The Utility of Characterizing Packet Loss as a Function of Packet Size in Commercial Routers

The Utility of Characterizing Packet Loss as a Function of Packet Size in

Commercial Routers

Jose Saldana

Julián Fernández-Navajas

José Ruiz-Mas

Eduardo Viruete Navarro

Luis Casadesus

Communication Technologies Group (GTC)

Aragon Institute of Engineering Research (I3A)

University of Zaragoza, Spain

INDEX

I. Introduction and Related Works

II. Tests and Results

III. Conclusions

CCNC 2012. Las Vegas, 15 Jan 2011

Introduction and Related Works

- The Internet was designed as a best-effort

network

- The first deployed services were e-mail, ftp, etc.,

which did not have critical delay requirements.

- So they used big packets in order to have a good

bandwidth efficiency.



- Nowadays, real-time services have very critical

delay requirements

- So they require a high rate of packets, which

implies small sizes



- Efficiency problem

One IPv4/TCP packet 1500 bytes

η=1460/1500=97%

One IPv4/UDP/RTP packet of VoIP with two samples of 10 bytes

η=20/60=33%



- Packet loss: one of the main impairments in

current networks

- Wireless: The radio link is the critical one

- Wired: Dropping in router queues

- TCP-based services can recover lost packets

- UDP-based services: interactive applications can

be seriously affected, as there is no time for

retransmission



- ¿Is there a relationship Packet loss vs. Packet

size?

- Wireless environment: YES. The bigger the packet,

the bigger the probability of a corrupted bit.

- Wired environment: The main cause of packet loss is

dropping in router queues, so we have to study the

behavior of the queues.

- Many routers were designed for “classical”

services, so they are not prepared to deal with

small packets



- Buffers of commercial routers have different

implementations

- Buffers can be measured in bytes, packets, or

maximum queuing delay

- Bandwidth is not the only limit: packets per second

limit is also present

- drop-tail, RED, etc

- The implementation may not strongly affect

delay and jitter, but may affect packet loss



- Characterization of packet loss vs packet size

packet loss

packet size



- Useful to make some decisions:

- Multiplexing or not

- Samples per packet

- Number of TS (Transport Stream) included in a packet

probability

packet size

One TS per packet



- Useful to make some decisions:

- Multiplexing or not

- Samples per packet

- Number of TS (Transport Stream) included in a packet

probability

packet size

Five TS per packet



What is better?

More TS per packet increases efficiency

But if it increases packet loss…

probability

packet size

Five TS per packet



- So

- Let’s study packet loss vs. packet size

behavior of the router (and the Internet, as

future work)

- Help when making decisions

- Give us an idea of the internal (and non-

publicly available) router implementation

- The interest is the shape of the graphs



“Sizogram”

packet loss

packet size

INDEX



III. Conclusions


Tests and Results

- Definition of a test able to reflect this

- An “empirical” test, suitable to be performed

with different routers and networks

- The router will be considered a “black box”

- We do not know its internal structure


Tests and Results

- Background traffic: 100% of the capacity

- 50 % packets 40 bytes

- 10% packets 576 bytes

- 40% packets 1,500 bytes


Tests and Results

0

2000

4000

6000

8000

10000

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

kb

ps

Packet size

Offered traffic

variable size

1500 bytes

576 bytes

40 bytes

Background

10Mbps


Tests and Results

0

2000

4000

6000

8000

10000

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

kb

ps

Packet size

Offered traffic

variable size

1500 bytes

576 bytes

40 bytes

Variable size

100kbps


Tests and Results

- Test traffic: two orders of magnitude smaller

0

2000

4000

6000

8000

10000

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

kb

ps

Packet size

Offered traffic

variable size

1500 bytes

576 bytes

40 bytes


Tests and Results

- Duration of each test:

- From 200 to 3,000 seconds for each point, in order to

have 25,000 packets of variable size traffic

- Matlab simulations


Tests and Results

- Two buffer implementations:

- byte-sized: 10kB, 20kB, 50kB, 100kB

- packet-sized: 16, 33, 83, 166 packets

- Can be considered “tiny buffers”


Tests and Results: Byte-sized

0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

10%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)

packet loss, buffer byte-sized

10kB

20kB

50kB

100kB



0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

10%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10kB

20kB

50kB

100kB

Monotonically

increasing



0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

10%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10kB

20kB

50kB

100kB

Bigger packet –

Bigger packet loss



0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

10%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10kB

20kB

50kB

100kB

The slope grows as the

buffer gets smaller


0%

1%

2%

3%

4%

5%

6%

7%

8%

9%

10%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)

packet loss, buffer packet-sized

16packets

33packets

83packets

166packets

Tests and Results: Packet-sized

No variation with packet

size, as expected


Tests and Results: pps limit

0%

5%

10%

15%

20%

25%

30%

35%

40%

45%

50%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10 kb

10kB, 2000pps

10kB, 2500pps

10kB, 3000 pps

10kB, 3500pps


0%

5%

10%

15%

20%

25%

30%

35%

40%

45%

50%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10 kb

10kB, 2000pps

10kB, 2500pps

10kB, 3000 pps

10kB, 3500pps


Offered 1,970 pps


0%

5%

10%

15%

20%

25%

30%

35%

40%

45%

50%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10 kb

10kB, 2000pps

10kB, 2500pps

10kB, 3000 pps

10kB, 3500pps


Increases significantly

with respect to no pps

limitation


0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10kB

10kB, tproc=20us

10kB, tproc=40us

10kB, tproc=60us

10kB, tproc=80us

10kB, tproc=100us

Tests and Results: processing time


0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10kB

10kB, tproc=20us

10kB, tproc=40us

10kB, tproc=60us

10kB, tproc=80us

10kB, tproc=100us


Graphs have the same

shape


0%

2%

4%

6%

8%

10%

12%

14%

16%

18%

20%

100 200 300 400 500 600 700 800 900 1000 1100 1200 1300 1400 1500

packet

loss

packet size (bytes)


10kB

10kB, tproc=20us

10kB, tproc=40us

10kB, tproc=60us

10kB, tproc=80us

10kB, tproc=100us


Packet loss increase is

not so severe

INDEX



III. Conclusions


Conclusions

- We have studied the relationship between packet

size and packet loss in routers

- Depends on buffer implementation

- A methodology has been defined to find the

packet loss histogram as a function of packet

size

- This can give us an idea of the internal

implementation of the router, and help us when

making some decisions


Conclusions

- Byte-sized buffers increase loss with size

- Packet-sized buffers do not vary

- Extension of this to Internet paths and

commercial routers

Thank you

[email protected]

The Utility of Characterizing Packet Loss as a Function of Packet Size in Commercial Routers

Education

Transcript of The Utility of Characterizing Packet Loss as a Function of Packet Size in Commercial Routers