multi-link topology: recent resultsbalaji/presentations/au-ashvin... · 2010-03-08 · 15 20 25 30...
TRANSCRIPT
Multi-link topology: Recent results
Ashvin Lakshmikantha, Balaji Prabhakar
UIUC and Stanford University
Multi-link topology: Recent results – p.1/19
Outline
Study flow-level performance.
Description of the Simulations
Interpretation of the results
Conclusions
Multi-link topology: Recent results – p.2/19
Introduction
So far......Study of N long-lived flows:
First step in the analysis of a protocol.Model amenable to control theoretic analysis.
Deterministic analysis: Helpful in tuning parameters.
Link utilization ⇐⇒ per-flow throughput
Fairness properties: Can be studied by comparing the throughput achieved byvarious flows.
Real network-like situation:Flows arrive and depart.
Finite (but random) file sizes.
Number of flows in the network random.Use flow completion time as a metric to evaluate network performance andfairness.
bandwidth =flowsize
FCT
Multi-link topology: Recent results – p.3/19
Introduction
Flow completion time:
Time taken for a flow to transmit the entire file.Depends on network load.
Good metric to measure end user performance.
Fairness??Has to be qualified indirectly.Ex: Measure the variance in the completion times of the same file at varioustimes.Ex: Measure flow completion times of flows on different paths, but samecongestion point.
Good network performance :
Number of flows in the network is bounded.Flows face a finite delay.
Multi-link topology: Recent results – p.4/19
Simulation: Goals
Study the effect of congestion spreading in multi-link topology.
Study the robustness of BCN protocolTo changes in− flow sizes− starting transmission rate− turning off switch-signalled rate increases
Using− flow completion time− fairness (variance of FCT)
Multi-link topology: Recent results – p.5/19
Network Topology
10Gbps
10Gbps
10Gbps
Fixed Load
Varying Load
Workgroup Switch
Catalyst
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Figure 1: Topology
Multi-link topology: Recent results – p.6/19
Simulation parameters
Traffic parameters:
Arrival process Poisson
File size distribution Exponential
Mean file size 1MB
RTT 100µs
Total load 50% − 80%
BCN parameters:
W 4
Gi 2
Gd1
128
Starting rate 1Gbps
Drift Multiplicative
a 10/sec
Buffer size 100pkts = 150KB
X0FF 75pkts
XON 25pkts
sampling probability 0.03
Qe 16pkts (24KB)
Multi-link topology: Recent results – p.7/19
Effect of BCN
5
10
15
20
25
30
35
40
0.6 0.65 0.7 0.75 0.8 0.85
Flo
w c
ompl
etio
n tim
es (
mse
c)
Offered Load
Flow completion times with BCN messaging
Incr loadFixed load
0
20
40
60
80
100
120
0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76 0.78
Flo
w c
ompl
etio
n tim
es (
mse
c)
Offered Load
Flow completion times without BCN messages
Increasing LoadSteady Load
10Gbps
10Gbps
10Gbps
Fixed Load
Varying Load
Workgroup Switch
Catalyst
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Multi-link topology: Recent results – p.8/19
Effect of mean flow size and flow size distribution
Traffic parameters:
Arrival process Poisson
File size distribution Hyper-exponential
Short-flow size 20pkts
Short-flows percentage 90%
Long-flow size 320pkts
Long-flow percentage 10%
Mean file size 50KB
RTT 100µs
Total load 50% − 80%
BCN parameters:
W 4
Gi 2
Gd1
128
Starting rate 1Gbps
Drift Multiplicative
a 10/sec
Buffer size 100pkts = 150KB
X0FF 75pkts
XON 70pkts
sampling probability 0.03
Qe 16pkts (24KB)
Multi-link topology: Recent results – p.9/19
Effect of BCN
0.1
1
10
100
0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76
Flo
w c
ompl
etio
n tim
e (m
s)
Offered Load
Flow completion times of flows with starting rate 1Gbps
sf-fixedloadlf-fixedload
sf-varyingloadlf-varyingload
0.1
1
10
100
0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7
Flo
w c
ompl
etio
n tim
e (m
s)Offered Load
Flow completion times of flows with starting rate 1Gbps (No BCN)
sf-fixedloadlf-fixedload
sf-varyingloadlf-varyingload
10Gbps
10Gbps
10Gbps
Fixed Load
Varying Load
Workgroup Switch
Catalyst
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Multi-link topology: Recent results – p.10/19
Interpretation of the results
Mean flow size = 1MB: gain in throughput is about 10%.
Mean flow size = 50KB: gain in throughput is about 5%.
Flow completion time: Using BCN messages improves the overall FCT.
Fairness:Without BCN, FCT depends on the loading of uncongested links too !
With BCN, FCT depends on most congested links.
BCN messages, helps improve the fairness.
Multi-link topology: Recent results – p.11/19
Effect of BCN on the FCT
0.001
0.01
0.1
1
1 10 100 1000 10000
Flo
w c
ompl
etio
n tim
e (s
econ
ds)
File Size (packets)
Load=0.55Load=0.60Load=0.65Load=0.70Load=0.75
0.001
0.01
0.1
1
1 10 100 1000 10000
Flo
w c
ompl
etio
n tim
e (s
econ
ds)
File Size (packets)
Load=0.55Load=0.60Load=0.65Load=0.70
Multi-link topology: Recent results – p.12/19
Interpretation of the results
Without BCN, no priority for short-flows. All flows are worse off equally.
With BCN, the short-flows completion time remain same irrespective of the loading.
At 70% loading FCT of 80% of the flows remain unchanged.
Multi-link topology: Recent results – p.13/19
Bursty loading
10Gbps
10Gbps
10Gbps
L1
L2
Workgroup Switch
Catalyst
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Multi-link topology: Recent results – p.14/19
Effect of BCN on the FCT
0.1
1
10
100
0.55 0.56 0.57 0.58 0.59 0.6 0.61 0.62 0.63 0.64 0.65 0.66
Flo
w c
ompl
etio
n tim
e (m
s)
Offered Load
Flow completion times of flows with starting rate 1Gbps (with BCN)
sf-load1lf-load1
sf-load2lf-load2
1
10
100
1000
0.55 0.56 0.57 0.58 0.59 0.6 0.61 0.62 0.63 0.64 0.65 0.66
Flo
w c
ompl
etio
n tim
e (m
s)
Offered Load
Flow completion times of flows with starting rate 1Gbps (No BCN)
sf-load1lf-load1
sf-load2lf-load2
10Gbps
10Gbps
10Gbps
L1
L2
Workgroup Switch
Catalyst
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Multi-link topology: Recent results – p.15/19
Effect of switch-signalled rate increase
10Gbps
10Gbps
10Gbps
Fixed Load
Varying Load
Workgroup Switch
Catalyst
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Multi-link topology: Recent results – p.16/19
Effect of switch-signalled rate increase
Switch increase is turned off. RPs respond only to switch decrease messages.
0.1
1
10
0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76
Flo
w c
ompl
etio
n tim
e (m
s)
Offered Load
Flow completion times of flows with starting rate 10Gbps:Short flows
sf-fixedloadsf-varyingload
sf-fixedload-self-incr-onlysf-varyingload-self-incr-only
1
10
100
0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76
Flo
w c
ompl
etio
n tim
e (m
s)
Offered Load
Flow completion times of flows with starting rate 10Gbps: Long flows
sf-fixedloadsf-varyingload
sf-fixedload-self-incr-onlysf-varyingload-self-incr-only
Multi-link topology: Recent results – p.17/19
Effect of starting rate
0.1
1
10
100
0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68 0.7 0.72 0.74 0.76
Flo
w c
ompl
etio
n tim
e (m
s)
Offered Load
Flow completion times of flows with starting rate 1Gbps
sf-fixedloadlf-fixedload
sf-varyingloadlf-varyingload
0.1
1
10
100
1000
0.54 0.56 0.58 0.6 0.62 0.64 0.66 0.68
Flo
w c
ompl
etio
n tim
e (m
s)Offered Load
Flow completion times of flows with starting rate 10Gbps
sf-fixedloadlf-fixedload
sf-varyingloadlf-varyingload
10Gbps
10Gbps
10Gbps
L1
L2
Workgroup Switch
Catalyst
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Cisco 7500 SERIES
CiscoSystems
Multi-link topology: Recent results – p.18/19
Conclusions
Studied the effect of BCN system in a multi-link topology.
Studied the performance of BCN under flow arrivals and departures.
Studied the effect ofFlow size distribution.Mean flow size.BCN increase messages.
Link pauses.
on the performance of BCN.
Qualitatively studied the fairness properties of BCN.
Multi-link topology: Recent results – p.19/19