A Concurrent-Push Parallel Server Architecture - Scalable Continuous Media Streaming Systems - page 187

6

w/o AGSS

4

w/ AGSS

2

w/ AGSS & Sub-Schedule Striping

0

0 5 10 15 20

Number of Servers

Figure 10.7b Client buffer requirement versus number of servers

10.6.3 System Response Time
We first plot scheduling delay versus the number of servers in the system in Figure 10.8a.
The case for sub-schedule striping is not plotted, as sub-schedule striping has no effect on the
scheduling delay. Note that the worst-case scheduling delay is substantially reduced by AGSS,
especially for large number of servers. Moreover, the average scheduling delay with AGSS is
even smaller and stays relatively constant, regardless of the number of servers in the system.
For example, with AGSS striping in a 16-server system, the average delay is only 1.26 seconds
for system utilization as high as 90% even though the worst-case scenario is 9.18 seconds.
The worst-case delay is even larger (13.98 seconds) without AGSS. Hence with AGSS, we can
maintain a reasonably short scheduling delay by operating the system to within, say, 90% of
the total capacity.
Figure 10.8b plots the prefill delay versus the number of servers in the system. The results
show that the prefill delay is also reduced by AGSS because the worst-case transmission jitter
T F is larger than the clock jitter

. More importantly, by using sub-schedule striping, the prefill
delay becomes completely independent of the number of servers in the system.
Finally, we plot the total system response time in Figure 10.9. Clearly the proposed AGSS
and sub-schedule striping can effectively maintain a small system response time (1.8 seconds
for N S =

τ

16 at 90% utilization) even if the number of servers is large.

10.6.4 Scalability
The results in the previous sections have shown that both the client buffer requirement and
the system response time can be kept low irrespective of the number of servers in the system.
Server buffer requirement is the only factor that increases with more servers. This factor will
certainly limit the ultimate scalability of the system. Nowadays, it is common to install 256MB
or more memory in a PC-based server as memory price has dropped substantially. Under our
system parameters and ignoring operating system overhead, a 256MB memory size will limit