rice university, spin group spin.rice.edu multiscale traffic processing techniques for network...
TRANSCRIPT
Rice University, SPiN Group spin.rice.edu
Multiscale Traffic Processing Techniques for Network Inference and Control
R. Baraniuk R. Nowak E. Knightly R. Riedi
X. Wang V. Ribeiro S. Sarvotham Y. Tsang
NMS PI meeting Baltimore April 2002
Signal Processing in Networking
Rice University, SPiN Group spin.rice.edu
Chirp Probing
Effort 1
Rice University, SPiN Group spin.rice.edu
Objective: Reduced complexity, multiscale
link models with known accuracy
Innovative Ideas Multifractal analysis Multiplicative modeling Multiscale queuing Chirps for probing
Impact Congestion control
Workload balancing at servers Dynamical streaming Pricing on connection basis
Rice University, SPiN Group spin.rice.edu
Chirp Probe Cross-Traffic Inference
Rice University, SPiN Group spin.rice.edu
New Ideas
Probing multiple hopsChirp-Sandwiches to probe routers
“down the path”IP-tunneling
Monitor packets at intermediate point of pathVerification for Chirping and Sandwiches
Network calculus (max-plus algebra)
Probing buffer at core router Passive inference
Rice University, SPiN Group spin.rice.edu
Tech Transfer
- Stanford (SLAC) Chirps in PingER as freeware for monitoring- C+ code for sending Chirp Probes- 10 Msec precision- Visualization tools for link properties
IP tunneling- Infrastructure at Rice ready
- CAIDA (chirping as a monitoring tool)
- UC Riverside (Demystify Self-similarity)
- Sprint Labs (queue-sizes at core routers)
Rice University, SPiN Group spin.rice.edu
Connection-level Analysis and
Modeling of Network Traffic
Effort 2
Rice University, SPiN Group spin.rice.edu
Aggregate Traffic at small scales
• Trace:– Time stamped headers
• Large scales– Gaussian – LRD (high variability)
• Small scales– Non-Gaussian – Positive process– Burstiness
Objective :– Origins of bursts
Auckland Gateway (2000)Aggregate Bytes per time
Gaussian: 1%Real traffic: 3%
Kurtosis - Gaussian : 3 - Real traffic: 5.8
Mean
99%
0 0.5 1 1.5 2 2.5 3
x 105
0
50
100
150
200
250
300
350
400
450histogram
Gaussian: 1%Real traffic: 3%
0 2000 4000 60000
0.5
1
1.5
2
2.5
3x 10
5
time (1 unit=500ms)nu
mbe
r of
byt
es
Mean
99%
Rice University, SPiN Group spin.rice.edu
• ON/OFF model – Superposition of sources– Connection level model
• Explains large scale variability: – LRD, Gaussian– Cause: Costumers – Heavy tailed file sizes !!
Bursts in the ON/OFF framework
• Small scale bursts:– Non-Gaussianity– Conspiracy of sources ??– Flash crowds ??
(dramatic increase of active sources)
Rice University, SPiN Group spin.rice.edu
Non-Gaussianity: A Conspiracy?
• The number of active connections is close to Gaussian; provides no indication of bursts in the load
• Indication for:- No conspiracy of sources
- No flash crowds
Load: Bytes per 500 ms
Number of active connections
Mean
99%
Mean
99%
Rice University, SPiN Group spin.rice.edu
Non-Gaussianity: a case study
Typical bursty arrival(500 ms time slot)
Typical Gaussian arrival(500 ms time slot)
Histogram of load offered in same time bin per connection:Considerable balanced “field” of connections
10 Kb
Histogram of load offered in same time bin per connection:One connection dominates
150 Kb
15 Kb
Rice University, SPiN Group spin.rice.edu
Non-Gaussianity and Dominance
Systematic study: time series separation• For each bin of 500 ms:
remove packets of the ONE strongest connection• Leaves “Gaussian” residual traffic
Overall traffic Residual traffic1 Strongest connection
= +Mean
99%
Rice University, SPiN Group spin.rice.edu
Separation on Connection Level
Definition:• Alpha connections:
Peak rate > mean arrival rate + 1 std dev
• Beta connections: Residual traffic
• C+ analyzer (order of sec)• Separation is more efficient (less accurate)
using multi-scale analysis
Rice University, SPiN Group spin.rice.edu
Beta Traffic Component
• Constitutes main load• Governs LRD properties of overall traffic• Is Gaussian at sufficient utilization (Kurtosis = 3)
• Is well matched by ON/OFF model
Beta trafficVariance time plot
Mean
99%
Traffic percentile: 1%
Recall: beta = non-dominatingRecall: beta = non-dominating
0 0.5 1 1.5 2 2.5
x 105
0
50
100
150
200
250
300
350
400
Histogram of Beta traffic
Beta Traffic Kurtosis 3.4
Number of connections = ON/OFF
Mean
99%
Rice University, SPiN Group spin.rice.edu
Origins of Alpha Traffic
• Mostly TCP (flow & congestion controlled).• HTTP, SMTP, NNTP, etc.• Alpha connections tend to have the same source-
destination IP addresses.• Any large volume connection with the same e2e IP
addresses as an alpha connection is also alpha.
Systematic cause of alpha:Large file transfers over high bandwidth e2e paths.
Rice University, SPiN Group spin.rice.edu
Simple Connection Taxonomy
This is the only systematic reason
Bursts arise from largetransfers over fast links.
.RTTcwnd
bandwidth .RTTcwnd
bandwidth
Rice University, SPiN Group spin.rice.edu
Cwnd or RTT?
Correlation coefficient=0.68
RTT has strong influence on bandwidth and dominance.
103
104
105
106
10-1
100
101
102
peak-rate (Bps)
1/R
TT
(1/
s)
Correlation coefficient=0.01
103
104
105
106
102
103
104
105
peak-rate (Bps)
cwnd
(B
)
Colorado State University trace, 300,000 packets
cwnd 1/RTT
ratepeak cwnd
1/RTTratepeak
Beta Alpha Beta Alpha
Rice University, SPiN Group spin.rice.edu
Examples of Alpha/Beta Connections
0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4
-200
0
200
400
600
800
1000
1200
1400
one beta connection
packet arrival time (second)
pack
et s
ize (
byte
s)
forward direction
reverse direction
41.3 41.35 41.4 41.45
-200
0
200
400
600
800
1000
1200
1400
one alpha connection (96078, 196, 80, 59486)
packet arrival time (second)
pack
et s
ize
(byt
es)
forward direction
reverse direction
Alpha connections burst because of short round trip time, not large TCP window
Modeling of Alpha Traffic• ON/OFF model revisited:
High variability in connection rates (RTTs)
Low rate = beta High rate = alpha
fractional Gaussian noise stable Levy noise
+
=
+
+
=
Modeling of Alpha Traffic• ON/OFF model revisited:
High variability in connection rates (RTTs)
Low rate = beta High rate = alpha
fractional Gaussian noise stable Levy noise
0 2000 4000 60000
0.5
1
1.5
2
x 105
time (1 unit=500ms)
num
ber
of b
ytes
0 2000 4000 60000
0.5
1
1.5
2
2.5
3x 10
5
time (1 unit=500ms)
num
ber
of b
ytes
Rice University, SPiN Group spin.rice.edu
Rice University, SPiN Group spin.rice.edu
Impact: Simulation
• Simulation: ns topology to include alpha links
Simple: equal bandwidth Realistic: heterogeneous end-to-end bandwidth
• Congestion control• Design and management
Rice University, SPiN Group spin.rice.edu
Impact: Performance
• Beta Traffic rules the small Queues• Alpha Traffic causes the large Queue-sizes
(despite small Window Size)
All Alpha Packets Queue-size with Alpha Peaks
Rice University, SPiN Group spin.rice.edu
Multiscale Nature of Traffic
• LRD– Large time scales – approx. Gaussian – Client behavior– Bandwidth over Buffer
• Multifractal– Small time scale– Mixture Gaussian - Stable– Network topology– Control at Connection level
Structure: Multiplicative Additive
packetscheduling
sessionlifetime
networkmanagement
round-triptime
< 1 msec msec-sec minutes hours
Rice University, SPiN Group spin.rice.edu
Ongoing Work
• Versatile efficient additive/multiplicative models (fast synthesis, meaning of model parameters)
• Impact of Alpha traffic on Performance• Goal: Obtain performance parameters
directly from network and user specifications
• Monitoring tool for network load (freeware) using chirps (Caida, SLAC)
• Verification/”internal” measurements using IP-tunneling
Rice University, SPiN Group spin.rice.edu
Summary• Alpha traffic (peak rate > burst threshold)
– Few connections. Responsible for bursts– Origin: Large transfer over high bandwidth paths– High bandwidth from short RTT
• Beta traffic (residual): – Main load. Responsible for LRD– Origin: Crowd with limited bandwidth– Gaussian at sufficient utilization
• Connection level analysis: parameter estimation drastically simplifies using multi-scale analysis
• Statistical modeling: requires novel multi-scale models (mixtures of additive and multiplicative trees)