1 / 24

CA-RTO: A Contention-Adaptive Retransmission Timeout

CA-RTO: A Contention-Adaptive Retransmission Timeout. I. Psaras, V. Tsaoussidis, L. Mamatas Demokritos University of Thrace, Xanthi, Greece. This study was presented in the International Conference on Computer Communications and Networks (ICCCN 2005). Contributions of this work.

josie
Download Presentation

CA-RTO: A Contention-Adaptive Retransmission Timeout

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. CA-RTO: A Contention-Adaptive Retransmission Timeout I. Psaras, V. Tsaoussidis, L. Mamatas Demokritos University of Thrace, Xanthi, Greece This study was presented in the International Conference on Computer Communications and Networks (ICCCN 2005)

  2. Contributions of this work • Our perspective: • When contention increases, the timeout becomes the scheduler for the link • Our observations: • When contention increases, timeout decreases! • Congestion events cause synchronization, at least to some extend. • Our solutions: • We integrate a contention adaptive parameter into the timeout algorithm • Unnecessary retransmissions are reduced • We introduce randomness to avoid synchronization COMputer NETworks Group (COMNET)

  3. Research Steps • We investigated the timeout behavior towards various types of contention • We observed that when contention increases, RTO undertakes the role of the transmission scheduler for the link • We focused on scenarios with high contention • RTT can not always capture efficiently the network conditions • RTO is a bad scheduler for the link • We fixed the timeout value to correspond to different contention levels • We concluded that different levels of contention call for distinct timeout adjustments • We ran simulation results COMputer NETworks Group (COMNET)

  4. TCP Retransmission Timeout • The timer is adaptive to varying delay • The timeout is calculated every RTT • Given a sample RTT measurement M and the history of average RTT A, • the distance from the average is measured: • Diff = M – A, • the average is updated: • A = A + gDiff (where g=0.125), • the RTT Deviation is calculated: • Dev = Dev +d(|Diff| – Dev) (where d = 0.25) • and finally, the RTO value is adjusted to: • RTO = A + 4D COMputer NETworks Group (COMNET)

  5. Observations: Anomalies of the TCP Retransmission Timer • SCENARIO Setup • Dumbbell Topology • DropTail: 50 pkts • BxD = 10 packets • Contention Increase • 0 – 250s: 1 flow • 250 – 500s: 100 flows COMputer NETworks Group (COMNET)

  6. Contention Grows, Timeout Shrinks Contention Increase Scenario RTT RTO = A + 4D COMputer NETworks Group (COMNET)

  7. Contention Grows, Timeout Shrinks Contention Increase Scenario RTO = A +4 D COMputer NETworks Group (COMNET)

  8. Contention Grows, Timeout Shrinks COMputer NETworks Group (COMNET)

  9. Contention Grows, Timeout Shrinks • In this case: 4D  0 • Hence, RTO = A • Timeout Decreases instead of Increasing (there is no Deviation) • Smoothed RTT (A) does not differentiate between different flows • Synchronization is possible (although buffers are always full) • Fairness is not guaranteed COMputer NETworks Group (COMNET)

  10. The Proposed Algorithm (CA-RTO) • CA-RTO: RTO = A + 4D + c*p • We incorporate contention: • c = 1/cwnd_ • cont_diff_ = max_cwnd_ - cwnd_ • cont_diff_ = cont_diff_ / 100 • We introduce retransmission randomness: • p = Random(0, cont_diff_) • CA-RTO: RTO = A + 4D + c*p COMputer NETworks Group (COMNET)

  11. Behavior of theProposed Algorithm (CA-RTO) • The max_cwnd_ ever reached is 200pkts • Step 1: c = 1/cwnd_ • Step 2: cont_diff_ = max_cwnd_ - cwnd_ cont_diff = cont_diff_/100 COMputer NETworks Group (COMNET)

  12. Behavior of theProposed Algorithm (CA-RTO) • Step 3: p = Random(0, cont_diff_) • Finally: CA-RTO = RTO + c*p COMputer NETworks Group (COMNET)

  13. Behavior of theProposed Algorithm (CA-RTO) Contention Increase Scenario CA-RTO = RTO + c*p COMputer NETworks Group (COMNET)

  14. Behavior of theProposed Algorithm (CA-RTO) Small congestion window: • gives big value to parameter c c = 1/cwnd_ • may result in big cont_diff_ cont_diff_ = max_cwnd_ - cwnd_ Big congestion window: • gives small value to parameter c • results in small cont_diff_ We try to capture high contention We get aware of dynamic network conditions, e.g. contention increase We do not want to affect RTO’s performance in low contention scenaria COMputer NETworks Group (COMNET)

  15. Possible Further Enhancements • The algorithm may “punish” flows with small windows by extending their RTO value (e.g. during startup)  maybe a parameter indicating the history of cwnd_ has to be integrated instead of the current cwnd_ • The current mechanism used to capture contention may not be very accurate • The randomization factor used to split flows in time can be further improved COMputer NETworks Group (COMNET)

  16. Evaluation Methodology • We target high contented links/networks • We simulate large numbers of flows transmitting in low capacity channels Hence: • Fair-share is small • Flows are operating with small windows • Buffers are always full COMputer NETworks Group (COMNET)

  17. Evaluation Methodology • We use the dumbbell network topology • Bandwidth x Delay = 10 or 100 packets • We use both DropTail and RED queuing policies • We implement CA-RTO in TCP-Reno and compare the two versions • We measure Goodput, Throughput, Fairness and Number of Retransmitted Packets COMputer NETworks Group (COMNET)

  18. Experimental ResultsScenario 1 • B x D = 10 packets • Buffer size = 50 packets, DropTail Fairness Retransmitted Packets (up to 0.2 Index Points) (up to 4000 less retransmissions) COMputer NETworks Group (COMNET)

  19. Experimental ResultsScenario 1 Goodput (in Bps) Throughput (in Bps) COMputer NETworks Group (COMNET)

  20. Experimental ResultsScenario 2 • B x D = 100 packets • Buffer size = 100 packets, DropTail Fairness Retransmitted Packets (at least 0.15 Index Points) (25 % less retransmissions, 4500pkts) COMputer NETworks Group (COMNET)

  21. Experimental ResultsScenario 3 • bw_bb = 100Mbps, B x D = 250 pkts • Buffer size = 100 packets, DropTail • CA-RTO affects TCP’s performance only if needed… Goodput COMputer NETworks Group (COMNET)

  22. Experimental ResultsScenario 4 • B x D = 10 packets • Buffer size = 50 packets, DropTail • Packet Error Rate: 10% Fairness Index Retransmitted Packets COMputer NETworks Group (COMNET)

  23. Conclusions • Current RTO presents some behavioral anomalies when contention increases • A Contention-Adaptive RTO proves to be more efficient in terms of successful retransmissions. That calls for further investigation of the energy potential of CA-RTO • A Randomization Factor in the RTO schedules the participating flows in a more fair manner COMputer NETworks Group (COMNET)

  24. CA-RTO: A Contention-Adaptive Retransmission Timeout Thank you!!  COMputer NETworks Group (COMNET)

More Related