A Staggered-Push Parallel Server Architecture - Scalable Continuous Media Streaming Systems - page 217

12

A Staggered-Push Parallel
Server Architecture

The concurrent-push architecture discussed in the previous two chapters is designed to take
advantage of network-level QoS services and is also readily scalable to over ten thousand
concurrent streams. In practice, however, not all broadband networks support resource
reservation or QoS control. In this chapter we develop an alternative approach to schedule
the transmissions of media data from the parallel servers - a staggered-push approach.While
media servers in the concurrent-push architecture transmit media data simultaneously in
proportionally reduced bit-rate, the media servers in staggered push effectively transmit
data to the same client in turns, with only one of the servers sending data at any time. As a
result, the data traffic will be more bursty and this is also why staggered push cannot easily
take advantage of network resource reservation services.
In return, the staggered-push architecture can achieve linear scalability, i.e., the per-
stream server resource requirements (e.g., buffer requirement) are invariant to the scale
(number of servers) of the system. This chapter presents this staggered-push architecture,
explains its admission, scheduling, and buffer management schemes, and analyzes its per-
formance.

12.1 Introduction
In this chapter we present and analyze quantitatively a staggered-push architecture for schedul-
ing disk retrieval and network transmission in parallel video servers. We prove a remarkable
property of the staggered-push architecture - the system can be scaled up linearly to an ar-
bitrary number of servers as long as the network has sufficient capacity. We discover that for
loosely coupled servers like PC or workstation clusters, server-clock asynchrony could lead
to inconsistent schedule assignments among different servers. To tackle this problem, we in-
troduce an external admission scheduler to centralize admission control and perform schedule
assignments. Apart from inconsistent schedule assignments, we discover that server-clock
asynchrony could also lead to overlapping between data transmitted from different servers.