A Concurrent-Push Parallel Server Architecture - Scalable Continuous Media Streaming Systems - page 189

15

10

5

Concurrent Push (Max)
Concurrent Push w/ AG SS (Max)
Concurrent Push w/ AGSS (Average)
Concurrent Push w/ AGSS & SSS (Max)
Concurrent Push w/ AGSS & SSS (Average)

0

0 5 10 15 20

Number of Servers

Figure 10.9 System response time versus number of servers

the scalability of the parallel video server architecture to a maximum of 408 servers serving
a total of 3,672 concurrent video sessions at 90% utilization. If 1GB memory is available, the
architecture can be scaled up to 14,400 concurrent video sessions using a client-server ratio
of 250 at 90% utilization.
A second, more subtle limiting factor is due to the sub-schedule striping scheme. Under this
scheme, the client must resequence the incoming data by copying U -bytes stripe units into the
client buffer (Figure 10.6). Hence the processing overhead will likely increase with smaller
striping size. Our previous experiences showed that processing overhead remains practical for
software implementations running in even low-end PCs for striping size as small as 1KB. This
limits N S to 64. For larger systems, we can use more powerful server hardware with a larger
client-server ratio to avoid reaching this limit. In the previous example with 1GB memory,
we increase the client-server ratio to 250 to limit to a total of 64 servers. Clearly the rapid
improvement in CPU speed will undoubtedly extend this limit.

10.7 Summary
In this chapter, we have presented and analyzed a concurrent-push parallel video server archi-
tecture for designing scalable video-on-demand systems. The proposed architecture employs
fixed-size block striping and the server-push service model. To schedule disk retrievals and
transmissions, we introduced a concurrent-push scheduling algorithm where video data are
continuously transmitted from all servers to a client station. This constant-bit-rate traffic pro-
duced by the algorithm enables us to take advantage of the quality-of-service guarantees
provided by the networks. To extend the scalability of the architecture, we introduced the
Asynchronous Grouped Sweeping Scheme and the Sub-Schedule Striping Scheme into the
architecture. Results showed that the resultant architecture can be scaled up to more than ten
thousand concurrent users with acceptable buffer requirement and system response time.
Building video-on-demand systems upon parallel server architecture not only breaks through
the capacity limit of a single server, but also opens the way to fault-tolerant system designs.