57.7ms increase in measured RTT at
the TCP layer. The effect of packet loss
can be devastating to the responsiveness of TCP applications, but it is clear
that passively measuring network level
RTTs is still feasible, and distinct from
the perceived latency experienced by
applications that can be introduced by
TCP’s in-order delivery semantics.
Experiment 3: Bufferbloat. Misunderstandings around the relationship
between loss prevention and network
performance have led to excessive
buffering being introduced to forwarding and routing hardware as a
loss-avoidance strategy. Often (but not
exclusively) this affects commodity
customer premises equipment (CPE),
and thus directly affects end users.
However, excessive buffering works
against TCP’s loss-detection algorithm by increasing delay and thus delaying the time taken for a TCP stack
to identify loss and back-off; that is,
the additional delay introduced by
large buffers can disrupt TCP’s congestion-control mechanism.
figure 4. Rtt histograms representing the difference in Rtts.
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
(a) Range of application-layer Rtts
reported with different buffer sizes.
(b) Range of tCP Rtts reported
with different buffer sizes.
10kb buffer
100kb buffer
200kb buffer
300kb buffer
10kb buffer
100kb buffer
200kb buffer
300kb buffer
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
140
120
100
80
60
40
20
0
0 50 100 150 200 250 300
Rtt (ms)
fre
quen
c
y
fre
quen
c
y
fre
quen
c
y
fre
quen
c
y
fre
quen
c
y
fre
quen
c
y
fre
quen
c
y
fre
quen
c
y