Date post: | 20-May-2018 |
Category: |
Documents |
Upload: | hoangthuan |
View: | 227 times |
Download: | 2 times |
On the Effects ofPre-computed Route Caching
inMultiple Peer Group
ATM-PNNI Networks
byKamalesh S Kalarickal
Department of Electrical Engineering and Computer Science 2
Outline of the Presentation
• Introduction• Pre-computed Routing Strategy• Mechanics of Experimentation• Preliminary Results• Pre-computation for Multiple Peer Groups• Route Cache Update Heuristics• Conclusion
Department of Electrical Engineering and Computer Science 3
Overview of PNNI Routing• PNNI hierarchical
addressing.• Topology database of
network resources.• Link state flooding via
PTSEs.• PNNI uses source routing.• Route re-computation at
border node per peer group.
• Crankback on failures.
A.1A.2
Src UNI
Dst UNI
1
2
4 12
3
DTL A.1.1,A.1.2,A.1.3,A.1.4,A.2
DTLA.2.1,A.2.2
Level 88
Level 96
Department of Electrical Engineering and Computer Science 4
Stages in On-Demand Routing1. Representation of network as a graph.2. Generic Call Admission Control to prune
the graph based on QoS requirements of the call.
3. Dijkstra’s Single Shortest Path Algorithm, minimizing a selected routing constraint.
4. Designated Transit List (DTL) is created, containing route to the destination.
Department of Electrical Engineering and Computer Science 5
Problem Statement
• On-Demand Routing in ATM PNNI networks involves expensive computation.
• This is carried out for every incoming call.• It is also redone at the ingress node of every
other peer group in the path.• This results in increased call setup time.
Department of Electrical Engineering and Computer Science 6
Proposed Solution• Pre-Compute routes to destinations in
advance and store in a route cache.• When a call arrives at the source node, the
cache is searched to find an appropriate route.
• If a route is found, it is used.• Otherwise on-demand routing is carried out
and the route is inserted into the route cache.
Department of Electrical Engineering and Computer Science 7
Pre-Computed Routing• Reduction in time for route discovery as
compared to on-demand routing, results in lower setup time.
• But this occurs at the expense of the quality of routes due to aging of route cache entries.
• Hence update the route cache to reflect changed network state.
Department of Electrical Engineering and Computer Science 8
Comparison
On-Demand Routing.Advantages:• More up to date
information for routing decision. Hence route failures are lower.
Disadvantage:• Increased average call
setup time due to overhead of route computation for each call.
Pre-Computed Routing.Advantage:• Lower average routing
time leads to lower average call setup time.
Disadvantage:• Decreased call acceptance,
due to aging of information used to pre-compute routes.
Pre-computed Routing Strategy
Implementation and preliminary results
Department of Electrical Engineering and Computer Science 10
Stages in Pre-Computed Routing
1. Representation of network as a graph.2. Max-Min B/W range is divided into equivalence
classes of bandwidths.3. For each quantized QoS level, do GCAC.4. For every unique node in the database, run
Dijkstra’s Single Shortest Path Algorithm, and compute route.
5. DTL, if obtained, is stored along with the effective bandwidth of the route, in the cache.
Department of Electrical Engineering and Computer Science 11
Route Cache Policy
• Routes are pre-computed and stored in the cache during initialization and also during a route cache update.
• Pre-computed route processing is assumed to occur on a dedicated co-processor.
• While the route cache is being updated, it is not available for pre-computed route lookup.
Department of Electrical Engineering and Computer Science 12
Route Cache Policy ..(cont’d)
• Selecting from Route Cache.– First, a check is made for the destination node
address.– If the set of DTLs is not empty, it is searched
for a route that will satisfy the QoS requirements of the call.
– If more than one such route exists, the route that leaves maximum residual bandwidth (load balancing) is chosen.
Department of Electrical Engineering and Computer Science 13
Route Cache Policy ..(cont’d)
• Invalidation of Cache Entries.– Can be set to occur when there is a Crankback
event indicating a bad routing decision.
• Replacement of Cache entries.– Occurs when an invalidated entry is filled by a
subsequent on-demand route computation.– Occurs when the cache is entirely rewritten
during an update.
Mechanics of Experimentation
Metrics, Traffic, Evaluation, Analysis
Department of Electrical Engineering and Computer Science 15
Metrics & Parameters• Average Bandwidth
Acceptance Ratio.• Average Call Setup
Time.• Location of Failures.
• Network Topology.• Peer group size.• Traffic Characteristics.
• Cache Update Policy
Department of Electrical Engineering and Computer Science 16
Flat Network Topology
Department of Electrical Engineering and Computer Science 17
Hierarchical Network Topology
Department of Electrical Engineering and Computer Science 18
Traffic ParametersArrival DistributionDestination ChoiceNode Type
Poisson [Y ]uniform among 9 other Type Y nodes
Type Y
Poisson [X]uniform among 9 other Type X nodes
Type X
Poisson [U]uniform among 35 other Type U nodes
Type U
40%
30%
30%
%Mean Duration (exponential)
Bandwidth Range (uniform distribution)
Application
1/60s5Mbps - 15MbpsHigh
1/45s1.5Mbps –3.6MbpsMedium
1/30s64Kbps - 1.5MbpsLow
Department of Electrical Engineering and Computer Science 19
Experimental Evaluation Setup
• Warm-up ratio of 10% for all runs.• Default PNNI Parameter values:
– Aggregation Policy: Symmetric Star.– Routing Policy: Widest-Min.hop.
Department of Electrical Engineering and Computer Science 20
Statistical Analysis
• Multiple runs with different random seeds, to compute average values.
• Test for statistical significance.– For the 95 % confidence interval (alpha = 0.05), the
data values lie on average within 2 % of the population mean.
• Hypothesis testing of difference of means using the Student’s t-distribution (for small sample sizes).
Department of Electrical Engineering and Computer Science 21
Establishing Value of Route Caching
• As a flat “single peer group” network scales, the average setup time should decrease for pre-computed routing as compared to on-demand routing.
• For a fixed topology, the average call acceptance ratio should decrease, when there are NO updates to the route cache.
Department of Electrical Engineering and Computer Science 22
Preliminary Results
Department of Electrical Engineering and Computer Science 23
Effect of Topology Scaling on Average Call Setup Times
R2 = 0.9866
050
100150200250300350400450500
24 32 40 48 56 64 72 80 88 96
Number of Nodes
Avg
. Set
up T
ime
(ms)
On-DemandPre-ComputedDifferenceExpon. (Difference)
Department of Electrical Engineering and Computer Science 24
Percentage Difference in Avg. Setup Times for Flat Network
0
10
20
30
40
50
60
24 32 40 48 56 64 72 80 88 96
Number of Nodes
Perc
enta
ge I
mpr
ovem
ent S
etup
T
ime(
%)
% Difference
Department of Electrical Engineering and Computer Science 25
Average Bandwidth Acceptance Ratio with Increasing Load
0
0.2
0.4
0.6
0.8
1
1.2
0.33 0.5 1 2
Mean Call Arrival Rate (calls/seconds)
Ave
rage
Ban
dwid
th A
ccep
tanc
e R
atio On-Demand
Pre-Computed
Pre-computed routing for Multiple Peer Group topologies
Motivation, Evaluation, and Results
Department of Electrical Engineering and Computer Science 27
Motivation
• Multiple peer grouping is the way by which PNNI handles the problem of network scaling.
• No previous study on pre-computed routing for MPG to our knowledge.
• Effect of PTSE aging and its repercussions on the quality of the routes more interesting within a MPG hierarchy.
Department of Electrical Engineering and Computer Science 28
Hierarchical MPG Topology
DTL A.1.1,A.1.2,A.1.3,A.2,B
B.2.2.1
B.2
Dst UNI
12
A.1
Src UNI
1 2
A.2
12
3
DTLA.2.1,A.2.2,B
Level 88
Level 96 B.1
12
DTLB.1.1,B.1.2,B.2
DTLB.2.1,B.2.2
A.1.1.1
A B
Level 80
Level 88
Department of Electrical Engineering and Computer Science 29
Route Caching in MPG networks
• Effect of Number of Peer Groups on average call setup time.
• Effect of Number of Peer Groups on average call acceptance ratio.
• Location of call failures as the Number of Peer Groups is changed.
Department of Electrical Engineering and Computer Science 30
Average Setup Time with Peer Group Size (no cache updates)
250270290310330350370390410430450
1 2 5 10
Number of Peer Groups
Ave
rage
Set
up T
ime
(ms)
On-DemandPre-Computed
Department of Electrical Engineering and Computer Science 31
Percentage Difference in Avg. Setup Times for MPG
02468
101214161820
1 2 5 10
Number of Peer Groups
Per
cent
age
Dif
fere
nce
PrC
ove
r O
nD (%
)
% Diff.
Department of Electrical Engineering and Computer Science 32
Average Bandwidth Acceptance with Peer Group Size
0.520.540.56
0.580.6
0.620.64
0.660.68
1 2 5 10
Number of Peer Groups
Ave
rage
Ban
dwid
th A
ccep
tanc
e R
atio On-Demand
Pre-Computed
Department of Electrical Engineering and Computer Science 33
Average Database Size Vs. Peer Group Size
0
5
10
15
20
25
30
Ave
rage
Top
olog
y D
atab
ase
Size
(KB
)
1 2 5 10
Number of Peer Groups
DB Size
Department of Electrical Engineering and Computer Science 34
Location of Call Failures with Peer Group Size (On-Demand)
05
1015202530354045
1 2 5 10
Number of Peer Groups
Per
cent
age
of F
aile
d C
alls
(%)
in Foreign PGin Local PGat Source
Department of Electrical Engineering and Computer Science 35
Location of Call Failures with Peer Group Size (Pre-Computed)
05
1015202530354045
1 2 5 10
Number of Peer Groups
Per
cent
age
of C
all F
ailu
res
(%)
in Foreign PGin Local PGat Source
Department of Electrical Engineering and Computer Science 36
Comparison of Failure Locations (Pre-Computed Vs On-Demand)
0
510
15
20
25
30
3540
45
OnD PrC OnD PrC OnD PrC OnD PrC
Number of Peer Groups
Per
cent
age
of C
all F
ailu
res
(%)
in Foreign PG
in Local PG
at Source
1 2 105
Route Cache Update Heuristics
Department of Electrical Engineering and Computer Science 38
Characteristics of a Route Cache Update Heuristic
• Intended to refresh the pre-computed cache with routes based on new topology data.
• The cache update policy should try and follow the network state as closely as possible.
• It should aim to retain the lower setup times obtained due to caching and attain the call success ratio of on-demand routing.
Department of Electrical Engineering and Computer Science 39
Types of Heuristics
• No updates (trivial case).• Timer based updates.• Crankback based invalidation.• PTSE count based updates.• Combinations of above policies: ptse_timer,
ptse_crankback, crankback_timer, ptse_timer_crankback.
Department of Electrical Engineering and Computer Science 40
Comparison of Heuristics for Updating the Cache
Metrics:• Average call setup time.• Average call acceptance ratio.• Combined metric = average call rejection
ratio * average call setup time.
Department of Electrical Engineering and Computer Science 41
Timer based updates
• Simplest and least intuitive of all update heuristics.
• Ideal value of timer??• Hypothesis: For an MPG topology,
synchronizing the timer with the re-aggregation rate will yield the best setup times.
Department of Electrical Engineering and Computer Science 42
Average Setup Time Vs. Update Timer Period
385
390
395
400
405
410
415
60 100 150 200
Cache Update Period (ms)
Setup Time (ms)
Re-Aggregation Period : 100 ms
Department of Electrical Engineering and Computer Science 43
Average Bandwidth Acceptance Ratio Vs.Update Timer Period
88.589
89.5
90
90.5
91
91.5
92
92.5
60 100 150 200
Cache Update Period (ms)
B/W Acceptance Ratio
Re-Aggregation Period : 100 ms
Department of Electrical Engineering and Computer Science 44
Crankback Based Invalidation
• Logical update condition, since a crankback happens when a route failure occurs.
• Indication of inaccurate information or staleness of cache entry.
• Results in invalidation of cache entry.• Subsequent on-demand routing and re-
filling of the cache entry.
Department of Electrical Engineering and Computer Science 45
PTSE Count Based Update
• PTSEs are flooded when there is significant change and/or after sufficiently long idle period.
• Counting PTSE’s received for any Peer Group/ Logical Level, is a good indicator of state change in that Peer Group/ Level.
• This metric is associated with network specific components, and is available essentially for free.
• PTSE count, upon reaching a specified value, is used to initiate a route cache update.
Department of Electrical Engineering and Computer Science 46
Combined Heuristics
• Each individual heuristic detects network changes by considering mutually independent information.
• Combining them should lead to a better heuristic for initiating route cache updates.
Department of Electrical Engineering and Computer Science 47
Avg. Setup Time for Different Cache Update Heuristics
200
250
300
350
400
450
none
timer pts
e cbk
ptsetim
er
timerc
bk
ptsecb
k all
on-de
mand
Ave
rage
Set
up T
ime
(ms)
Setup Time
Types of heuristics
Department of Electrical Engineering and Computer Science 48
Avg. Call Acceptance Ratio for Diff. Cache Update Heuristics
687072747678808284
none
timer pts
ecbk
ptsetim
er
timerc
bk
ptsecb
k all
on-de
mand
Ave
rage
Cal
l Acc
epta
nce
Per
cent
age
(%)
Call Acceptance
Types of heuristics
Department of Electrical Engineering and Computer Science 49
Performance of Route Cache Update Heuristics
60
70
80
90
100
110
120
none
timer pts
ecbk
ptsetim
er
timerc
bk
ptsecb
k all
on-de
mand
Scal
ed v
alue
of C
ombi
ned
Met
ric
Combined Metric
Types of heuristics
Department of Electrical Engineering and Computer Science 50
Revalidating results
• Applying ptse_crankback heuristic to the chosen MPG topology.
• Study effect of update heuristic on setup times and call acceptance ratio.
• Investigate location of failures.
Department of Electrical Engineering and Computer Science 51
Setup Time in Multiple Peer Group with Route Cache Updates
250270290310330350370390410430450
1 2 5 10
Number of Peer Groups
Ave
rage
Set
up T
imes
(ms)
On-DemandPre-Computedptse_cbk
Department of Electrical Engineering and Computer Science 52
B/W Acceptance in Multiple Peer Group with Route Cache Updates
0.50.520.540.560.580.6
0.620.640.660.680.7
1 2 5 10
Number of Peer Groups
Ave
rage
Ban
dwid
th A
ccep
tanc
e R
atio On-Demand
Pre-Computedptse_cbk
Department of Electrical Engineering and Computer Science 53
Best of Both Worlds !!
Department of Electrical Engineering and Computer Science 54
Location of failures
• Hypothesis: With route cache updates, the number of failures in foreign peer groups should be a smaller fraction of the total number of failures, as compared to no route cache updates.
Department of Electrical Engineering and Computer Science 55
Location of Call Failures (no updates)
05
1015202530354045
1 2 5 10
Number of Peer Groups
Per
cent
age
of C
all F
ailu
res
(%)
in Foreign PGin Local PGat Source
Department of Electrical Engineering and Computer Science 56
Location of Failures (with ptse_crankback heuristic)
05
10
15202530
3540
1 2 5 10
Number of Peer Groups
Per
cent
age
of C
all F
ailu
res
(%)
in Foreign PGin Local PGat Source
Department of Electrical Engineering and Computer Science 57
Comparison of Failure Locations(No Update Vs. Ptse_cbk)
05
1015202530354045
no up
date
ptse c
bk
no up
date
ptse c
bk
no up
date
ptse c
bk
no up
date
ptse c
bkNumber of Peer Groups
Lo
ca
tio
n o
f C
all
Fa
ilu
res
(%
)
Foreign PG Failed
Local PG Failed
Source Failed
1 PG 2 PG 5 PG 10 PG
Department of Electrical Engineering and Computer Science 58
On-Demand Processing Times
Department of Electrical Engineering and Computer Science 59
Pre-Computed Processing Times
Department of Electrical Engineering and Computer Science 60
On-Demand Processing Time
Department of Electrical Engineering and Computer Science 61
Pre-Computed Processing Time
Conclusion
Lessons Learned, Future Work, Q&A
Department of Electrical Engineering and Computer Science 63
Lessons Learned
• Pre-computed route caching is effective in reducing the average call setup time as compared to on-demand routing, even as the network scales.
• On-demand routing gives better call acceptance as compared to pre-computed routing, for increasing traffic load.
• Route cache updates helps in increasing the call acceptance ratio of pre-computed routes, while achieving lower call setup times.
Department of Electrical Engineering and Computer Science 64
Lessons Learned
• In Multiple Peer Group topologies, periodic timer based route cache updates are most effective when synchronized with the re-aggregation timer.
• PTSE count heuristic is an effective means of tracking network state.
• Combining ptse count and crankback initiated invalidation shows best overall performance among the chosen heuristics.
Department of Electrical Engineering and Computer Science 65
Future Work
• Route Cache Policy enhancements:– Efficient route pre-computation.– Better cache data structures for storage of
routes.– Check relevance of a chosen cache entry before
it is used for connection establishment.– Investigating effects of crankback retries and
alternate routing.
Department of Electrical Engineering and Computer Science 66
Future Work (cont’d)
• Hybrid Routing algorithm:– on-demand computations to first determine the
foreign peer groups to traverse in a route,– then consult a route cache for crossing the
source peer group to reach the first border node.
• Use of PTSE count to indicate when to carry out re-aggregation within a peer group.
Department of Electrical Engineering and Computer Science 67
Q&A
Department of Electrical Engineering and Computer Science 68
Thank You!!