Opportunistic Traffic Scheduling Over Multiple Network Path Coskun Cetinkaya and Edward Knightly.
Post on 05-Jan-2016
214 Views
Preview:
Transcript
Opportunistic Traffic Scheduling Over Multiple Network Path
Coskun Cetinkaya and Edward Knightly
Edward Knightly
Multi-Path Routing
Establishes and simultaneously uses multiple parallel paths– Key advantage is efficiency
Routing protocol assigns weights to paths– OSPF, QoS routing, traffic engineering
Edward Knightly
Existing Splitting Techniques
Per packet round robin forwarding– Simplest and most frequently used– Degrades TCP throughput due to re-ordering
Per flow hashing – Fine splitting granularity and no TCP re-ordering– Per-TCP-flow lookup limits implementation feasibility
Destination prefix based forwarding– Coarse-granularity splitting and no TCP re-ordering– Unpredictable load splitting that may not match desired
weights
All ignore path quality in splitting decision
Edward Knightly
Our Thesis
Observe– Routing weights change slowly (from traffic engineering) – Quality of paths changes continuously
Opportunistic Multipath Scheduling– Exploits short-term capacity variations on different paths via
scheduling packets to opportunistically favor low-delay paths– Obey weights at long time scales to ensure “global”
objectives
Hypothesis– Improve throughput/delay, no per-flow lookup, satisfy
weights– TCP throughput improvements due to RTT reduction will
overwhelm re-ordering effects
Edward Knightly
System Model
Design: scheduling/traffic splitting policy
Objective: minimize mean delay of multipath traffic– Decrease RTT and loss rate increase TCP throughput
Subject to: mean traffic on path i = i (path weight)
Multipath traffic
…
Cross traffic
Splitter
Edward Knightly
Xk = size of packet k I(sk,i) = 1 of packet k is scheduled on path i, 0
otherwise For equal capacity paths minimizing delay is
equivalent to minimizing the expected queue length
Mathematical Formulation
)),((
}{..
}{min
1ik
kki
ki
ki
ik
k
CisIXAQQ
isPts
ksQE
Edward Knightly
Optimal Scheduler
Assumptions:– Cross-traffic and multi-path traffic are stationary
processes queue length is stationary – Multi-path traffic does not change path conditions
Using a wireless scheduling analogy [LCS02], we can show that the optimal scheduler is threshold based:
iiijjji
i
ii
vQvQPtsv
vQQS
}min{..
)min(arg)(***
**
Contrast to “join the shortest queue” policy which ignores weights
Edward Knightly
Evaluate the performance under self-similar cross traffic
Queue size distribution is Weibull:
Expected queue size (and delay):
Round Robin Optimal Scheduler
Performance of the Optimal Policy
iix
i exQP }{
)1(2 ii H
iH
i
ii
iiii H
Hm
Hma
2
2
)1)(1(
)1(2
1
i
i
i
iRRQE
/1
1 )1(}{
*
* )(}{
v
vx
opt dxeQEi
ii
)max( **ivv
Edward Knightly
On-Line Computation of v*
In practice, we do not know the queue length or its distribution
Threshold update:– stochastic approximation technique [KC78,LCS02]
Scheduling decision:– Qi
k estimated via probes
)),((11 isIvv ki
ki
ki k
}{minarg ki
ki
ik vQs
Edward Knightly
Evaluation Scenario
Two paths with capacity 10 Mb/sec Cross-traffic: self-similar with mean rate m[0.3,
0.9], variance coefficient a[0.5,4], and Hurst Parameter H[0.5,0.9]
Multi-path traffic is constant-rate or TCP Gain defined as
}{
}{1
RR
OMS
QE
QEG
…
Edward Knightly
Model: gain depends only on H and # paths and is 50%
Higher N more path diversity higher gain
Large H long-time scale path correlation higher gain
Homogeneous Paths: Model
)1(2/1
11
HNG
Edward Knightly
Simulated gains higher than predicted by model – Model serves as lower bound– Queue distribution is asymptotic lower bound, tighter for larger queues
Delay increases with increasing mean (m) and variance coefficient (a) Gain (relative) is highest under higher H, lower m, lower a
Homogeneous Paths: Simulation
Edward Knightly
Gain increases with path diversity (increasing ratio of variance coefficient) – OMS exploits different path properties subject to
weights
Heterogeneous Paths: Impact of Variance Coefficient Ratio
H=0.6
m=0.7
Edward Knightly
So far, assumed path information is immediately available at scheduler/splitter
RTT-scale delay to obtain buffer state (via probes or ECN) Gain decreases as information delay increases High gain for measured values of traffic (0.7 < H < 0.85)
and delay (1 < RTT < 100 msec)
Effect of Information Delay
m=0.9
a=0.5
Edward Knightly
When can OMS do worse than RR? Three combined factors:– iid traffic having no long-time-scale bursts– High information delay– High ratio of multi-path traffic to cross-traffic (scheduled
traffic itself determines conditions)
Limits of OMS
Edward Knightly
TCP Multi-Path Traffic
With RR, multipath traffic achieves only 20% to 38% of fair share– High cost of mis-ordering and delay
TCP/OMS significantly outperforms TCP/RR– TCP/OMS requires an aggregate level of only 10 cross-traffic
flows to achieve maximum performance– OMS impact overwhelms effect of TCP variants
10 msec probing interval
32 kb/s probing overhead
(0.32% of capacity)
Edward Knightly
Probing Interval and TCP Traffic
Base case probing interval: 10 msec interval and 32 kb/sec Faster 1 msec probing yields higher-than-fair share for
multi-path flows Slower probing (e.g., 3.2 kb/sec) reduces performance
Edward Knightly
Summary
Multipath routing promises increased efficiency and performance
Today’s traffic splitting ignores path dynamics and– inhibits TCP throughput via reordering,– requires expensive per-TCP flow lookups, or– cannot achieve weights via prefix splitting
Opportunistic Multipath Scheduling– Improves throughput/delay via a measurement
based opportunistic policy that satisfies routing weights
– Gains overwhelm occasional misordering
http://www.ece.rice.edu/networks
top related