Recovery Strategies for Distributed Multimedia Applications

This work demonstrates the potentialities of a simulated fault-injection based approach to the design of effective recovery strategies for distributed multimedia systems. Fault injection experiments are conducted on a simulated model of a real prototype of a Video on Demand (VoD) distributed application. The model is built by combining the basic objects of a simulation tool for dependability (reliability, availability, and performability) prediction and evaluation of distributed multimedia systems, currently under development at the University of Naples. Results demonstrate that the approach, and the tool are suited for evaluating the effectiveness of potential recovery strategies, in the early stages of the design process, when the architecture of the system has not been defined yet, or in the early prototype phase, when some components have been already implemented, but there is still potential for alternative solutions and/or substantial modifications. 1 Application Architecture The application used as a case study is called Distributed Video Architecture (DiVA) and is currently under development at the University of Naples. DiVA [4] is an interactive client-server system for retrieval and real-time delivery of video sources from a remote distributed database. Figure 1 – Components of the DiVA system It can be operated on a whatever collection of heterogeneous computing nodes connected by an IP-based network infrastructure. DiVA general structure is illustrated in Figure 1. Video/audio documents are distributed over the network. At each computing node, documents are managed by the local Film Server (FS). A central component, called Video Server (VS), provides the Video Clients (VCs) with functions to search the database and select the clips. Clients and servers are synchronized throughout the transmission of audio/video data. Multiple sessions may be active concurrently. Typically, a user would require the list of available films and then choose one of them for display. Video is stored in MPEG format [3]. 2 Simulation Environment The simulation environment consists of a set of objects, which are a generalization of fundamental components typically found in most distributed multimedia systems for VoD applications. Thanks to the flexibility of object oriented technology, these basic entities provide a framework which can be easily extended and/or customized, in order to more accurately model specific aspects of the application under study. The current implementation of the environment is based on Depend [1], a simulation tool for dependability analysis developed at the University of Illinois, which supports hierarchical simulation [2]. The environment is described in detail in [6]. 3 Experimental Results The objective of this study was twofold. First, we wanted to develop an effective recovery strategy capable of providing the user with an alternative URL in case of failure of the server process. Second, we wanted to demonstrate the potentialities of the simulation tool we have developed. In order to do this, we performed extensive fault injection experiments on a simulated model of the DiVA system. The configuration of the the real system was as follows: 3⁄4 A Video Server process running on a Sun Sparc 4, located in Milan; 3⁄4 Four Video Client processes running on two Parallax equipped Sparcs and two Pentium II PCs, located in Milan; 3⁄4 Two Video Client processes running on two Sparcs, located in Naples; 3⁄4 Two Video Archive processes running on two Sparcs, located in Milan; 3⁄4 A Video Archive process running on a Pentium II PC, located in Naples. The computers in Milan and Naples were connected via a 100 Mb/sec and a 10 Mb/sec Ethernet switch, respectively. The link between the two LANs had a bandwidth of 2 Mb/sec. The simulated model was built by customizing the tool‘s built-in objects according to the characteristics of the experimental testbed described above. Performance parameters such as the link capacity, the decoding and rendering time, and the disk I/O time for simulation setup were derived from a trace file collected on the field. We evaluated the impact of the introduction of possible transmission recovery strategies by investigating the behavior of the system while faults were injected to the Film Servers. The results for one of such procedures are reported in the following. Faults have been modeled according to exponential laws. The assumed ranges for fault parameters were as follows: λs − fault rate: from 0.05 [1/h] to 50 [1/h]; μs − fault mean duration: from 0.00025 [h] to 0.25 [h]; σs − standard deviation from mean duration: 0.1*μs. Fault injection experiments were conducted under varying load conditions, in order to take into account the impact of load on system ability to recover from server failures. The Video Client Video Server