Continuous Media Storage and Retrieval - Scalable Continuous Media Streaming Systems - page 50

.

Group 2

.

. . . . . .

Group 1

Transmission

. . .

. . . .

. . . . . .

. . .

Retrieval

. . .

Group 1 Group 2 Group 1 Group 2

Figure 3.6 Illustration of the Grouped Sweeping Scheme with two groups

rounds of the g different groups. Figure 3.6 illustrates this GSS scheduler with two groups of
media streams.
There are two important characteristics in GSS. First, a media stream can join any one of the
groups as long as the group is not already running at full capacity. Second, transmission of a
data block lasts for a macro service round (i.e., g micro service rounds) to sustain a continuous
transmission of media data to the client while waiting for the next media block to be retrieved.
Using GSS, the retrieval of a data block can be completed in a shorter time than in the case
of SCAN/CSCAN. For example, suppose at full capacity the server can serve k concurrent
media streams. The service round length in SCAN will then be equal to t max

round ( k ). However, if

/

we employ GSS with two groups, then each group will serve up to k 2 requests and so the

micro-round length will be equal to t max

round ( k

/ 2) while the macro-round length will be equal to

2 t max 2).

This shorter service round length is desirable in terms of buffer requirement. A buffer is
occupied from the time a service round begins to the time the data retrieval is completed. Thus,
comparing GSS to CSCAN, it is clear that the buffer-holding time is reduced from t max

round ( k

/

round ( k )to

t max 2), thus allowing the buffer to be reused more quickly for the next service round. With

g groups, it can be shown [1] that the buffer requirement of GSS is given by

round ( k

/

k k

g

Q

B GSS = + (3.8)

Comparing GSS to CSCAN's buffer requirement 2 kQ , it is clear that we can reduce the buffer
requirement by using GSS with more groups (i.e., increasing g ).
On the other hand, a new user joining the system running GSS may also experience shorter
admission delay. This is possible because the new stream can join any of the g groups to
receive service. If the new stream arrives in the middle of, say, micro-round serving group
i , for example, then it can simply join the next group (( i 1) mod g ) provided that it is not

already fully utilized. In this case the average delay is only 1.5 t max

+

round ( k

/ g ), which is shorter than

the corresponding delay under CSCAN (i.e., 1.5 t max

round ( k )).
From the previous discussion it seems that we should use GSS with as many groups as
possible to reduce buffer requirement and admission delay. Taking it to the limit for a server