WB-RTO: A Window-Based Retransmission Timeout Ioannis Psaras, Vassilis Tsaoussidis Demokritos University of Thrace, Xanthi, Greece
Motivation and Contribution We observe that retransmission scheduling affects transmission scheduling WB-RTO results in: 50% less retransmitted pkts Higher Goodput, and Better Fairness than TCP-RTO TCP-RTO WB-RTO COMputer NETworks Group (COMNET) 2
Motivation-Contribution Perspective: When contention increases, the timeout becomes the scheduler for the link. Motivation: When contention is high, all flows measure similar RTTs. TCP-RTO should not be solely based on RTT measurements. Congestion events cause retransmission synchronization. Algorithm: Approximation of the current level of network contention Estimation of the contribution of each flow to congestion Allowance for asynchronous retransmissions when timeout happens. COMputer NETworks Group (COMNET) 3
Motivation Queuing Policy: DropTail DBP = Buffer Size = 10 pkts 5 participating flows 1500 sec total simulation time Flows ideal rate = 2 pkts/wnd We trace: Seqno progress, RTT, RTO Sequence Number Dumbbell Network Topology RTT (in secs) COMputer NETworks Group (COMNET) 4
Motivation For the TCP-RTO performance, we observe: RTT stabilization Similar timeout values RTO (in secs) ~3000 retransmitted packets in 1500 seconds Un-Fairness We investigate the impact of retransmission synchronization on system behavior (e.g. overhead) TCP Performance COMputer NETworks Group (COMNET) 5
Outline of the presentation The current Retransmission Timeout Algorithm Recent Related Work The proposed algorithm: WB-RTO The algorithm Expected Behavior Evaluation Plan Experimental Results COMputer NETworks Group (COMNET) 6
The current Retransmission Timeout Algorithm Upon each ACK arrival, the sender: Calculates the RTT Variation: Updates the expected RTT prior to calculating the timeout: Calculates the Retransmission Timeout value: COMputer NETworks Group (COMNET) 7
Recent Related Work Eifel RTO Algorithm Uses the timestamp option to detect a spurious timeout. Forward RTO Algorithm Uses the first 3 ACKs after the timeout to decide if the timeout was spurious or not. Peak-Hopper RTO Algorithm Uses 2 timers: one is aggressive and one is conservative. Each time it decides which one to follow. CA-RTO: A Contention-Adaptive RTO Integrates a contention-adaptive parameter and introduces random retransmission scheduling. COMputer NETworks Group (COMNET) 8
The 3 stages of WB-RTO 1. Contribution to Congestion penalty charge 2. Estimation of Contention determine the scale of possible values 3. Calculation of Timeout COMputer NETworks Group (COMNET) 9
Window-Based RTO (1/4): Proportional Timeout Estimation of the contribution of the flow to congestion: c = f(cwnd, max cwnd ) Compare the current cwnd_ with the max_cwnd_: If cwnd_ < max_cwnd_ / 2, c = 1: minimal charge If max_cwnd_ / 2 < cwnd_ < (3/4)* max_cwnd_, c = 1,5: medium charge If (3/4)* max_cwnd_ < cwnd_ < max_cwnd_, c = 2: major charge COMputer NETworks Group (COMNET) 10
Window-Based RTO (2/4): Contention Estimation Flow classification according to its cwnd_ history (awnd_): ai = g(awnd, Thresholdi) where, awnd=average window, Thresholds 1 to 4 represent different levels of network contention: Threshold 1 corresponds to very high contention Threshold 4 corresponds to low contention Example: 1. awnd_ < 5: a1 = 10 2. 5 < awnd_ < 10: a2 = 5 3. 10 < awnd_ < 30: a3 = 3 4. 30 < awnd_ < 50: a4 = 2 COMputer NETworks Group (COMNET) 11
Adjust the scale COMputer NETworks Group (COMNET) 12
Window-Based RTO (3/4): Timeout Adjustment Calculation of the Window-Based RTO: or WB RTO = random(rtt, c ai) WB RTO = random(rtt, f(cwnd, max cwnd ) g(awnd, Thresholdi)) 7. rtt, to avoid timeout expiration prior to the estimated RTT measurement 8. Parameter c captures the contribution of the flow to congestion 9. Parameter a approximates the current level of flow contention 10. Randomization guarantees asynchronous retransmission attempts COMputer NETworks Group (COMNET) 13
Window-Based RTO (4/4): Expected Behavior High penalties result in high timeout values. As awnd_ increases timeout settles to smaller values. So, Large windows do not always mean large timeout values. WB-RTO vs awnd_ COMputer NETworks Group (COMNET) 14
Performance Evaluation Plan WB-RTO is implemented in TCP-Reno Evaluation Scenarios Motivation Part Scenario Scenario 1: Standard/Proposed Parameters Scenario 2: Modified Parameters COMputer NETworks Group (COMNET) 15
An Important Note WB-RTO does not improve the Goodput performance of TCP significantly We focus on concurrent Retransmissions hence we pay more attention on the combination of the retransmission effort and the Goodput performance, rather than on the Goodput performance alone. COMputer NETworks Group (COMNET) 16
Scenario 1 Queuing Policy: DropTail Bottleneck BW = 10Mbps Bottleneck Delay = 10ms Buffer Size = 50 pkts 1500 sec total simulation time Goodput (in B/s) Retransmitted Packets COMputer NETworks Group (COMNET) 17
Scenario 1 Contention grows->wb-rto allows for better multiplexing Behavior Captured by fairness index More flows are getting service Fairness Goodput per Flow (B/s) COMputer NETworks Group (COMNET) 18
Scenario 2 Queuing Policy: DropTail Bottleneck BW = 10Mbps Bottleneck Delay = 10ms Buffer Size = 50 pkts 1500 sec total simulation time Scale now different Goodput (in B/s) Retransmitted Packets COMputer NETworks Group (COMNET) 19
Scenario 2 Fairness initially drops when the scale is not adjusted appropriately Fairness Goodput per Flow (B/s) COMputer NETworks Group (COMNET) 20
Interaction with AQM (i.e. RED) (1) Topology: Dumbbell Queuing Policy: RED DBP = Buffer Size = 40 pkts min_thresh = 4 pkts max_thresh = 12 pkts 1500 sec total simulation time Goodput (in B/s) Number of Timeouts Retransmitted Packets COMputer NETworks Group (COMNET) 21
Interaction with AQM (i.e. RED) (2) We observe: Similar Goodput Significant difference in Retransmission Effort (50%) WB-RTO results in 66% less timeout expirations TCP-RTO causes inefficient queue utilization The average queue length always overcomes the max_thresh, when using TCP- RTO TCP-RTO WB-RTO COMputer NETworks Group (COMNET) 22
Satellite Scenario (1) Topology: Cross-Traffic Bottleneck Queuing Policy: RED The rest of the buffers use DT bw_bottleneck = 20Mbps bw_delay = 300ms Buffer Size = 200 pkts min_thresh = 20 pkts max_thresh = 60 pkts 150 sec total simulation time PER = 0,0001 No blackout 3 blackouts on the backbone link After 3 blackouts COMputer NETworks Group (COMNET) 23
Satellite Scenario (2) We observe that: TCP-RTO interprets the blackout as a congestion signal WB-RTO does not extend the timeout, due to low contention and hence exploits bandwidth faster TCP still waits for the extended timeout to expire, while WB-RTO resumes transmission immediately TCP-RTO WB-RTO COMputer NETworks Group (COMNET) 24
Traffic Diversity (Mice and Elephants) (1) Topology: Dumbbell Bottleneck Queuing Policy: RED bw_bottleneck = 10Mbps bw_delay = 30ms Buffer Size = 40 pkts Goodput (KB/s) Goodput per flow (KB/s) Retransmitted Packets COMputer NETworks Group (COMNET) 25
Traffic Diversity (Mice and Elephants) (2) We observe: Simultaneous timeout events for TCP-RTO All flows timeout during the Slow-Start Flows 7-9 timeout simultaneously 10 times during the experiment Short flows: 83 vs 50 timeouts Long flows: 43 vs 12 timeouts We conclude that: most of the timeouts are spurious WB-RTO achieves an important goal: it reduces the number of timeouts TCP-RTO WB-RTO COMputer NETworks Group (COMNET) 26
Conclusions RTT measurements cannot always reflect the level of network contention TCP-RTO should not be solely based on RTT samples A contention-aware RTO proves to be more efficient, since it is aware of current network conditions. A randomization factor in the RTO schedules retransmissions in a fairer manner WB-RTO cancels some of TCP miss-responses with noncongestion errors COMputer NETworks Group (COMNET) 27
Other References [1] CA-RTO: A Contention-Adaptive Retransmission Timeout for TCP, Ioannis Psaras, Vassilis Tsaoussidis, IEEE ICCCN 05 [2] Why TCP Timers (still) Don t Work Well, Ioannis Psaras, Vassilis Tsaoussidis, Computer Networks (COMNET), Elsevier Science, to appear 2007 http://comnet.ee.duth.gr/comnet/ COMputer NETworks Group (COMNET) 28