Upload
morgan-doyle
View
212
Download
0
Embed Size (px)
Citation preview
Exploiting the Path Propagation Time Differencesin Multipath Transmission with FEC
Maciej Kurant EPFL, Switzerland
Infocom 2009, April 19-25, Rio de Janeiro
Acknowledgements:Patrick Thiran (EPFL), Dan Jurca (DoCoMo, Munich), Pascal Frossard (EPFL)
2
T
source destination
relay
Multipath transmission with FEC
Requirements:• few losses• small delay (no retransmission possible)
schedule
data to send
protection (FEC)
received reconstructed
3
destination
relay
source
T=5ms
ExampleAssumptions:1. Paths are independent
4
0.553%
0.148%
destination
relay
source
T=5ms
1.000%
1%, 10ms
1%
10ms
Example
Schedule:
Time t =0
Assumptions:1. Paths are independent2. Continuous-time Gilbert model for packet losses: • average loss rate • average loss burst length
Effective loss rate :
State of the art - IMMEDIATE
5
1.000%
0.553%
0.148%
0.113%
destination
relay
source
∆t >> 0 (here assume 50ms)
We use ∆t to SPREAD the packets on the green path, so that the total block delay is not changed.
∆t = 50ms
Example
State of the art - IMMEDIATE
Schedule:
Time t =0
slow!
Effective loss rate :
6
1.000%
0.553%
0.148%
0.113%
destination
relay
source
Example
We use ∆t to SPREAD the packets on the green path, so that the total block delay is not changed.
The optimal rates on the paths may change.0.016%
The gain over state of the art is close to one order of magnitude!
Schedule:
Time t =0 Effective loss rate :
slow!
State of the art - IMMEDIATE
7
Contributions
8
1. Exact solution of the model
Previous attempts (using approximations):• L. Golubchik et al, “Multi-path continuous media streaming. what are the benefits?” Performance Evaluation Journal, 2002.
• E. Vergetis, R. Guerin, and S. Sarkar, “Realizing the benefits of user-level channel diversity,” SIGCOMM Comput. Commun. Rev., 2005.
• Y. Li, Y. Zhang, L. Qiu, and S. Lam, “SmartTunnel: Achieving reliability in the internet,” Proc. of INFOCOM’07, 2007.
• …
Solution (effective loss rate ):
destination
relay
source
Assumptions:1. Paths are independent2. Continuous-time Gilbert model for packet losses: • average loss rate • average loss burst length
data to sendprotection (FEC)
Model:
schedule
9
2. Observation that ∆t>0
∆t = t2 - t1 [ms]
t1
t2
∆t = max(t2,t3) - t1 [ms]
t1
t2
t3
10
Time t =0
SPREAD rules:
• On each path, spread the packets evenly in time within all the available time budget
• Find the best packet rates for each path
IMMEDIATE rules: • Send each packet immediately after it is generated at the source• Find the best packet rates for each path• Assign packets to paths to minimize the loss rate
3. Proposition of SPREAD
T
Usually very good,
but not necessarily optimal
Might schedule a packet departure
before it is generated!
11
i.e., a relative loss rate improvement over the state of the art
path propagation time difference [ms] source packet generation period [ms] number of packets per FEC block
)(
)(*
*
SPREAD
IMMEDIATE
B
B
4. Evaluation: Loss rate improvement
12
path propagation time difference [ms]
FE
C b
lock
del
ay g
ain
[ms]
4. Evaluation: Minimizing delays
“Minimize the FEC block delay of SPREAD,
but keep .”)()( ** IMMEDIATESPREAD BB
Reverse the problem, i.e.:
Useful, e.g., to minimize the effect of jitter.
13
5. Trace-driven evaluation
14
Limitations
• Simple performance metric • Not (yet) adaptive• Assumes path independence• Assumes that we control when packets
are actually sent• Needs more experimental verification
But the general principle remains.
15
Conclusions
• Paths often have different propagation times.
• This can be exploited in multipath FEC transmission to reduce:– the effective loss rate– the total FEC block delay.
• To achieve this, we:– solved the multipath FEC model,– proposed and evaluated SPREAD.
16
SPREADSpace Packets Regularly Exploiting Asymmetry in Delays
MASTErPIECEMultipAth Scheduler That Exploits PropagatIon diffErenCEs
bETTERExploit Transmission Time diffERences
SPLENDiDSpace Packets evenLy and Exploit Network Delay Differences
Sponsored article
Sponsored article
Sponsored article
Sponsored article
Not so easy to find?
Try AcronymCreator.net !
17
Thank you !
18
19
20