Review - Distributed Schedule Management in the Tiger Video Server

This paper introduces Tiger, a video server designed to be scalable,
fault tolerant and robust. The Tiger system is designed to handle
thousands of concurrent streaming requests, with the idea that no two
requests will ask for the same block at the same time. To achieve this
goal, the authors present a cluster of commodity computers on an ATM
network fabric that store video files striped across all servers and
across all disks. They use a distributed scheduling algorithm to remove
the single point of failure of a global scheduler and show that they can
achieve relatively good performance.

The authors address an important problem of video streaming. The idea of
striping the file across all servers is neat in that it distributes the
load evenly across all servers; especially helpful for the very popular
videos. They also use the idea of a hallucination such that the
scheduling algorithm is distributed and help scale the system. They also
show good performance (and little control data 0.16Mbps for 600 streams)
and show that their system fares well in the face of failures.

The weaknesses of this paper lies in the fact that the authors assume
independent failures. How true is that? Also, they absolutely try to
maintain a homogeneous system (replacing with identical disks on
failure). It is unclear how the system performs in a heterogeneous
system, a relistic situation since there will be entropy introduced as
the system ages. I believe that the performance will be limited by the
slowest/oldest node in the system which is not desirable as resources
will be wasted.

Although scalability is important in any cluster, the main problem for
video servers is caching and in this system, this is impossible to do.
Caching would be a good improvement for any video file servers. This
system also uses mirroring for data availability, clearly inefficient.
This can be improved using erasure coding for example. It would also be
good to see a system that works on a WAN, where the assumption of
bounded latency no longer holds and where control packets can be lost.
Received on Thu Dec 01 2005 - 00:46:12 EST