Measurement Study of
Low-bitrate Internet Video
Streaming
Dmitri Loguinov
Hayder Radha
City University of New York and Philips Research
1
Motivation
• Market research by Telecommunications Reports
International (TRI) from August 2001
• The report estimates that out of 70.6 million households in
the US with Internet access, an overwhelming majority
use dialup modems (Q2 of 2001)
households (million)
60
52.2
45
30
15
9.1
4.9
3.1
1.2
0.1
Internet TV
Satellite
0
Paid dialup
ISPs
Free ISPs
Cable
DSL
2
Motivation (cont’d)
• Consider broadband (cable, DSL, and satellite) Internet
access vs. narrowband (dialup and webTV)
• 89% of households use modems and 11% broadband
subscribers (percent)
100%
88.45%
80%
60%
40%
11.55%
20%
0%
dialup
broadband
3
Motivation (cont’d)
• Customer growth in Q2 of 2001:
– +5.2% dialup
– +0.5% cable modem
– +29.7% DSL
– Even though DSL grew fast in early 2001, the situation is now
different with many local companies going bankrupt and TELCOs
increasing the prices
• A report from Cahners In-Stat Group found that despite
strong forecasted growth in broadband access, most
households will still be using dial-up access in the year
2005 (similar forecast in Forbes ASAP, Fall 2001, by IDC
market research)
• Furthermore, 30% of US households state that they still
have no need or desire for Internet access
4
Motivation (cont’d)
• Our study was conducted in late 1999 – early 2000
• At that time, even more Internet users connected through
dialup modems, which sparked our interest in using
modems as the primary access technology for out study
• However, even today, our study could be considered upto-date given the numbers from recent market research
reports
• In fact, the situation is unlikely to change until year 2005
• Our study examines the performance of the Internet from
the angle of an average Internet user
5
Motivation (cont’d)
• Note that previous end-to-end studies positioned endpoints at universities or campus networks with typically
high-speed (T1 or faster) Internet access
• Hence, the performance of the Internet perceived by a
regular home user remains largely undocumented
• Rather than using TCP or ICMP probes, we employed
real-time video traffic in our work, because performance
studies of real-time streaming in the Internet are quite
scarce
• The choice of NACK-based flow control was suggested
by RealPlayer and Windows MediaPlayer, which dominate
the current Internet streaming market
6
Overview of the Talk
• Experimental Setup
• Overview of the Experiment
• Packet Loss
• Underflow Events
• Round-trip Delay
• Delay Jitter
• Packet Reordering
• Asymmetric Paths
• Conclusion
7
Experimental Setup
• MPEG-4 client-server real-time streaming architecture
• NACK-based retransmission and fixed streaming bitrate (i.e.,
no congestion control)
• Stream S1 at 14 kb/s (16.0 kb/s IP rate), Nov-Dec 1999
• Stream S2 at 25 kb/s (27.4 kb/s IP rate), Jan-May 2000
server
campus
net work
Internet
T1 link
National
dial-up ISP
UUNET
(MCI)
modem
link
client
8
Overview
• Three ISPs (Earthlink, AT&T WorldNet, IBM Global Net)
• Phone database included 1,813 dialup points in 1,188 cities
• The experiment covered 1,003 points in 653 US cities
• Over 34,000 long-distance phone calls
• 85 million video packets, 27.1 GBytes of data
• End-to-end paths with 5,266 Internet router interfaces
• 51% of routers from dialup ISPs and 45% from UUnet
• Sets D1p and D2p contain successful sessions with streams
S1 and S2, respectively
9
Overview (cont’d)
• Cities per state that participated in the experiment
16
14 8
4
5
12
3
8
5
24
3
24
24
5
6
25
14
4
6
29
22
22
13
8
23
13
20
9
3
4
7
12
7
19
11
Legend
13
13
(total 653)
21
5
27
22
16
7
6
25
20 to 29
14 to 20
9 to 14
6 to 9
1 to 6
(14)
(6)
(10)
(9)
(12)
4
10
Overview (cont’d)
• Streaming success rate during the day shown below
• Varied between 80% during the night (midnight – 6 am) to
40% during the day (9 am – noon)
11
Overview (cont’d)
• Average end-to-end hop count 11.3 in D1p and 11.9 in D2p
• The majority of paths contained between 11 and 13 hops
• Shortest path – 6 hops, longest path – 22 hops
12
Packet Loss
• Average packet loss was 0.5% in both datasets
• 38% of 10-min sessions with no packet loss
• 75% with loss below 0.3%
• 91% with loss below 2%
• During the day, average packet loss varied between 0.2%
(3 am - 6 am) and 0.8% (9am - 6pm EDT)
• Per-state packet loss varied between 0.2% (Idaho) to
1.4% (Oklahoma), but did not depend on the average RTT
or the average number of end-to-end hops in the state
(correlation –0.16 and –0.04, respectively)
13
Packet Loss (cont’d)
14
Packet Loss (cont’d)
• 207,384 loss bursts and 431,501 lost packets
• Loss burst lengths (PDF and CDF):
15
Packet Loss (cont’d)
• Most bursts contained no more than 5 packets (however,
the tail reached to over 100 packets)
• RED was disabled in the backbone; still 74% of loss bursts
contained only 1 packet apparently dropped in FIFO queues
• Average burst length was 2.04 packets in D1p and 2.10
packets in D2p
• Conditional probability of packet loss was 51% and 53%,
respectively
• Over 90% of loss burst durations were under 1 second
(maximum 36 seconds)
• The average distance between lost packets was 21 and 27
seconds, respectively
16
Packet Loss (cont’d)
• Apparently heavy-tailed distributions of loss burst lengths
• Pareto with  = 1.34; however, note that data was nonstationary (time of day or access point non-stationarity)
17
Underflow events
• Missing packets (frames) at their decoding deadlines cause
buffer underflows at the receiver
• Startup delay used in the experiment was 2.7 seconds
• 63% (271,788) of all lost packets were discovered to be
missing before their deadlines
• Out of these 63% of lost packets:
– 94% were recovered in time
– 3.3% were recovered late
– 2.1% were never recovered
• Retransmission appears quite effective in dealing with
packet loss, even in the presence of large end-to-end delays
18
Underflow events (cont’d)
• 37% (159,713) of lost packets were discovered to be
missing after their deadlines had passed
• This effect was caused by large one-way delay jitter
• Additionally, one-way delay jitter caused 1,167,979 data
packets to be late for decoding
• Overall, 1,342,415 packets were late (1.7% of all sent
packets), out of which 98.9% were late due to large one-way
delay jitter rather than due to packet loss combined with
large RTT
• All late packets caused the “freeze-frame” effect for 10.5
seconds on average in D1p and 8.5 seconds in D2p (recall
that each session was 10 minutes long)
19
Underflow events (cont’d)
• 90% of late retransmissions missed the deadline by no more
than 5 seconds, 99% by no more than 10 seconds
• 90% of late data packets missed the deadline by no more
than 13 seconds, 99% by no more than 27 seconds
20
Round-trip Delay
• 660,439 RTT samples
• 75% of samples below 600 ms and 90% below 1 second
• Average RTT was 698 ms in D1p and 839 ms in D2p
• Maximum RTT was over 120 seconds
• Data-link retransmission combined with low-bitrate
connection were responsible for pathologically high RTTs
• However, we found access points with 6-7 second IP-level
buffering delays
• Generally, RTTs over 10 seconds were considered
pathological in our study
21
Round-trip Delay (cont’d)
• Distributions of the RTT in both datasets (PDF) were similar
and contained a very long tail
22
Round-trip Delay (cont’d)
• Distribution tails closely matched hyperbolic distributions
(Pareto with  between 1.16 and 1.58)
23
Round-trip Delay (cont’d)
• The average RTT varied during the day between 574 ms (3
am – 6 am) and 847 ms (3 pm – 6 pm) in D1p
• Between 723 ms and 951 ms in D2p
• Relatively small increase in the RTT during the day (by only
30-45%) compared to that in packet loss (by up to 300%)
• Per-state RTT varied between 539 ms (Maine) and 1,053
ms (Alaska); Hawaii and New Mexico also had average
RTTs above 1 second
• Little correlation between the RTT and geographical
distance of the state from NY
• However, much stronger positive correlation between the
number of hops and the average state RTT:  = 0.52
24
Delay Jitter
• Delay jitter is one-way delay variation
• Positive values of delay jitter are packet expansion events
• 97.5% of positive samples below 140 ms
• 99.9% under 1 second
• Highest sample 45 seconds
• Even though large delay jitter was not frequent, once a
single packet was delayed by the network, the following
packets were also delayed creating a “snowball” of late
packets
• Large delay jitter was very harmful during the experiment
25
Packet Reordering
• Average reordering rates were low, but noticeable
• 6.5% of missing packets (or 0.04% of sent) were reordered
• Out of 16,852 sessions, 1,599 (9.5%) experienced at least
one reordering event
• The highest reordering rate per ISP occurred in AT&T
WorldNet, where 35% of missing packets (0.2% of sent
packets) were reordered
• In the same set, almost half of the sessions (47%)
experienced at least one reordering event
• Earthlink had a session where 7.5% of sent packets were
reordered
26
Packet Reordering (cont’d)
• Reordering delay Dr is time between detecting a missing
packet and receiving the reordered packet
• 90% of samples Dr below 150 ms, 97% below 300 ms, 99%
below 500 ms, and the maximum sample was 20 seconds
27
Packet Reordering (cont’d)
• Reordering distance is the number of packets received
during the reordering delay (84.6% of the time a single
packet, 6.5% exactly 2 packets, 4.5% exactly 3 packets)
• TCP’s triple-ACK avoids 91.1% of redundant retransmits
and quadruple-ACK avoids 95.7%
28
Path Asymmetry
• Asymmetry detected by analyzing the TTL of the returned
packets during the initial traceroute
• Each router reset the TTL to a default value (such as 255)
when sending a “TTL expired” ICMP message
• If the number of forward and reverse hops was different, the
path was “definitely asymmetric”
• Otherwise, the path was “possibly (or probably) symmetric”
• No fail-proof way of establishing path symmetry using endto-end measurements (even using two traceroutes in
reverse directions)
29
Path Asymmetry (cont’d)
• 72% of sessions operated over definitely asymmetric paths
• Almost all paths with 14 or more end-to-end hops were
asymmetric
• Even the shortest paths (with as low as 6 hops) were prone
to asymmetry
• “Hot potato” routing is more likely to cause asymmetry in
longer paths, because they are more likely to cross AS
borders than shorter paths
• Longer paths also exhibited a higher reordering probability
than shorter paths
30
Conclusion
• Dialing success rates were quite low during the day (as low
as 40%)
• Retransmission worked very well even for delay-sensitive
traffic and high-latency end-to-end paths
• Both RTT and packet-loss bursts appeared to be heavytailed
• Our clients experienced huge end-to-end delays both due to
large IP buffers as well as persistent data-link
retransmission
• Reordering was frequent even given our low bitrates
• Most paths were in fact asymmetric, where longer paths
were more likely to be identified as asymmetric
31