Get documents about "Infocom07 fad
Shared by: HC120912051825
-
Stats
- views:
- 2
- posted:
- 9/11/2012
- language:
- Unknown
- pages:
- 37
Document Sample
A Suite of Schemes for User-level
Network Diagnosis without Infrastructure
Yao Zhao, Yan Chen
Lab for Internet and Security Technology,
Northwestern University
1
Motivation
• How do end users, with no special
privileges, identify packet loss
inside the network with one or two
computers?
2
Motivation
• How do end users, with no special
privileges, identify packet loss
inside the network with one or two
computers?
• Take-home
– We propose three user-level loss rate
diagnosis approaches
– The combo of our approaches and
Tulip [SOSP03] is much better than
any single approach
3
Outline
• Motivation
• Related Works
• Lossy Link Diagnosis
– Fragmentation Aided Diagnosis (FAD)
• Algebraic FAD
• Opportunistic FAD
– Striped Probe Analysis (SPA)
• Evaluations
• Conclusions
4
Related Work I
• Internet Tomography
– Multicast based (not practical)
– Unicast based
• Mimic multicast S
The more
cooperating end
hosts, the shorter the
virtual links
L1 L2 L3 L4
5
Related Work II
• Tulip [SOSP03]
– Leverage on consecutive IPID
– Tend to underestimate forward loss
rates
• Suffer from the packet loss correlation
S D S D S D
x x
x x
Forward Loss Reverse Loss ?
6
Outline
• Motivation
• Related Works
• Lossy Link Diagnosis
– Fragmentation Aided Diagnosis (FAD)
• Algebraic FAD
• Opportunistic FAD
– Striped Probe Analysis (SPA)
• Evaluations
• Conclusions
7
Link Diagnosis=> Forward Path Diagnosis
• If we can infer the loss rates of
forward path F1 and F2, we can
infer the link loss rate of l3
F1
l1 l2 l3 D
S R1 R2
F2
• The more diagnosable forward path
segments, the better the diagnosis
granularity
8
Basic Idea of FAD
P
S P R N
R
P1
P2
P2
S P1 R N
R
9
Algebraic FAD
• Let pf and pr be the loss rate of the
forward and reverse path respectively
P
(1 - pf)×(1 - pr)=1 – p (1)
R
P1
P2
(1 - pf)2×(1 - pr)=1 – p’ (2)
R
p and p’ are measurable. Solve pf and pr
using (1) and (2)
10
How to Achieve FAD
IP Fragmentation
– Fragment a packet longer than MTU
– Required to be supported in IPv4
– Some routers disable it for security reason
• Support of IP Fragmentation
– 64,320 router IP addresses probed by using
Traceroute
– About 80% of routers support IP fragmentation
• Degree of Rate Limiting on Responses
– 99% of routers allow a rate of 100 probes/s for
ICMP Echo, ICMP Timestamp and TCP probes
– Response to UDP probe is severely rate-
limited
11
Opportunistic FAD
F1 + F2 P
aaaaaaaa bbbbbbbb aaaaaaaabbbbbb
F1 + F’2 P’
aaaaaaaa ccccccccc aaaaaaaacccccccc
12
Opportunistic FAD
No Loss Forward Loss
S N S N
x
F1+F2
F1+F’2
Similar to Tulip, but OFAD allows
large gap between fragments
13
Striped Probe Analysis (SPA)
S
p1 p1
p2
S R D R
p2 p3
p3 D S
• S sends a probe to D and we get the path p1->p2
• S sends UDP packet with a certain TTL so that R
returns an ICMP TTL-Exceeded response. Hence
we get path p1->p3
14
Striped Probe Analysis (SPA)
S
P1 P2
p1
R
p2 p3
D S
(1) Loss on shared link
15
Striped Probe Analysis (SPA)
S S
P1 P2
p1 p1
R R
p2 p3 p2 p3
D S D S
(1) Loss on shared link (2) Loss on non-shared link
• Success rate of p1≈n1×n2 / (n×n12)
– n: number of striped probes sent,
– n1: number of P1 received by D,
– n2: number of P2 received by S,
– n12: number of cases that both P1 and P2 are received
• Unbiased if packet loss has perfect correlation and loss rates of
different links are independent 16
Summary
Requirement Accuracy
Tulip Inaccurate w/ strong loss
Consecutive IPID (70%)
[SOSP03] correlation
FAD (AFAD Accurate w/ weak or short loss
IP fragmentation (80%)
& OFAD) correlation
ICMP TTL-Exceeded.
Accurate w/ strong loss
SPA Access from both end
correlation
hosts
The current Internet usually has strong but short loss correlation.
17
Outline
• Motivation
• Related Works
• Lossy Link Diagnosis
– FAD
– SPA
• Evaluations
• Conclusions
18
Evaluation Metrics
• Diagnosis Granularity
– Weighted average of the lengths of the
path’s diagnosable segments
– For example, an 8-hop path has two
diagnosable segments of length 3 and 5,
and then the granularity of the path is
(32 + 52)/8 = 4.25
• Accuracy
– Estimation error:
– Relative error:
19
Diagnosis Granularity
SPA is
best
FAD ≈Tulip
Combo of
FAD and Tulip
is better
20
Path-Level Accuracy Evaluation
FAD > Tulip >
SPA
OFAD, Tulip and
SPA tends to
underestimate
loss rates
21
More Evaluations
• Consistency Check
• Packet Probe Size Selection
• Lossy Link Distribution
• More in the technical report
http://www.cs.northwestern.edu/~yzh734/
22
Conclusions and Recommendations
• We propose AFAD, OFAD and SPA which
can conduct loss rate diagnosis without
infrastructure
• Tulip, FAD and SPA have different working
scenarios
– The combination of them can achieve low
diagnosis granularity and high accuracy
• Recommendations
– OFAD+SPA, if we can control the two ends of
an end-to-end path
– OFAD+Tulip, if we can only control the source
23
24
Thanks!
Questions?
25
Path-Level Accuracy of Combined Schemes
26
27
Path-Level Accuracy Evaluation
28
Path-Level Accuracy of Combined Schemes
29
IP Fragmentation Is Widely Supported
• Router Collection
– 64,320 router IP addresses probed by using traceroute from a
machine
• Support of Different Probes
Echo Timestamp UDP TCP Any
1 source 85.3% 69.2% 64.5% 71.7% 88.2%
10 sources 87.3% 72.3% 70.7% 73.3% 90.1%
• Support of IP Fragmentation
– 90.3% of responsive routers support IP fragmentation
– Altogether about 80% of routers support FAD.
• Degree of Rate Limiting on Responses
– 99% of routers allow a rate of 100 probes/s for ICMP Echo,
ICMP Timestamp and TCP probes
– UDP probe is severely rate-limited
30
Packet Transmission Correlation
• Choose 100 PlanetLab hosts and
randomly measure 5000 paths
• Little loss correlation with enough gap
31
Forward Path Diagnosis => Link Diagnosis
• If we can infer the loss rates of
forwarding path l1 and P1, we can
infer the link loss rate of l2 too.
D
32
Opportunistic FAD
P1 P1
P2 P2
x
R12
P’2 P’2
R’12
(1) (2)
• n: number of R12 received, n’: number of
R’12 received
• Xi = 0 when forward packet i is lost and
Xi =1 otherwise
• P(X2=1)≈P(X2=1|X1=1)≈n/(n+n’)
33
Striped Probe Analysis (SPA)
S
l1 l2
l1
S R D
R
l3 l2 l3
D S
• No fragmented packets needed !
• S sends a probe to D and we get the path l1->l2
• S sends UDP packet with a certain TTL so that R
returns an ICMP TTL-Exceeded response. Hence
we get path l1->l3
34
Striped Probe Analysis (SPA)
S
P1 P2
l1
R
l2 l3
D S
35
Striped Probe Analysis (SPA)
S S
P1 P2
l1 l1
R R
l2 l3 l2 l3
P1 P2
D S D S
(1) No loss (2) Loss on shared link
36
Striped Probe Analysis (SPA)
S S S
P1 P2
l1 l1 l1
R R R
l2 l3 l2 l3 l2 l3
P1 P2
D S D S D S
(1) No loss (2) Loss on shared link (3) Loss on non-shared link
• Success rate of l1≈n1×n2 / (n×n12)
– n: number of striped probes sent,
– n1: number of P1 received by D,
– n2: number of P2 received by S,
– n12: number of cases that both P1 and P2 are received
• Unbiased if packet loss has perfect correlation and loss rates of different
links are independent 37
