This paper presents a cluster-based video server to achieve good
scalability and fault-tolerance. Since a popular video file easily becomes
a hotspot, so Tiger divides a file into blocks, and scatters them on
different cubs; and thus the access to the different parts of the same
file could be concurrently scheduled. Moreover, to avoid conflicts on disk
requests, a distributed scheduler algorithm is presented to specify the
service order of viewers, and each cub follows their scheduler. In this
way, Tiger can support a larger number of viewers based on a cluster of
cheap machines.
This scheme achieves good overall throughput by increasing the latency of
the start time of a streaming because a new viewer may have to wait for an
idle slot in the scheduler. This may lose some impatient viewers. Also, the
distributed scheduler algorithm is more complex, by introducing ownership
and relaxed consistency model, than a simple central controlled scheduler,
and thus the scalability is achieved by sacrificing the simplicity. In
addition, to achieve ideal load balancing, the requests to the same file
should be evenly spaced while this is rare in reality.
In brief, this paper sounds neat at its published time. If it could
provide more discussion about mirroring techniques v.s. coding techniques
for choosing replication algorithms, and analyze more about disk access
bottleneck v.s. bandwidth bottleneck for streaming, this papers would be
more profound.
Received on Wed Nov 30 2005 - 22:15:57 EST
This archive was generated by hypermail 2.2.0 : Thu Dec 01 2005 - 00:46:13 EST