Elsevier

Information and Software Technology

QoS-based evaluation of file systems and distributed system services for continuous media provisioning☆

Abstract

This paper presents a QoS-based performance analysis of file systems and distributed object services for Continuous Media (CM) provisioning, as well as the details and implementation experiences of a continuous media file system and associated CM servers. For this we have implemented QoS-driven CM servers and the Presto continuous media file system (PFS) in the context of a distributed multimedia application development environment, and validate the performance of PFS against that of the conventional Unix file system through an experimental evaluation. Using our CM server prototype, we next examine the effect of continuous media data delivery on the three different kinds of network protocols such as CORBA, UDP/IP, and TCP/IP, with respect to QoS provisioning and throughput.

Introduction

Given the very large size of continuous media data [2] (e.g. video), most multimedia applications must access data from secondary and/or tertiary storage during execution. This makes the performance of continuous media file systems critical to overall system performance. Additionally, access to continuous media requires that some Quality of Service (QoS) parameters (specified by the application) be met. These include specification of the retrieval rate and its variation, allowable timing drift, acceptable averages and bursty errors, and the degree of synchronization. The Unix File System (UFS) has been the landmark achievement in file system design, and practically every modern file system borrows heavily from it. However, such conventional file systems do not have any notion of QoS, and thus fail on two counts. First, they are unable to provide any guarantees regarding QoS, and second they are not able to take advantage of the application's tolerance (specified by the QoS parameters) in optimizing the file system performance. As a solution to those problems, we developed a prototyped continuous media file system, the Presto File System (PFS) [11], which stores files of continuous media streams such as audio and video efficiently.

While the mechanisms provided by the UFS have been sufficient for most applications, there are important classes of applications where it has not been so [33]. An increasingly important class of applications where UFS is not suitable are those requiring storage and retrieval of Continuous Media (CM); i.e. audio, video, animation, etc. [32]. The following are some unmet file system needs of continuous media:

•

Treatment of files as a sequence of bytes by current file systems preclude modeling and thereby prevent optimized timely retrieval of time sensitive CM data such as audio and video. Untimely jittery retrieval causes user dissatisfaction [32], and in conventional file systems, these increase with the load.

•

Given the high level of redundancy and temporal autocorrelation of continuous media, its human consumption is very tolerant to controlled errors. The system should be able to understand the specifications of acceptable tolerances, also called Quality of Service, and should be able to take advantage of them in resource management.

Based on the above observations, the following issues must be addressed in designing a continuous media file system:

•

Placement and storage structures: The placement of CM data and an indexing mechanism of file systems should be such that they are capable of dealing with large sizes with real time retrieval needs that are embodied in QoS metrics.

•

Scheduling: The goal of disk/server scheduling for continuous media is to satisfy QoS requirements by meeting the deadlines of periodic I/O requests generated by some server resident CM stream manager with minimum buffers.

The key to designing high performance file systems to support CM requires that we use information about: (i) the inherent temporal nature of continuous media, and (ii) QoS oriented nature of media access, to develop new solutions to these problems.

In this paper, we present the design and implementation of a novel continuous media file system, and describe QoS metrics, which we will use for our simulation on QoS provisioning. Experimental evaluations on throughput are also performed.

Based on the experience on the development of a new file system, we design and implement QoS-based continuous media file servers, which use PFS and UFS as their underlying file systems. Using our CM server prototype, we examine the effect of continuous media data delivery on the three different kinds of network protocols (i.e. CORBA, UDP/IP, and TCP/IP) with respect to QoS provisioning and throughput by performing extensive experimental evaluation. Network speed has been continuously increasing; therefore, services like Video on Demand, Teleconferencing, and Distance Learning are very likely to be popular in everyday life. Yet, given the limitations of the current network bandwidth, straightforward TCP implementations are not suitable for such bandwidth-sensitive applications [4], [6], [27]. TCP has its own flow control mechanisms, error detection and retransmissions, all of which add extra time as well as network bandwidth overhead, to the transmission. This causes unexpected and unpredictable delay and jitter time when transferring CM data, when timing is one of the most critical requirements of CM applications. However, most CM applications do not need highly reliable transmission. Losing some frames is less important than having too much delay jitter or losing synchronization between streams. Thus, we could say that UDP is more suitable for CM applications than TCP. Specifically, we can take advantage of the fact that even UDP is a lossy protocol, the number of frames lost is not that many; in most cases, it is still in an acceptable range. This fact gives us an important observation: adding some control channel and having more intelligent dropping mechanisms will give us good performance in terms of both timing quality and number of lost frames. Our CM servers provide a QoS-driven dropping mechanism in order to reduce the drift and loss QoS values as much as possible, compared to the non-dropping and pure dropping mechanisms.

We ported our original Socket(TCP-UDP/IP)-based CM server [17] on a CORBA environment. We use version 2.0 of Orbix from IONA Technologies as CORBA system [1]. This CORBA-based CM server replaces all C socket calls with stubs and skeletons generated from a pair of CORBA interface definition language (IDL) specifications. The IDL specification uses sequence parameters, for the data buffer rather than string parameters, which are a bit slow.

To the best of our knowledge, our study on PFS is the first one to evaluate file system performance for managing continuous media in terms of how well the user-specified QoS is attained. We have compared the experimental behavior of the Unix file system, as instantiated in the Solaris 2.5 operating system, with that of the Presto file system. The experimental data used conformed to that specified for motion JPEG video. Our principal findings are [18] as follows. (1) For general file operations, i.e. non-continuous media access, the maximum aggregate throughput obtained from PFS was about 80% more than that for UFS. While the throughput of either file system increases by increasing the size of the block accessed from the disk in each read (at the expense of having larger memory buffers), the range in which PFS can take advantage of this is much larger than that for UFS. (2) For motion JPEG data, where the mean size of a frame is about 18   KB, where PFS outperforms UFS by about 25%. (3) For QoS parameters related to time drifts, PFS outperforms UFS by a huge margin. (4) For QoS parameters related to frame loss, the loss suffered by PFS is less than one third of that suffered by UFS.

We present a general methodology of QoS study to characterize continuity metrics for lossy continuous media. We show a development of QoS-based continuous media servers and present a performance evaluation of these prototyping environments in video streams with respect to a set of QoS metrics designed to measure continuity losses and timing drifts: loss and drift QoS factors. To reduce QoS loss and drift factors in our CM server, we adopt a new strategy, called QoS-Driven Dropping Mechanism, and it achieves good performance compared to the other two prior approaches such as non-dropping and pure-dropping mechanisms. In a series of experiments on our QoS provisioning in CM streams, we observe that TCP/IP is not suitable for distributed CM applications which require high network bandwidth and timing-criticality, and that UDP/IP is one of the alternatives. Besides the traditional C-socket (TCP-UDP/IP)-based CM server mechanisms, we implemented our CM server on CORBA environments. It turns out that the CORBA-based implementation shows better performance on QoS provisioning than the TCP-version, but UDP still works the best. It is mainly because the data copying overhead and the higher fixed overhead of CORBA implementations such as demultiplexing and memory management considerably limit the performance.

The rest of the paper is organized as follows. In Section 2, we describe the QoS metrics used in the experiments. Section 3 describes the architecture of PFS and its implementation, and Section 4 does so for the CM server. Section 5 presents a detailed experimental comparison of PFS and UFS for basic file operations, and for continuous media access. In Section 6, we present the results of the experimental evaluation of our CM Server system, based on the QoS metrics described in Section 3 for several variable factors such as file systems, network protocols, server types, and number of streams. We describe the prior related work in Section 7. Section 8 presents the concluding remarks.

Section snippets

QoS model for continuous media

In order to achieve high performance in designing our CM server, we should consider the resource constraints as well as the properties of CM streams. These CM streams have their own features and special QoS metrics. Since we decided to design our CM server and clients based on the lossy UDP, we adopted the QoS metrics suitable for the lossy protocol. These QoS metrics play an important role in our QoS-driven CM server (in particular, QoS Manager).

PFS design and implementation

This section describes design objectives and requirements of PFS and compares it with conventional file systems (e.g. UFS).

CM server design and implementation

Our CM server system is a typical client/server application. It includes one CM server and some CM clients. The CM server serves the clients concurrently. Fig. 3 depicts the architecture of the CM server, which plays out multiple (or single) streams to requesting clients across the network. The CM server has four components. The Network Manager responds to the clients' connection requests. The QoS Manager is responsible for admission control and I/O scheduling. Each Proxy Server communicates

Comparison of PFS and UFS on standard file operations

In this section, we demonstrate the advantages of PFS over UFS based on experiments that use standard file operations. The experiments were carried out on a Sun Ultra Sparcstation with a 4GB SCSI disk. The focus of our experiments was to verify that for standard file operations, the performance of PFS is comparable to that of UFS.

Performance evaluation of distributed system services

In this section we compare the performance results of the three kinds of CM server architectures (non-dropping, pure dropping, and QoS-driven dropping) based on the three different network protocols (TCP, UDP, CORBA), which were evaluated by experiments using QoS metrics such as ALF, CLF, ADF and CDF.

Related work

Storage and transmission are both critical to the successful provisioning of multimedia services in a distributed environment. However, in general these have been studied separately, which is reflected in our literature survey below.

Conclusions

In this paper, we described the Presto File System implementations, the CM server architecture and the techniques working for both C/Sockets and the CORBA network environments. We have specified and represented QoS metrics that are applicable to lossy channels like UDP and have done our CM server performance evaluation based on those metrics. These QoS metrics in general include the specification of retrieval rates and its variation, allowable timing drift, acceptable average and bursty errors,

Acknowledgements

The ideas implemented in the CM servers have benefited a lot by discussion with our colleagues, Raja Harinath, Hung Q. Ngo of University of Minnesota, and Deepak R. Kenchamanna-Hosekote of IBM Almaden Research Center, CA. We would like to thank all of them for their valuable suggestions and some part of implementation on this work.

References (37)

  • Orbix 2 Programming Guide, IONA Technologies,...
  • et al.

    A files system for continuous media

    ACM Trans. Comput. Syst.

    (1992)
  • M. Chen, D. Kandlur, P. Yu, Support for fully interactive playout in a disk-array-based video server, in: Proceedings...
  • Z. Chen, S.-M. Tan, R.H. Campbell, Y. Li, Real time video and audio in the world wide web system, in: Fourth...
  • A. Dan, D. Sitaram, P. Shahabuddin, Dynamic batching policies for an on-demand video server, in: Proceedings ACM...
  • A. Eleftheriadis, S. Pejhan, D. Anastassiou, Algorithms and performance evaluation of the Xphone multimedia...
  • C. Federighi, L. Rowe, The design and implementation of the UCB distributed video-on-demand system, in: Proceedings...
  • D Gemmell et al.

    Principles of delay sensitive multimedia data storage and retrieval

    ACM Trans. Inf. Syst.

    (1992)
  • D Gemmell et al.

    Multimedia network file servers: multichannel delay sensitive data retrieval

    Multimedia Syst.

    (1994)
  • R. Haskin, F. Schmuck, The tiger shark file system, in: COMPCON 96,...
  • J. Huang, J. Richardson, D.R. Kenchamanna-Hosekote, J. Srivastava, Presto: final technical report. Technical report,...
  • K Jeffay et al.

    Adaptive, best-effort delivery of digital audio and video across packet-switched networks

    Lect. Notes Comput. Sci.

    (1993)
  • D. Kenchammana-hosekote, J. Srivastava, Scheduling continuous media in a video-on-demand server, in: Proceedings IEEE...
  • D. Kenchammana-hosekote, J. Srivastava, Retrieval techniques for compressed video streams, in: Proceedings of...
  • W. Lee, B. Sabata, Admission control and QoS negotiations for soft-real time applications, in: Proceedings of the Sixth...
  • W. Lee, J. Srivastava, W.-H. Lee, Adaptive disk scheduling algorithms for video servers, in: Proceedings of the 28th...
  • W. Lee, D. Su, H.Q. Ngo, J. Srivastava, A QoS-driven networked continuous media server, in: Proceedings of SPIE...
  • W. Lee, D. Su, J. Srivastava, D. Kenchammana-hosekote, D. Wijesekera, Experimental evaluation of PFS continuous media...

    • Cited by (3)

    View full text

    Copyright © 2000 Elsevier Science B.V. All rights reserved.