One-way delay over time
Time is represented by the packet sequence number. The assumption
is that the traffic is issued by a constant bit rate source (300 Kbps)
and that the inter packet gap is constant, so that for each time interval
only a single packet is received.
Comments:
Note: noBE: without best-effort, BE: with best-effort
- Measurement results of one-way delay and instantaneous packet
delay variation
show that the premium buffer size (PBS) does not impact
the one-way delay evolution in time both with
and without background traffic.
This is true for any packet size in the following
set: [5000, 10000, 15000, 20000, 45000] bytes. The only exeption is
PBS equal to 1500 bytes with best-effort. In fact,
only in this case when best-effort traffic is added the
maximum one-way delay can be up to 19.23 msec
(figure 8),
while without best-effort traffic the maximum one-way delay is
of 9 msec. Then, we could conclude that in this case
the scheduler does not isolate strams properly. This issue
needs further investigation.
- (noBE) Independently of the packet size, delay seems
to be periodic, where the frequence is a function of
the packet size (see
figure 7). In particular, within
each cycle, again latency varies periodically with
a different frequence. This is illustrated in figure
10.
- Without best-effort traffic latency varies within a given
range which depends on the packet size. For example,
for 1024 bytes packets, the range is
[4.57, 7.00] msec, as
figure 10 shows.
A similar (periodic) evolution in
time is measured when best-effort traffic is added
(see figure 11).
However, in this case, while the min value is almost
identical (4,63 msec), the maximum latency
depends on the PBS. For PBS = 1500 bytes the maximum
latency can be up to 14.50 msec, while for other
PHB values the maximum is lower, for example 11.00 msec
as figure 12 shows.
Time contraints for premium
traffic do not seem to be met even if premium is
treated as high priority traffic by the scheduler. PB
size is only 1500 bytes and the weight is such that
the premium queue never builds up. High latency values
do not seem to be due to queueing effects. This problem
is under investigation.
Figure 7: one-way delay in time for PBS = 1500 bytes.
For each packet size latency
is periodic. Then again within each cycle latency varies
with a different frequence.
Figure 8: one-way delay in time for PBS = 1500 bytes
with background traffic. The exact periodicity of figure 7
is lost, however, a less periodic behavior is still visible
for a given packet size.
Figure 9: one-way delay in time as in figure 8 but with
PBS = 10000 bytes. The maximum delay is around 13.57 msec, while
for PBS = 1500 bytes, the maximum is 19.23 msec.
The following graphs are extracts of a single curve (corresponding to
a packet size of 1024 bytes) from the previous graphs. These curves
clearly show the periodicity in one-way delay over time.
Figure 10: zoom on one-way delay in time for packet size equal to
1024 bytes (PBS = 1500 bytes).
Two cycles are visible. Values are in the range [4.57, 7.00] msec.
Figure 11: zoom on one-way delay in time with background traffic for packet size equal to
1024 bytes (PBS = 1500 bytes). Cycles are less visible in this case and
values are in the range [4.5, 14.5] msec.
Figure 12: like figure 11 but with a PBS = 10000 bytes.
Values are in the range [4.5, 11.00] msec.
In comparison to figure 11, the maximum one-way delay value is smaller.
This result is against our expectations, since larger buffer are
supposed to introduce queuing delays.
Similar figures are achived for PBS in the following set:
[5000, 10000, 15000, 20000, 45000] bytes.
Last modified: Nov 22, 1999