Search

Training jobs of deep neural networks (DNNs) can be accelerated through distributed variants of stochastic gradient descent (SGD), where multiple nodes process training examples and exchange updates. The total throughput of the nodes depends not only on their computing power, but also on their networking speeds and coordination mechanism (synchronous or asynchronous, centralized or decentralized), since communication bottlenecks and stragglers can result in sublinear scaling when additional nodes are provisioned. In this paper, we propose two classes of performance models to predict throughput of distributed SGD: fine-grained models, representing many elementary computation/communication operations and their dependencies; and coarse-grained models, where SGD steps at each node are represented as a sequence of high-level phases without parallelism between computation and communication. Using a PyTorch implementation, real-world DNN models and different cloud environments, our experimental evaluation illustrates that, while fine-grained models are more accurate and can be easily adapted to new variants of distributed SGD, coarse-grained models can provide similarly accurate predictions when augmented with ad hoc heuristics, and their parameters can be estimated with profiling information that is easier to collect. [ABSTRACT FROM AUTHOR]

Consistent (pause-free) quality of service is required in peer-to-peer (P2P) video streaming systems. In this paper, we aim to eliminate the problem of playback pauses in such systems via the use of positive incentives for peers to contribute high upload rates. We model our problem as a market, where the market stakeholders consist of multiple content providers, advertisement providers, and network peers; the positive incentives for peers in the market are reduced advertisement (ad) viewing durations. From a system design perspective, one of our primary goals is to compute the market equilibria that include appropriate ad viewing durations, offering sufficient incentives for network peers to continue contributing. Our simulation-based studies demonstrate that we mitigate the “playback pause” problem for peers by up to 80% as compared to existing approaches, generate sufficient utility for advertisers to be part of the market, and enable content providers to achieve their desired utility by providing sufficient incentives for all peers to stay in the system without violating ad provider agreements. [ABSTRACT FROM PUBLISHER]

This paper considers a stochastic optimization approach for job scheduling and server management in large-scale, geographically distributed data centers. Randomly arriving jobs are routed to a choice of servers. The number of active servers depends on server activation decisions that are updated at a slow time scale, and the service rates of the servers are controlled by power scaling decisions that are made at a faster time scale. We develop a two-time-scale decision strategy that offers provable power cost and delay guarantees. The performance and robustness of the approach is illustrated through simulations. [ABSTRACT FROM AUTHOR]

Increasingly, peer-to-peer (P2P) network users expect to be able to search objects by semantic attributes based on their preferences for multimedia content. Partial match search (i.e., search through the use of multimedia content semantic information) has become an essential service in P2P systems. In this paper, we propose SocioNet, a social-based overlay that clusters peers based on their preference relationships as a small-world network. In SocioNet, peers mimic how people form a social network and how they query, by preference, their friends or acquaintances. Hence, SocioNet benefits from two desirable features of a social network: interest-based clustering and small-world properties (i.e., high cluster coefficient among all peers yet short path lengths between any two peers). To realize an interest-based small-world SocioNet, we also investigate the following practical design issues: 1) similarity estimation: we define a quantifiable similarity measure that enables clustering of similar peers in SocioNet; 2) distributed small-world overlay adaptation: peers maintain a small-world overlay under network dynamics; and 3) query strategy under the small-world overlay: we analyze appropriate settings for the Time-to-Live (TTL) value, for TTL-limited flooding, that provides a satisfactory success ratio and avoids redundant message overhead. We use simulations and a real database called AudioScrobbler [1], which tracks users' listening habits, to evaluate the performance of SocioNet. The results show that SocioNet assists peers in locating content at peers with similar interests through short path lengths, and hence, achieves a higher success ratio (than nonsmall-world interest-based overlays and noninterest-based small-world overlays) while reducing message overhead significantly. [ABSTRACT FROM AUTHOR]

In this paper, we consider performance evaluation of a system which shares K servers (or resources) among N heterogeneous classes of workloads, where server allocation and deallocation for class i is dictated by a class specific threshold- based policy with hysteresis control. In particular, the server activation time for class i is noninstantaneous. There are many systems and applications whare a multiclass threshold-based queueing system can be of great use. One important utility of using threshold-based approaches is In situations where applications may incur server usage costs. In these cases, one needs to consider not only the performance aspects but also the resulting cost/performance ratio. The motivation for using hysteresis control is to reduce the unnecessary cost of server setup (or activation) and server removal (or deactivation) whenever there are momentary fluctuations in workload. Moreover, servers in such systems and applications are often needed by multiple classes of workloads and, hence, it is desirable to find good approaches to sharing server resources among the different classes of workloads, preferably without statically partitioning the server pool among these classes. An important and distinguishing characteristic of our work is that we consider the modeling and analysis of a multiclass system with noninstantaneous server activation. The main contributions of this work are 1) in developing an efficient approximation method for solving such models, 2) in verifying the convergence of our iterative method, and 3) in evaluating the resulting accuracy of the technique for computing performance measures of interest, which can subsequently be used in making system design choices. [ABSTRACT FROM AUTHOR]

We consider the problem of collecting a large amount of data from several different hosts to a single destination in a wide-area network. This problem is important since improvements in data collection times in many applications such as wide-area upload applications, high-performance computing applications, and data mining applications are crucial to performance of those applications. Often, due to congestion conditions, the paths chosen by the network may have poor throughput. By choosing an alternate route at the application level, we may be able to obtain substantially faster completion time. This data collection problem is a nontrivial one because the issue is not only to avoid congested link(s), but to devise a coordinated transfer schedule which would afford maximum possible utilization of available network resources. Our approach for computing coordinated data collection schedules makes no assumptions about knowledge of the topology of the network or the capacity available on individual links of the network. This approach provides significant performance improvements under various degrees and types of network congestions. To show this, we give a comprehensive comparison study of the various approaches to the data collection problem which considers performance, robustness, and adaptation characteristics of the different data collection methods. The adaptation to network conditions characteristics are important as the above applications are long lasting, i.e., it is likely changes in network conditions will occur during the data transfer process. In general, our approach can be used for solving arbitrary data movement problems over the Internet. We use the Bistro platform to illustrate one application of our techniques. [ABSTRACT FROM AUTHOR]

Quality of service in streaming of continuous media over the Internet is poor, which is partly due to variations in delays, bandwidth limitations, and packet losses. Although continuous media applications can tolerate some missing data, non-recoverable information loss degrades these applications' QoS. Consequently, a number of application areas (e.g., those related to the entertainment industry) have backed away from streaming of their content over the Internet. Inability to control the resulting visual and auditory quality of the resulting streamed presentation is an important reason for such a trend. We believe that this trend can be reversed. To this end, our work focuses on providing high- quality streaming through the exploitation of multiple paths existing in the network. By high quality, we mean with significant bandwidth requirements, of relatively long duration, and without information loss or hiccups in data delivery. In this article we present evidence that multi- path streaming is a promising approach. [ABSTRACT FROM AUTHOR]

Investigates the analytical performance models of system sizing and resource allocation in interactive video-on-demand (VOD) systems. Resource requirement of the VOD system; Benefit of data sharing technique; Reduction of the system cost.

Presents a K-server threshold-based queuing system with hysteresis in which the number of servers is governed by a forward threshold. Applications for a threshold-based queuing system; Description of the proposed model; Use of a bounding technique to evaluate the performance of a Video-on-Demand storage server.

Discusses alternative schemes which provide a high degree of reliability at low disk storage, bandwidth and memory costs for video-on-demand (VOD) multimedia servers. Issues and trade-offs associated with providing fault-tolerance in multidisk VOD systems; Sequentiality of video delivery; Buffer memory requirements.

Describes how the clients of media-on-demand servers may be partitioned into sync-classes, or sets of clients who can be serviced through the allocation of a single set of resources. Use of a very large class of cost functions to determine which partition to choose; Algorithms for computing optimal splits; Reasons why the proposed framework is generic.

This comment points out several mathematical errors in the proof of Therorem 3, and gives the correct expression of B3 . [ABSTRACT FROM AUTHOR]

Introduces several articles about multimedia systems. Scheduling algorithms for maximizing profits of a broadcast-based electronic delivery service for purchasing digital products; On-demand multicast delivery of streaming media; Approach for indexing animated objects; Construction of communication subgraphs for a distributed virtual environment system.