Upload
gemma-beard
View
20
Download
1
Embed Size (px)
DESCRIPTION
Overload Design Team Status. Jonathan Rosenberg Cisco. Team Overview. Design team formed January 2007 Working on simulations of overload mechanisms Composed of experts in this area Many non-IETF participants Bi-Weekly Calls. Current Status. Agreed on simulation model - PowerPoint PPT Presentation
Citation preview
Overload Design Team Status
Jonathan Rosenberg
Cisco
Team Overview
• Design team formed January 2007
• Working on simulations of overload mechanisms
• Composed of experts in this area
• Many non-IETF participants
• Bi-Weekly Calls
Current Status
• Agreed on simulation model
• Several indpendendent simulators developed
• Simulations performed to determine baseline behavior– Drop requests on overload, no 503 [sim1]– Send 503 on overload, no retry [sim2]– Retry on 503 [sim3]
Baseline Results (Volker)
Baseline Results (Shen)
Simulation Comparison with Stateless 503
0
20
40
60
80
100
120
140
0 200 400 600 800 1000 1200
Load
Go
od
pu
t
Simulation 1Simulation 2 StatelessSimulation 3 Stateless
Stateful vs. Stateless 503
• Found that stateless 503 server transactions seriously worsen performance
• Team agreed to use stateless 503 model despite this because faster 503 processing is critical overall
Simulation Comparison with stateful 503
0
20
40
60
80
100
120
140
0 200 400 600 800 1000 1200
Load
Go
od
pu
t
Simulation 1Simulation 2 StatefulSimulation 3 Stateful
Simulation Comparison with Stateless 503
0
20
40
60
80
100
120
140
0 200 400 600 800 1000 1200
Load
Go
od
pu
tSimulation 1Simulation 2 StatelessSimulation 3 Stateless
Nortel Algorithm
• Use 503/Retry-After and turn off traffic to server during interval
• Retry-After value determined algorithmically from queue sizes
• Equals amount of time for queue to drain to low watermark
ResultsGoodput in the Network
-200
2040
6080
100120
140160
0 500 1000 1500
call rate (calls/s)
go
od
pu
t (c
alls
/s)
stateful 503, sim3
Nortel's OverloadControl Algorithmproposal
stateful 503, sim2
AT&T Algorithm
• During each Measurement Interval, Core Proxy estimates:– Mean processing rate t,– Queue length at end of measurement interval qt,
• Core Proxy calculates the retry after period as the time needed for Core Proxy to process the queued messages that exceed the target queueing delay:
Retry-after-delay = qt/t - a x de
• Retry-after-delay updated each Control Interval and distributed to Edge Proxies in 503 responses,
Results
0
20
40
60
80
100
120
140
0 200 400 600 800Offered load (cps)
Car
ried
lo
ad (
cps)
AT&T Sim3 Statefull 503 gput
AT&T Sim3 Stateless 502 gput
AT&T Sim3 Statefull 503 with ovld gput
AT&T Sim3 Stateless 503 with ovld gput
Columbia Algorithm
• Sending elements send their current load forward
• Receiving elements send back desired load
• Receiving elements can compute fair share bandwidth allocations and fine grained load adjustment
• No simulation results yet
Bell Labs Algorithm
• Overload Control Feedback Loop in SIP– SIP extension that realizes an overload control
feedback loop to upstream neighbors.– Transport vehicle for feedback generated by an
overload control algorithm.• Orthogonal to the specific control algorithm used.
– draft-hilt-sipping-overload-03.txt
• Overload Control Algorithm– Algorithm that determines fine grained load feedback
to adjust incoming load.
• No simulation results yet.
Next Steps
• Waiting for folks to finish implementing their algorithms
• Once done, can do the real work of comparing results and doing additional simulations
• Targeting proposal for Philadelphia