Information Technology Reference
In-Depth Information
to negligible levels (order of 10
−
1
percentage bit-rate variation averaged over 1 second interval)
without any impact on other parts of the system.
18.7.2 Experimental Results
We conducted extensive benchmarking experiments to collect three performance results,
namely start-up latency, aggregate bit-rate of all channels, and peak aggregate reception bit-rate
to compare with the theoretical calculations. In all experiments, we use the system parameters
of
L
42. The media stream is aMPEG-1 encoded system
stream multiplexing one video stream with one audio stream. We conducted benchmarks for a
total of 7 GCB system configurations, with the number of media segments
N
ranging from 50
to 1,000. For each configuration, we obtain the performance data by averaging data collected
from 20 benchmark runs. The results are summarized in Table 18.2.
We first consider start-up latency that is measured fromwithin the client software. The results
show that the experimental results agree closely with the theoretical calculations. The minor
differences are likely due to network delay and software processing delay. Next, we measured
the aggregate network bit-rate of all channels using a hardware protocol analyzer connected to
the Ethernet switch's mirroring port, which forwards all packets passing through the switch.
The measured results exhibit a consistent 5% increase in bandwidth usage compared to the
theoretical calculations. This increase is due to the header overheads in the application-layer
protocol (8 bytes), UDP (8 bytes), IP (24 bytes), and Ethernet (18 bytes). With a UDP datagram
payload of 1,400 bytes, the combined header overhead is equal to (8
=
4
,
401,
C
=
2
.
84,
m
=
2 and
b
=
1
.
+
8
+
24
+
18)
/
1458
=
4%, which closely matches the measured results.
Finally, we measure the aggregate reception bandwidth usage in the client access link, again
using a hardware protocol analyzer. Unlike the aggregate network bit-rate, the reception bit-
rate is not constant and does vary depending on which media segments are being received.
Nevertheless, we are more interested in the peak bandwidth usage and thus we measure the
maximum bandwidth usage averaged over a 10-second window. The results show similar
header overhead-induced bit-rate increases (
5%) for configurations with
N
up to 200. For
larger values of
N
, the differences widen further up to 9.51%. Our study of the log data shows
that two factors lead to the bit-rate increase.
∼
Table 18.2
Comparison of theoretical and experimental results (with
m
=
2)
Config Latency Aggregate Bit-rate of all channels Peak aggregate reception bit-rate
N
/
N
G
*
Theory Measured Theory Measured Difference (%) Theory Measured Difference (%)
50
/
19 176.04
176.10
5.8208
6.11
+
4.97
2.84
2.98
+
4.93
80
/
21 110.03
110.05
6.8302
7.17
+
4.97
2.84
2.98
+
4.93
100
/
22
88.02
88.04
7.2779
7.64
+
4.98
2.84
2.98
+
4.93
200
/
28
44.01
44.05
8.6611
9.10
+
5.07
2.84
3.00
+
5.63
500
/
33
17.60
18.00
10.6277
11.16
+
5.00
2.84
3.05
+
7.39
800
/
37
11.00
11.20
11.5828
12.16
+
4.98
2.84
3.11
+
9.51
1000
/
38
8.80
9.00
12.1096
12.71
+
4.96
2.84
3.09
+
8.80
Note.
*
N
and
N
G
are the number of video segments and number of channels respectively.
Search WWH ::
Custom Search