on the processing time for detection of skype traffic p.m. santiago del río, j. ramos, j.l....

25
On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A. Cuadra-Sánchez, M. Cutanda-Rodríguez Indra Sistemas International Workshop on TRaffic Analysis and Classification TRAC 2011 Istanbul July, 7th 2011

Upload: jade-ramsey

Post on 01-Jan-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

On the processing time for detection of Skype traffic

P.M. Santiago del Río, J. Ramos,

J.L. García-Dorado, J. AracilUniversidad Autónoma de Madrid

A. Cuadra-Sánchez, M. Cutanda-RodríguezIndra Sistemas

International Workshop on TRaffic Analysis and Classification TRAC 2011Istanbul July, 7th 2011

Page 2: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

2

Contents Introduction. Skypeness. Performance Evaluation. Conclusion and Future work.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 3: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

3

Contents Introduction.

• Detecting Skype.• Motivation.

Skypeness. Performance Evaluation. Conclusion and Future work.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 4: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

4

How to detect Skype traffic? Skype:

• Proprietary, obfuscated and encrypted protocol.• Random ports.• UDP packets (TCP only used in presence of UDP-restricted

firewalls).

Traffic Classification Techniques:• Port numbers:

– Not valid (random port numbers).• DPI:

– Computationally expensive (payload inspection).– Insufficient accuracy (encrypted content).

• Machine-learning:– Sufficient accuracy.– Sufficient performance (from a computational viewpoint)

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 5: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

5

Motivation Why detect Skype traffic?

• Popularity:– Users (increasing thanks to mobile smartphones).– Profits.

• Security reasons:– Lawful interception.– e.g. EU directive 2006/24/EC.

• Quality of Service (QoS).• Billing/accounting and restrictions.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 6: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

6

Motivation Why so fast?

• Increase in data transmissions speeds:– 10 Gb/s (and even faster…).– A lot of sorts of traffic.

• Current traffic classifications applications:– Accuracy: – Processing speed:

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 7: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

7

Motivation Why in commodity hardware?

• Alternative to specialized hardware (e.g. FPGAs):– Less flexibility.– More cost.

• Low-cost integration and minimal interference with current infrastructure.

• Good performance and limited cost.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 8: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

8

Contents Motivation. Skypeness.

• Detector Fundamentals• HW & SW Architecture

Performance Evaluation. Conclusion and Future work.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 9: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

9

Skypeness Goals:

1. Accurate

2. Fast

3. Cheap

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 10: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

10

Detector Fundamentals Based on Tstat Skype classifier:

• Packet length.• Interrarival times.• Bit rate.• Chi Square Test (packet payload inspection required)

UDP packets.• TCP only used in presence of UDP-restricted firewalls.

Smooth data (moving average).

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 11: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

11

Detector Fundamentals Detection algorithm:

• Proportion of packet windows whose mean packet size/inter-arrival/bit-rate are inside the valid intervals.

• If these proportions are greater than the threshold, the flow is marked as Skype.

• Ground-truth obtained in several scenarios (wired and wireless, real and emulated networks conditions, etc).

• Thresholds values optimized with C4.5 trees.On the processing time for detection of Skype traffic

TRAC 2011 Istanbul, July 7th

Characteristic Interval Threshold

Packet size [Bytes] [30,200] 0.75

Interarrival [ms] [in-1±15] 0.6

Bit rate [Kb/s] [0,150] 0.75

Page 12: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

12

Software Architecture 3 modules:

• Capturing and parsing packets.• Creating and updating flows.• Detecting Skype.

Modular architecture:• Provides flexibility.• Makes possible adding other tests.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 13: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

13

Software Architecture Capturing module details:

• Raw socket and mmap functions to map NIC receive queues at user-space.

• Parse IP/TCP-UDP headers, obtaining:– 5-tuple (IP src/dst, port src/dst, transport protocol).– Packet size– Arrival timestamp

Flow-builder module details:• Unidirectional flows.• Hash-based table.• Pre-allocated memory pool.• Sorted.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 14: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

14

Hardware Architecture Commodity hardware:

• 4 AMD Opteron processors.• 32x4 GB DDR3 memory.• Intel 10 Gigabit NIC:

– Multiqueue Tx and Rx (up to 16 per interface).

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 15: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

15

Hardware Architecture NUMA: Non Uniform Memory Access

• Memory split in several groups.• The lower the distance, the higher the performance.

Ubuntu Server 10.04 64 bits, 2.6.35 kernel

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 16: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

16

Contents Motivation. Skypeness. Performance Evaluation.

• Accuracy results.• Processing performance results.

Conclusion and Future work.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 17: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

17

Accuracy results Using 3 traces:

• T1 and T2: from 96-hour of only Skype traffic (PoliTo).• T3: Synthetic P2P (and no Skype) traffic (UAM).

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Trace Skype Other FP(%) FN(%)

T1Bytes ~8GB 0 - 0.41

Packets ~40M 0 - 0.79

Flows ~1K 0 - 11.33

T2Bytes ~230M 0 - 5.88

Packets ~3M 0 - 6

Flows ~160 0 - 6.29

T3Bytes 0 ~1GB 0 -

Packets 0 ~5.3K 0 -

Flows 0 ~50 0 -

Page 18: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

18

Processing performance results Receiving real traffic:

• Variable rate (100 Mb/s … 1 Gb/s).• Tcpreplay limitation to 1 Gb/s.• 1 Rx queue and 1 Skypeness instance:

– Only using 2 cores: 1 for receiving and 1 for detecting.– Located in NUMA nodes 1 and 4 respectively (worst case in

terms of distance).

• Using Trace 4: – 3G real traffic.– ~70M packets.– ~12M TCP/UDP flows.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 19: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

19

Processing performance results Receiving real traffic:

• No packet loss at 1 Gb/s.

But… how about 10 Gb/s?

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Bit rate [Mbps]

Packet rate [Kpps]

Max. Flow rate per second

Total Packet Loss Rate

100 30 ~26K 0

250 75 ~53K 0

500 150 ~90K 0

750 225 ~120K 0

1000 300 ~170K 0

Page 20: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

20

Processing performance results Assigning a Rx queue per socket:

• Up to16 Rx queues and 16 detecting processes.

Offline experiments:• Read from memory (instead of

from the NIC). – Bandwidth of DDR3: 170.6 Gb/s

» Much greater than a backbone link.

• Skypeness instances and traces are located in memory as far as possible in terms of NUMA distance (worst case).

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 21: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

21

Processing performance results 3.7 Gb/s a single instance. 45 Gb/s using 16 instances. No perfect scaling due to serialized access to

shared memory by NUMA.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 22: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

22

Contents Motivation. Skypeness. Performance Evaluation. Conclusion and Future work.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 23: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

23

Conclusion Skypeness detects Skype traffic:

• With sufficient accuracy:– False negative rate of 6%.– False positive rate of zero.

• At high speed:– 1Gb/s and up to 3.4 Gb/s per instance (from NIC and

memory respectively).– Total throughput of 45 Gb/s with 16 instances.

• Using commodity hardware.

Feasibility of Skype traffic detection at high-speed networks (10-40 Gb/s) using commodity hardware.

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 24: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

24

Future (and present) work Capturing packets from a 10 Gb/s NIC:

• Using PacketShader capture engine.

Online experiments at 10 Gb/s Apply methodology to

• other classes of traffic:– RTP (done!)– P2P

• other classification techniques:– DPI

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th

Page 25: On the processing time for detection of Skype traffic P.M. Santiago del Río, J. Ramos, J.L. García-Dorado, J. Aracil Universidad Autónoma de Madrid A

Thank you for your attention!

Questions?

On the processing time for detection of Skype trafficTRAC 2011 Istanbul, July 7th