IP fragmenting the application data but not providing good reliability control on the fragments. We hence propose READ (reliable and efficient application-data dissemination), a protocol based on higher-layer fragmentation with fragment-level reliability control. READ splits a data packet into fragments, and disseminates them separately at dynamically adaptive intervals. Receivers piggyback implicit NACKs when propagating the fragments, and retrieve missing fragments from neighbors. Through experiments, we show that READ consistently achieves high delivery ratio and short latency, outperforming all other examined protocols.

Flashback: A Peer-to-PeerWeb Server for Flash Crowds

We present Flashback, a ready-to-use system for scalably handling large unexpected traffic spikes on web-sites. Unlike previous systems, our approach does not rely on any intermediate nodes to cache content. Instead, the clients (browsers) create a dynamic, self-scaling peer-to-peer (P2P) Web-server that grows and shrinks according to the load. This approach translates into a challenging problem - a P2P data exchange protocol that can operate in churn rates where more than 90% of peers can leave the overlay in under 10 seconds. This is at least an order of magnitude higher churn rate than previously addressed research. Additionally, our system operates under two strict constraints - users are assured that they upload only as much as they download and second, end-user browsing experience is preserved, i.e., low latency downloads and zero configuration or download of any software. Various innovations were required to meet these challenges. Key among them are (a) A TCP-friendly, UDP protocol (Roulette) for tit-for-tat data exchange under extreme churn, (b) A novel data structure (NOIS) for partial-data management and (c) A distributed hole-punching protocol for automatic NAT traversal. Experimental results show the effectiveness and near optimal scaling of Flashback. For a Web-server (and clients) running on a DSL-like connection, end-user latency increases only one second for every doubling in Web-server load.

Gateway designation for timely communications in instant mesh networks

In this paper, we explore how to effectively create and use “instant mesh networks”, i.e., wireless mesh networks that are dynamically deployed in temporary circumstances (e.g., emergency responses) - in addition to enabling coverage for internal on-site communications, such a network will support information flow into and out of the deployment site through its gateway (i.e., the mesh router that connects to the external backhaul). We study optimizing the performance of communications (specifically in terms of latency) in an instant mesh network by intelligently selecting the gateway. We demonstrate that designating the proper gateway significantly enhances the timeliness of communications with the external backhaul. We mathematically model the “gateway designation problem” using the notion of centrality from graph theory. We propose a distributed algorithm, FACE (Fast Approximate Center Exploration), for locating the optimal gateway. FACE is an approximate algorithm that works in an efficient manner without compromising the optimality of solutions. A thorough performance evaluation shows that the gateways designated by FACE reduce latencies by up to 92% for various types of communications, and that FACE saves transmission cost and execution time by up to 71% in finding the gateways.