radhika mittal, justine sherry, sylvia ratnasamy, …...short flow completion time " •...
TRANSCRIPT
![Page 1: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/1.jpg)
RC3���Recursively Cautious Congestion Control
Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, Scott Shenker
UC Berkeley
![Page 2: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/2.jpg)
Roadmap
• Isn’t congestion control a solved problem? – Conflicting goals of high throughput and friendliness solved through priorities
• Scope for performance gains – Increases with increasing RTTxBW
• Design Details – Additional packets sent backwards from the end using multiple low priority levels
• Simulation Results – 40-80% reduction in FCT over baseline TCP implementation
• Linux Implementation and Evaluation – Simple modifications, agnostic to the underlying congestion control algorithm
• Challenges and Future
![Page 3: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/3.jpg)
Roadmap
• Isn’t congestion control a solved problem? – Conflicting goals of high throughput and friendliness solved through priorities
• Scope for performance gains – Increases with increasing RTTxBW
• Design Details – Additional packets sent backwards from the end using multiple low priority levels
• Simulation Results – 40-80% reduction in FCT over baseline TCP implementation
• Linux Implementation and Evaluation – Simple modifications, agnostic to the underlying congestion control algorithm
• Challenges and Future
![Page 4: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/4.jpg)
Short Flow Completion Time
• “Being fast really matters. Users really respond to speed.”
– 0.5 sec delay caused a 20% drop in traffic – Google
– 2 sec slowdown changed queries/user by -1.8% and revenue/user by -4.3% – Bing
– 5 sec speedup resulted in a 25% increase in page views and 7-12% increase in revenue – Shopzilla
- James Hamilton’s Blog
![Page 5: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/5.jpg)
RC3 in a nutshell
Send additional packets from the flow using low priority service (WQoS),
filling up only the spare capacity in the network • 40-80% Reduction in Flow Completion Time
• No harm to the regular high priority traffic • Better use of Network Resources
![Page 6: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/6.jpg)
Example Scenario
Network Provider Receiver Sender
![Page 7: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/7.jpg)
Network Provider Viewpoint
What if I get a burst of traffic in peak hours or a failure occurs?
Must overprovision 30-50% average link utilization
Sender Network Provider
Receiver
![Page 8: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/8.jpg)
Endhost Viewpoint
The network might be very
congested!
Must ramp-up cautiously
Sender Network Provider
Receiver
![Page 9: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/9.jpg)
TCP
Wasted Capacity Long FCT Slow Start
Congestion Avoidance Link Fully Utilized
![Page 10: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/10.jpg)
The Root Cause
Two Goals of Congestion Control – Fill the pipe for high throughput
– Do no harm to other flows
Traditional Approach – Single mechanism tries to balance the two conflicting goals
RC3: Decouple these goals using priorities – Fill the pipe at lower priority – Do no harm at higher priority
![Page 11: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/11.jpg)
RC3 in action
Additional Packets at Low Priority Fill the Pipe Regular TCP at High Priority
Flow Completes Sooner
![Page 12: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/12.jpg)
Example: FCT with Slow Start
7 packets flow (with initial congestion window of 1 segment) completes in 3RTTs under slow start
Sender Network Provider
Receiver
![Page 13: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/13.jpg)
Example: FCT with Slow Start
7 packets flow (with initial congestion window of 1 segment) completes in 3RTTs under slow start
Sender Network Provider
Receiver
![Page 14: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/14.jpg)
Example: FCT with Slow Start
7 packets flow (with initial congestion window of 1 segment) completes in 3RTTs under slow start
Sender Network Provider
Receiver
![Page 15: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/15.jpg)
Example: FCT with RC3
Remaining 6 packets sent at lower priority with the 1st packet Flow completes in 1RTT
Sender Network Provider
Receiver
![Page 16: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/16.jpg)
Roadmap
• Isn’t congestion control a solved problem? – Conflicting goals of high throughput and friendliness decoupled through priorities
• Scope for performance gains – Increases with increasing RTTxBW
• Design Details – Additional packets sent backwards from the end using multiple low priority levels
• Simulation Results – 40-80% reduction in FCT over baseline TCP implementation
• Linux Implementation and Evaluation – Simple modifications, agnostic to the underlying congestion control algorithm
• Challenges and Future
![Page 17: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/17.jpg)
Theoretical Model
Flow Size (N) i A×RTT
% F
CT
Re
du
ctio
n
TCP: 1 RTT
RC3: 1 RTT
TCP: RTTs
RC3: 1 RTT
log(N
i
) TCP: RTTs +
RC3: 1 RTT +
log(A⇥RTT
i
)N
A
N
A
Initial Congestion
Window
Available Capacity A = (1-u)BW
![Page 18: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/18.jpg)
Parameter Sensitivity: AxRTT
TCP: 1 RTT
RC3: 1 RTT
TCP: RTTs
RC3: 1 RTT
log(N
i
) TCP: RTTs +
RC3: 1 RTT +
log(A⇥RTT
i
)N
A
N
A
Flow Size (N) i
% F
CT
Re
du
ctio
n
A×RTT
Higher A×RTT
Initial Congestion
Window
Available Capacity A = (1-u)BW
![Page 19: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/19.jpg)
Roadmap
• Isn’t congestion control a solved problem? – Conflicting goals of high throughput and friendliness decoupled through priorities
• Scope for performance gains – Increases with increasing RTTxBW
• Design Details – Additional packets sent backwards from the end using multiple low priority levels
• Simulation Results – 40-80% reduction in FCT over baseline TCP implementation
• Linux Implementation and Evaluation – Simple modifications, agnostic to the underlying congestion control algorithm
• Challenges and Future
![Page 20: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/20.jpg)
WQoS Implementation
Routers offer several layers of worse service – Use Priority Queues – Support already present
Packets carry priority (possibly) in DSCP field – Priority 0 – default (highest) – Priority 1, Priority 2, Priority 3,….
![Page 21: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/21.jpg)
RC3 Design
RC3 runs two parallel control loops
– TCP control loop Transmits packets that obey unmodified TCP logic at highest priority
– Recursive Low Priority (RLP) control loop Transmits additional packets at low priority
![Page 22: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/22.jpg)
What packets are sent at low priority?
Minimum overlap between packets sent by the two control loops for maximum gains
– RLP starts from the last packet in buffer
– Goes in reverse order
0 1 3 2 N N-1 4 …. ….
High Priority Low Priority
TCP Control Flow RLP Control Flow
![Page 23: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/23.jpg)
Single Flow
Sender Network Provider Receiver
1 1 1 0 1 1 1 1 1 Bottleneck
BWxRTT = 9 packets ...
![Page 24: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/24.jpg)
Multiple Flows?
1 1 1 0
1 1 1 1 1
Sender 2
Sender 1
Network Provider Receiver
![Page 25: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/25.jpg)
Router’s Priority Queue
0 1 1 1 0 1 1 1 1 1 1 1
![Page 26: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/26.jpg)
Multiple Flows?
0 1 1 0 1 1 1 1 1 Sender 1
Sender 2
Network Provider Receiver
![Page 27: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/27.jpg)
Recursively Cautious Congestion Control
• Use multiple priority levels • Send exponentially larger number of packets at
each priority level
0 1 3 2 N …. .... N-X …. N-10X
Priority 1 Priority 0 Priority 2 Priority 3
TCP Control Flow RLP Control Flow
![Page 28: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/28.jpg)
RC3 Design: Quick Recap
Two parallel control loops – Regular TCP – Recursive Low Priority (RLP)
Minimum overlap between the two control loops – Send low priority packets from the end in reverse
order
Max-min fairness across flows – Use multiple priority levels
![Page 29: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/29.jpg)
Roadmap
• Isn’t congestion control a solved problem? – Conflicting goals of high throughput and friendliness decoupled through priorities
• Scope for performance gains – Increases with increasing RTTxBW
• Design Details – Additional packets sent backwards from the end using multiple low priority levels
• Simulation Results – 40-80% reduction in FCT over baseline TCP implementation
• Linux Implementation and Evaluation – Simple modifications, agnostic to the underlying congestion control algorithm
• Challenges and Future
![Page 30: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/30.jpg)
Simulation Setup
• Multi-hop Internet-2 network topology – 10 core nodes, 100 end hosts
• 1Gbps bottleneck bandwidth • 40ms average RTT
• Baseline is 30% average link utilization • Pareto flow size distribution with Poisson inter-arrival • Initial Congestion Window of 4 segments
![Page 31: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/31.jpg)
Comparing baseline simulation results with the theoretical model
Flow Size (N)
i A×RTT
% F
CT
Re
du
ctio
n
![Page 32: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/32.jpg)
Baseline
![Page 33: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/33.jpg)
Flow Size < 4MSS - No RC3 packets sent
Benefits because high priority congestion is reduced
Baseline
![Page 34: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/34.jpg)
Model does not account for queuing
delays and drops
Baseline
![Page 35: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/35.jpg)
Average Over Flows
Average Over Bytes
Regular TCP 0.135s 0.443s
RC3 0.076s 0.114s
% Reduction 43.54% 74.35%
Baseline
![Page 36: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/36.jpg)
Comparing RC3 with other schemes
![Page 37: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/37.jpg)
RC3 in comparison
• Increasing the initial congestion window
• Rate Control Protocol (RCP)
![Page 38: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/38.jpg)
RC3 in comparison
• Increasing the initial congestion window
• Rate Control Protocol (RCP)
![Page 39: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/39.jpg)
Comparison: Increasing InitCWnd
Flow Size (N) i A×RTT i'
TCP: 1 RTT
RC3: 1 RTT
TCP: RTTs
RC3: 1 RTT
log(N
i
) TCP: RTTs +
RC3: 1 RTT +
log(A⇥RTT
i
)N
A
N
A
% F
CT
Re
du
ctio
n
![Page 40: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/40.jpg)
Comparison: Increasing InitCwnd
![Page 41: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/41.jpg)
Short Flows penalized by Increasing Initial
Congestion Window; Benefit from RC3
Comparison: Increasing InitCwnd
![Page 42: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/42.jpg)
Long flows see stronger gains with
RC3
Comparison: Increasing InitCwnd
![Page 43: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/43.jpg)
RC3 in comparison
• Increasing the initial congestion window
• Rate Control Protocol (RCP)
![Page 44: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/44.jpg)
Comparison: RCP
![Page 45: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/45.jpg)
RCP penalizes short flows due to more
aggressive long flows and explicit pacing
Comparison: RCP
![Page 46: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/46.jpg)
Comparison: RCP
Long flows see similar performance gains
![Page 47: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/47.jpg)
Stress Testing RC3
• Varying Link Utilization • Varying RTTxBW • More Topologies
• Different Workload • Link Heterogeneity • Random Losses
• Varying Priority Assignments • Application Pacing • Comparison with traditional QoS
![Page 48: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/48.jpg)
Roadmap
• Isn’t congestion control a solved problem? – Conflicting goals of high throughput and friendliness decoupled through priorities
• Scope for performance gains – Increases with increasing RTTxBW
• Design Details – Additional packets sent backwards from the end using multiple low priority levels
• Simulation Results – 40-80% reduction in FCT over baseline TCP implementation
• Linux Implementation and Evaluation – Simple modifications, agnostic to the underlying congestion control algorithm
• Challenges and Future
![Page 49: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/49.jpg)
RC3 in Implementation • Implemented in Linux 3.2 kernel
• 121 additional LOC – Sending Data Packets: 74 LOC – Receiving Data Packets and Acks: 47 LOC
• Agnostic to the underlying TCP algorithm – Can be Tahoe, Reno, NewReno, BIC, CUBIC etc
![Page 50: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/50.jpg)
Evaluation
![Page 51: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/51.jpg)
Evaluation
Low priority out-of-order packets processed by slow path High per-packet CPU overhead
![Page 52: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/52.jpg)
Leveraging NIC Offloading
• TCP Segmentation Offload (TSO) – Multiple segments processed by sender stack as a single chunk
• Large Receive Offload (LRO)
– Multiple segments received aggregated into a single chunk
• RC3 supports offloading to reduce CPU overhead – Logically treat each chunk as a single packet at the sender – This allows aggregation of segments at the receiver
![Page 53: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/53.jpg)
Leveraging NIC Offloading
![Page 54: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/54.jpg)
Roadmap
• Isn’t congestion control a solved problem? – Conflicting goals of high throughput and friendliness decoupled through priorities
• Scope for performance gains – Increases with increasing RTTxBW
• Design Details – Additional packets sent backwards from the end using multiple low priority levels
• Simulation Results – 40-80% reduction in FCT over baseline TCP implementation
• Linux Implementation and Evaluation – Simple modifications, agnostic to the underlying congestion control algorithm
• Challenges and Future
![Page 55: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/55.jpg)
Where RC3 is of little help…
• Low delay bandwidth product
• Very heavily utilized links
• Small queue buffer size at the bottleneck
• Application pacing
![Page 56: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/56.jpg)
Deployment Concerns
• Partial Priorities Support
• Middleboxes [Honda et. al. 2011, Flach et al 2013]
• Wireless
![Page 57: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/57.jpg)
Future
• Performance gains increase with BWxRTT – Likely to increase with time
• Futuristic datacenter bandwidth of 100Gbps – 45% reduction in average FCT (over flows) – 66% reduction in average FCT (over bytes)
![Page 58: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/58.jpg)
Summary
• Send additional packets from a flow using several layers of low priority service
• Uses only the spare capacity in the network without affecting the regular traffic
• Gives 40-80% reduction in FCTs over baseline TCP
![Page 59: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/59.jpg)
Thank you!
• Send additional packets from a flow using several layers of low priority service
• Uses only the spare capacity in the network without affecting the regular traffic
• Gives 40-80% reduction in FCTs over baseline TCP
http://netsys.github.io/RC3/
![Page 60: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/60.jpg)
Back Up!
![Page 61: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/61.jpg)
What about dropped low priority packets?
• Low priority packets are transmitted only once • Losses recovered by TCP control loop
• SACK indicates which segments are missing (optional)
![Page 62: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/62.jpg)
Throughput
![Page 63: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/63.jpg)
Varying Link Load
![Page 64: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/64.jpg)
Varying RTTxBW
![Page 65: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/65.jpg)
Drop Rates in Baseline Simulations
Network
Load Drop % (Regular
TCP)
Drop % (RC3)
High Priority
Low Priority
Total
0.3 0.5 0.3 13.15 13.45
0.5 0.84 0.58 28.46 29.04
0.7 1.42 1.09 36.84 37.94
![Page 66: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/66.jpg)
Varying Loss Rates
![Page 67: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/67.jpg)
Topologies
![Page 68: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/68.jpg)
Workload
![Page 69: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/69.jpg)
Link Heterogeneity
![Page 70: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/70.jpg)
Application Pacing
![Page 71: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/71.jpg)
Priority Levels
![Page 72: Radhika Mittal, Justine Sherry, Sylvia Ratnasamy, …...Short Flow Completion Time " • “Being fast really matters. Users really respond to speed.”" " – 0.5 sec delay caused](https://reader034.vdocument.in/reader034/viewer/2022050117/5f4e4541984d9c4aa66df9ef/html5/thumbnails/72.jpg)
Some Queues DropTail