A Concurrent-Push Parallel Server Architecture - Scalable Continuous Media Streaming Systems - page 184

And the corresponding average scheduling delay given a system utilization of n is

M avg ( n ) Q
R V

D S = (10.38)

As the admission scheduler reduces the transmission jitter to equal to the clock jitter, the
new prefill delay can be obtained by replacing δ with τ in equation (10.23):

2 τ + T F +

f + − f − − T E

f +

D P = + (10.39)

T F

10.5 Sub-Schedule Striping Scheme
The AGSS algorithm presented in the previous section substantially reduces the server buffer
requirement as well as the scheduling delay. However, the client buffer requirement and,
consequently, the prefill delay are only slightly reduced as a side effect of the admission
scheduler. In this section, we consider another modification to the concurrent-push algorithm
that can substantially reduce the client buffer requirement and the prefill delay.
Specifically, the analysis in Section 10.3 reveals that the main reason for the increase in
client buffer requirement with the number of servers stems from the increase in the average
filling time in equation (10.1). This suggests that we can reduce the buffer requirement by
using smaller striping size Q . However, as the server retrieves data from the disk in units of Q
bytes, reducing the striping size will adversely affect disk retrieval efficiency.
To solve this problem, we propose decoupling the transaction size for disk retrieval and
transmission from the striping size - sub-schedule striping (SSS). In particular, we maintain
the disk transaction size at Q bytes but use a striping size (denoted by U ) inversely proportional
to the number of servers in the system (Figure 10.6):

U = Q / N S (10.40)

S 0 . . .

5 0

S 1 . . .

6 1

. . .

7 2 9 8 7 6 5 4 3 2 1 0

S 2

Playout

. . .

Video Client 8 3

S 3

. . .

S 4

9 4

Qbytes
(Transaction Size)

U bytes
(Striping Size)

Figure 10.6 Data organization in sub-schedule striping