Building an Internet Caching System for Multimedia Content Delivery


                            Songqing Chen


                       July 9 (Friday), 10:00 am
                              McGl 104


Committee: Phil Kearns
           Dimitrios Nikolopoulos 
           Haining Wang 
           Zhen Xiao (external member, AT&T Research)
           Xiaodong Zhang (Chair)


=========================================================


                                Abstract


The proxy has been widely and successfully used to cache the static Web
objects fetched by a client so that the subsequent clients requesting
the same Web objects can be served directly from the proxy instead of
other sources faraway, thus reducing the server's load, the network
traffic and the client response time.  However, with the dramatic
increase of streaming media objects emerging on the Internet, the
existing proxy cannot efficiently deliver them due to the following
reasons. First, streaming media objects are generally very large. To
directly cache them entirely as caching static Web objects will easily
exhaust the limited cache space and result in poor cache
performance. Second, clients requesting streaming media objects always
demand continuous streaming delivery without playback jitter and with a
small startup latency. Unfortunately, these rigorous real time
requirements cannot be satisfied by the existing proxy.

In this dissertation, we design, implement, and evaluate a
cost-effective and high performance proxy-based Internet caching system
for streaming media delivery. Addressing the conflicting performance
objectives for streaming media delivery and comprehensively considering
their trade-offs, we first propose an efficient segment-based streaming
media proxy system model. This model has guided us to design a practical
streaming proxy, called  Hyper-Proxy, aiming at delivering the streaming
media data to clients with minimum playback jitter and a small startup
latency, while achieving high caching performance.  Second, we have
implemented Hyper-Proxy by leveraging the existing Internet
infrastructure. Hyper-Proxy pushes the streaming functions from the
server to the proxy, which makes the through streaming traffic run on
UDP suffer from less data loss and no blocking. The evaluation of
Hyper-Proxy on the global Internet environment connecting Japan and US
west coast and local network environment (e.g. LAN) shows it can provide
satisfying streaming performance to clients while maintaining a good
cache performance.  Finally, to further improve the streaming delivery
efficiency, we propose a group of Shared Running Buffers (SRB) based
proxy caching techniques to effectively utilize proxy's memory. SRB
algorithms can significantly reduce the media server's load and network
traffic and relieve the bottlenecks of the disk bandwidth and the
network bandwidth.

The contributions of this dissertation are threefold: (1) we have
studied several critical performance trade-offs and provided insights
into Internet media content caching and delivery. Our understanding
further leads us to establish an effective streaming system optimization
model; (2) we have designed and evaluated several efficient algorithms
to support Internet streaming content delivery, including segment
caching, segment pref etching, and memory locality exploitation for
streaming; (3) having addressed several system challenges, we have
successfully implemented a real streaming proxy system and deployed it
in a large industrial enterprise.