An Information-theoretic Approach to Network Measurement

Download Report

Transcript An Information-theoretic Approach to Network Measurement

An Information-theoretic Approach to
Network Measurement and Monitoring
Yong Liu, Don Towsley, Tao Ye, Jean Bolot
1
Outline




motivation
background
flow-based network model
full packet trace compression
 marginal/joint
 coarser granularity
 netflow and SNMP
 future work
2
Motivation
 network monitoring: sensing a network
 traffic engineering, anomaly detection, …
 single point v.s. distributed
 different granularities
 full traffic trace: packet headers
 flow level record: timing, volume
 summary statistics:
byte/packet counts
 challenges
 growing scales:
high speed link, large topology
 constrained resources:
processing, storage, transmission
 30G headers/hour at UMass gateway
 solutions
 sampling: temporal/spatial
 compression: marginal/distributed
3
Questions
 how much can we compress monitoring traces?
 how much information is captured by different
monitoring granularity?
 packet trace/NetFlow/SNMP
 how much joint information is there in multiple
monitors?
 joint compression
 trace aggregation
 monitor placement
4
Our Contribution
 flow-based network models
 explore temporal/spatial correlation in network traces
 projection to different granularity
 information theoretic framework
 entropy: bound/guideline on trace compression
 quantitative approach for more general problems
 validation against measurement from operational
network
5
Entropy & Compression
 Shannon entropy of discrete r.v.
 compression of i.i.d. symbols (length M) by coding
 coding:
 expected code length:
 info. theoretic bound on compression ratio:
 Shannon/Huffman coding
 assign short codeword to frequent outcome
 achieve the H(X) bound
6
Entropy & Correlation
 joint entropy
 entropy rate of stochastic process

exploit temporal correlation

Lempel-Ziv Coding: (LZ77, gzip, winzip)
asymptotically achieve the bound for stationary process
 joint entropy rate of correlated processes
7

exploit spatial correlation

Slepian-Wolf Coding: (distributed compression)
encode each process individually, achieve joint entropy rate in limit
Network Trace Compression
 naïve way: treat as byte stream, compress by generic tools

gzip compress UMass traces by a factor of 2
 network traces are highly structured data

multiple fields per packet
• diversity in information richness
• correlation among fields


multiple packets per flow
•
•
packets within a flow share information
temporal correlation
•
•
most fields unchanged within the network
spatial correlation
multiple monitors traversed by a flow
 network models
 explore correlation structure
 quantify information content of network traces
 serves as lower bounds/guidelines for compression algorithms
8
Packet Header Trace
0
16
time stamp (sec.)
Timing
time stamp (sub-sec.)
vers. HLen
ToS
IPID
IP Header
31
TTL
total length
flags
protocol
fragment offset
header checksum
source IP address
destination IP address
source port
destination port
data sequence number
acknowledgment number
TCP Header
Hlen
TCP flags
checksum
9
window size
urgent pointer
Header Field Entropy
0
16
time stamp (sec.)
Timing
time stamp (sub-sec.)
vers. HLen
ToS
IPID
IP Header
31
TTL
total length
flags
protocol
time
fragment offset
header checksum
source IP address
destination IP address
source port
destination port
data sequence number
acknowledgment number
TCP Header
Hlen
TCP flags
checksum
10
window size
urgent pointer
flow id
Single Point Packet Trace
T0
F0
T1
F1
T3
Tm
F0
F0
Tn
Fn
 packet inter-arrival:
# bits per packet:
 temporal correlation introduced by flows
 packets from same flow closely spaced in time
 they share header information
 flow based trace:
T0
F0
 flow record:
F0
K
T3
Tm
T0
flow flow arrival
ID size time
11
F0
packet inter-arrival
F0
Network Models
flow-based model
 flow arrivals follow Poisson with rate
 flows are classified to independent flow classes
according to routing (the set of routers traversed)
 flow i is described by:
• flow inter-arrival time:
• flow ID:
• flow length:
• packet inter-arrival time within the flow:
 packet arrival stochastic process:
12
Entropy in Flow Record
 # bits per flow:
 # bits per second:
 marginal compression ratio
 determined by flow length (pkts.) and
variability in pkt. inter-arrival.
13
Single Point Compression: Results
C1-in
BB1-out
BB2-out
C2-in
router
Trace
H (total)
Model
Ratio
Compression
Algorithm
C1-in
706.3772
0.2002
0.6425
BB1-out
736.1722
0.2139
0.6574
BB2-out
689.9066
0.2186
0.6657
 Compression ratio lower bound calculated by entropy much lower
than real compression algorithm
 Real compression algorithm difference
 Records IPID, packet size, TCP/UDP fields
 Fixed packet buffer for each flow => many flow records for long flows
14
Distributed Network Monitoring
 single flow recorded by multiple
monitors
 spatial correlation:
traces collected at distributed
monitors are correlated
 marginal node view:
#bits/sec to represent flows seen
by one node, bound on single point
compression
 network system view:
#bits/sec to represent flows
cross the network, bound on joint
compression
 joint compression ratio:
quantify gain of joint compression
15
Baseline Joint Entropy Model
 “perfect” network
 fixed routes/constant link delay/no packet loss
 flow classes based on routes
 flows arrive with rate:
 # of monitors traversed:
 #bits per flow record:
 info. rate at node v:
 network view info. rate:
 joint compression ratio:
 dependence on # of monitors travered
16
Joint Compression: Results
C1-in
C2-in
17
BB1-out
BB2-out
router
Set of Traces
Joint Compression Ratio
{C1-in, BB1-out, C2-in, BB2-out}
0.5
{C1-in, BB1-out}
0.8649
{C1-in, BB2-out}
0.8702
{C2-in, BB1-out}
0.7125
{C2-in, BB2-out}
0.6679
Coarser Granularity Models
 NetFlow model
 similar to flow model:
 joint compression result similar to full trace
 SNMP model
 any link SNMP rate process is sum of rate processes of all
flow classes passing through that link
 traffic rates of flow classes are independent Gaussian
 entropy can be calculated by covariance of these processes
 information loss due to summation
 small joint information between monitors
 difficult to recover rates of flow classes from SNMP data
18
Joint Compression Ratio of Different
Granularity
C1-in
C2-in
19
BB1-out
BB2-out
router
Set of Traces
SNMP
NetFlow
Packet Trace
{C1-in, BB1-out}
1.0021
0.8597
0.8649
{C1-in, BB2-out}
0.9997
0.8782
0.8702
Conclusion
 information theoretic bound on marginal
compression ratio -- ~ 20% (time+flow id,
even lower if include other low entropy fields)
 marginal compression ratio high (not very
compressible) in SNMP, lower in NetFlow, and
the lowest in full trace
 joint coding is much more useful/nessassary
in full trace case than in SNMP
 “More entropy for your buck”
20
Future Work
 network impairments
 how many more bits for delay/loss/route change




model netflow with sampling
distributed compression algorithms
lossless v.s. lossy compression
entropy based monitor placement
 maximize information under constraints
21
Thanks!
22