Internet Engineering Task Force Mark Allman INTERNET DRAFT ICIR File: draft-allman-tcp-early-rexmt-03.txt Konstantin Avrachenkov INRIA Urtzi Ayesta France Telecom R&D Josh Blanton Ohio University December, 2003 Expires: June, 2004 Early Retransmit for TCP and SCTP Status of this Memo This document is an Internet-Draft and is in full conformance with all provisions of Section 10 of [RFC2026]. Internet-Drafts are working documents of the Internet Engineering Task Force (IETF), its areas, and its working groups. Note that other groups may also distribute working documents as Internet-Drafts. Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress." The list of current Internet-Drafts can be accessed at http://www.ietf.org/ietf/1id-abstracts.txt The list of Internet-Draft Shadow Directories can be accessed at http://www.ietf.org/shadow.html. Abstract This document proposes a new mechanism for TCP and SCTP that can be used to more effectively recover lost segments when a connection's congestion window is small. The "Early Retransmit" mechanism allows the transport to reduce, in certain special circumstances, the number of duplicate acknowledgments required to trigger a fast retransmission. This allows the transport to use fast retransmit to recover packet losses that would otherwise require a lengthy retransmission timeout. Terminology The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119]. 1 Introduction A number of researchers have pointed out that the loss recovery Expires: June 2004 [Page 1]
draft-allman-tcp-early-rexmt-03.txt December 2003 strategies employed by TCP [RFC793] and SCTP [RFC2960] do not work well when the amount of outstanding data a TCP sender has injected into the network is small. This can happen in a number of situations, such as: (1) The connection is "application limited" and has only a limited amount of data to send. This can happen any time the application does not produce enough data to fill the congestion window. A particular case when all connections become application limited is as the connection ends. (2) The connection is limited by the receiver-advertised window. (3) The connection is constrained by end-to-end congestion control when the connection's share of the path is small, the path has a small bandwidth-delay product or the transport is ascertaining the available bandwidth in the first few round-trip times of slow start. Many researchers have studied problems with TCP when the congestion window is small and have outlined possible mechanisms to mitigate these problems [Mor97,BPS+98,Bal98,LK98,RFC3150,AA02]. SCTP's loss recovery and congestion control mechanisms are based on TCP and therefore the same problems impact the performance of SCTP connections. When the transport detects a missing segment, the connection enters a loss recovery phase using one of two methods. First, if an acknowledgment (ACK) for a given segment is not received in a certain amount of time a retransmission timer fires and the segment is resent [RFC2988]. Second, the ``Fast Retransmit'' algorithm resends a segment when three duplicate ACKs arrive at the sender [Jac88,RFC2581]. However, because duplicate ACKs from the receiver are also triggered by packet reordering in the Internet, the sender waits for three duplicate ACKs in an attempt to disambiguate segment loss from packet reordering. When using small windows it may not be possible to generate the required number of duplicate ACKs to trigger Fast Retransmit when a loss does happen. Once in a loss recovery phase, a number of techniques can be used to retransmit lost segments. TCP can use slow start based recovery or Fast Recovery [RFC2581], NewReno [RFC2582], and loss recovery based on selective acknowledgments (SACKs) [RFC2018,FF96,RFC3517]. SCTP's loss recovery is not as varied due to the built-in selective acknowledgments. The transport's retransmission timeout (RTO) is based on measured round-trip times (RTT) between the sender and receiver, as specified in [RFC2988] (for TCP) and [RFC2960] (for SCTP). To prevent spurious retransmissions of segments that are only delayed and not lost, the minimum RTO is conservatively chosen to be 1 second. Therefore, it behooves TCP senders to detect and recover from as many losses as possible without incurring a lengthy timeout during which the connection remains idle. However, if not enough duplicate ACKs arrive from the receiver, the Fast Retransmit algorithm is Expires: June 2004 [Page 2]
draft-allman-tcp-early-rexmt-03.txt December 2003 never triggered---this situation occurs when the congestion window is small, if a large number of segments in a window are lost or at the end of a transfer as data drains from the network. For instance, consider a congestion window (cwnd) of three segments. If one segment is dropped by the network, then at most two duplicate ACKs will arrive at the sender, assuming no ACK loss. Since three duplicate ACKs are required to trigger Fast Retransmit, a timeout will be required to resend the dropped packet. [BPS+98] shows that roughly 56% of retransmissions sent by a busy web server are sent after the RTO timer expires, while only 44% are handled by Fast Retransmit. In addition, only 4% of the RTO timer-based retransmissions could have been avoided with SACK, which has to continue to disambiguate reordering from genuine loss. Furthermore, [All00] shows that for one particular web server the median transfer size is less than four segments, indicating that more than half of the connections will be forced to rely on the RTO timer to recover from any losses that occur. Thus, loss recovery without relying on the conservative RTO is beneficial for short TCP transfers. The Limited Transmit mechanism introduced in [RFC3042] allows a TCP sender to transmit previously unsent data upon the reception of each of the two duplicate ACKs that precede a fast retransmit. SCTP [RFC2960] uses SACK information to calculate the number of outstanding segments in the network. Hence, when the first two duplicate ACKs arrive at the sender they will indicate that data has left the network and allow the sender to transmit new data (if available) similar to TCP's Limited Transmit algorithm. By sending these two new segments the TCP sender is attempting to induce additional duplicate ACKs (if appropriate) so that Fast Retransmit will be triggered before the retransmission timeout expires. The "Early Retransmit" mechanism outlined in this document covers the case when previously unsent data is not available for transmission. Section 2 of this document outlines a small change to TCP and SCTP senders that will decrease the reliance on the retransmission timer, and thereby improve performance when Fast Retransmit cannot otherwise be triggered. Section 3 discusses related work. Section 4 sketches security issues. 2 Reduction of the Retransmission Threshold The Early Retransmit algorithm calls for lowering the threshold for triggering Fast Retransmit when the amount of outstanding data is small and when no unsent data segments are enqueued. We define variants of Early Retransmit for connections that do and do not support selective acknowledgments (SACK) [RFC2018]. (Note: SCTP includes SACK in the base protocol and so there is no need for the non-SACK variant of Early Retransmit in SCTP.) If the following two conditions hold the sender can use Early Expires: June 2004 [Page 3]
draft-allman-tcp-early-rexmt-03.txt December 2003 Retransmit (regardless of SACK support). (2.a) The amount of outstanding data (ownd) is less than 4*SMSS bytes. (2.b) There is either no unsent data ready for transmission at the sender or the advertised window does not permit new segments to be transmitted. When the above two conditions hold and the connection does not support SACK the duplicate ACK threshold used to trigger Fast Retransmit MAY be reduced to: ER_thresh = ceiling (ownd/SMSS) - 1 (1) duplicate ACKs, where ownd is in terms of bytes. When conditions (2.a) and (2.b) hold and the connection does support SACK Fast Retransmit MAY be used when ownd - SMSS bytes have been SACKed. In other words, when ownd is small enough that losing one segment would not trigger Fast Retransmit, the trigger for Fast Retransmit is reduced to receiving indications that all but one segment has arrived at the receiver. This mitigation is less robust in the face of reordered segments than the standard Fast Retransmit threshold. Research shows that a general reduction in the number of duplicate ACKs required to trigger fast retransmission of a segment to two (rather than three) leads to a reduction in the ratio of good to bad retransmits by a factor of three [Pax97]. However, this analysis did not include the additional conditioning on the event that the ownd was smaller than 4 segments. The SACK variant of the Early Retransmit algorithm is preferred to the non-SACK variant due to its robustness in the face of ACK loss (since SACKs are sent redundantly) and due to interactions with the delayed ACK timer. Consider a flight of three segments, S1...S3, with S2 being dropped by the network. When S1 arrives it is in-order and so the receiver may or may not delay the ACK, leading to two scenarios: (A) The ACK for S1 is delayed. In this case the arrival of S3 will trigger an ACK to be transmitted covering segment S1 (which was previously unacknowledged). In this case Early Retransmit without SACK will not prevent an RTO because no duplicate ACKs will arrive. However, with SACK the ACK for S1 will also include SACK information indicating that S3 has arrived at the receiver. The sender can then invoke Fast Retransmit on this ACK because ownd - SMSS bytes have been SACKed when the ACK arrives. (B) The ACK for S1 is not delayed. In this case the arrival of S1 triggers an ACK and the arrival of S3 triggers a second ACK (because it is out-of-order). Both ACKs will cover the same Expires: June 2004 [Page 4]
draft-allman-tcp-early-rexmt-03.txt December 2003 segment (S1). Therefore, regardless of whether SACK is used Early Retransmit can be performed by the sender (assuming no ACK loss). We note two "worst case" scenarios for Early Retransmit: (1) Persistent reordering of segments, coupled with an application that does not constantly send data, can result in large numbers of needless retransmissions when using Early Retransmit. For instance, consider an application that sends data two segments at a time, followed by an idle period when no data is queued for delivery by TCP. If the network consistently reorders the two segments, the sender will needlessly retransmit one out of every two unique segments transmitted (and one-third of all segments) when using the above algorithm. However, this would only be a problem for long-lived connections from applications that transmit in spurts. (2) Similar to the above, consider the case of 2 segment transfers that always experience reordering. Just as in (1) above, one out of every two unique data segments will be retransmitted needlessly, therefore one-third of the traffic will be spurious. Currently this document offers no suggestion on how to mitigate the above problems. Rather, the authors believe that the community's consensus is that Early Retransmit is scoped enough that the worst case problems are pathological and do not need mitigation at this time. However, Appendix A offers a survey of possible mitigations. 3 Related Work Deployment of Explicit Congestion Notification (ECN) [Flo94,RFC3168] may benefit connections with small congestion window sizes [RFC2884]. ECN provides a method for indicating congestion to the end-host without dropping segments. While some segment drops may still occur, ECN may allow TCP to perform better with small cwnd sizes because the sender will be required to detect less segment loss [RFC2884]. [Bal98] outlines another solution to the problem of having no new segments to transmit into the network when the first two duplicate ACKs arrive. In response to these duplicate ACKs, a TCP sender transmits zero-byte segments to induce additional duplicate ACKs. This method preserves the robustness of the standard Fast Retransmit algorithm at the cost of injecting segments into the network that do not deliver any data (and, therefore are potentially wasting network resources). 4 Security Considerations The security considerations found in [RFC2581] apply to this document. No additional security problems have been identified with Early Retransmit at this time. Expires: June 2004 [Page 5]
draft-allman-tcp-early-rexmt-03.txt December 2003 Acknowledgments We thank Sally Floyd for her feedback in discussions about Early Retransmit. We also thank Sally Floyd and Hari Balakrishnan who helped with a large portion of the text of this document when it was part of a separate document. Armando Caro and many members of the tsvwg mailing list provided good discussions that helped shape this document. Normative References [RFC793] Jon Postel. Transmission Control Protocol. Std 7, RFC 793. September 1981. [RFC2018] Matt Mathis, Jamshid Mahdavi, Sally Floyd, Allyn Romanow. TCP Selective Acknowledgement Options. RFC 2018, October 1996. [RFC2581] Mark Allman, Vern Paxson, W. Richard Stevens. TCP Congestion Control. RFC 2581, April 1999. [RFC2883] Sally Floyd, Jamshid Mahdavi, Matt Mathis, Matt Podolsky. An Extension to the Selective Acknowledgement (SACK) Option for TCP. RFC 2883, July 2000. [RFC2960] R. Stewart, Q. Xie, K. Morneault, C. Sharp, H. Schwarzbauer, T. Taylor, I. Rytina, M. Kalla, L. Zhang, V. Paxson. Stream Control Transmission Protocol. October 2000. [RFC2988] Vern Paxson, Mark Allman. Computing TCP's Retransmission Timer. RFC 2988, April 2000. [RFC3042] Mark Allman, Hari Balakrishnan, Sally Floyd. Enhancing TCP's Loss Recovery Using Limited Transmit. RFC 3042, January 2001. [RFC3522] Reiner Ludwig, Michael Meyer. The Eifel Detection Algorithm for TCP. RFC 3522, April 2003. Informative References [AA02] Urtzi Ayesta, Konstantin Avrachenkov, "The Effect of the Initial Window Size and Limited Transmit Algorithm on the Transient Behavior of TCP Transfers", In Proc. of the 15th ITC Internet Specialist Seminar, Wurzburg, July 2002. [All00] Mark Allman. A Server-Side View of WWW Characteristics. ACM Computer Communications Review, October 2000. [Bal98] Hari Balakrishnan. Challenges to Reliable Data Transport over Heterogeneous Wireless Networks. Ph.D. Thesis, University of California at Berkeley, August 1998. [BPS+98] Hari Balakrishnan, Venkata Padmanabhan, Srinivasan Seshan, Mark Stemm, and Randy Katz. TCP Behavior of a Busy Web Server: Expires: June 2004 [Page 6]
draft-allman-tcp-early-rexmt-03.txt December 2003 Analysis and Improvements. Proc. IEEE INFOCOM Conf., San Francisco, CA, March 1998. [FF96] Kevin Fall, Sally Floyd. Simulation-based Comparisons of Tahoe, Reno, and SACK TCP. ACM Computer Communication Review, July 1996. [Flo94] Sally Floyd. TCP and Explicit Congestion Notification. ACM Computer Communication Review, October 1994. [Jac88] Van Jacobson. Congestion Avoidance and Control. ACM SIGCOMM 1988. [LK98] Dong Lin, H.T. Kung. TCP Fast Recovery Strategies: Analysis and Improvements. Proceedings of InfoCom, March 1998. [Mor97] Robert Morris. TCP Behavior with Many Flows. Proceedings of the Fifth IEEE International Conference on Network Protocols. October 1997. [Pax97] Vern Paxson. End-to-End Internet Packet Dynamics. ACM SIGCOMM, September 1997. [RFC2582] Sally Floyd, Tom Henderson. The NewReno Modification to TCP's Fast Recovery Algorithm. RFC 2582, April 1999. [RFC2884] Jamal Hadi Salim and Uvaiz Ahmed. Performance Evaluation of Explicit Congestion Notification (ECN) in IP Networks. RFC 2884, July 2000. [RFC3150] Spencer Dawkins, Gabriel Montenegro, Markku Kojo, Vincent Magret. End-to-end Performance Implications of Slow Links. RFC 3150, July 2001. [RFC3168] K. K. Ramakrishnan, Sally Floyd, David Black. The Addition of Explicit Congestion Notification (ECN) to IP. RFC 3168, September 2001. [RFC3517] Ethan Blanton, Mark Allman, Kevin Fall, Lili Wang. A Conservative Selective Acknowledgment (SACK)-based Loss Recovery Algorithm for TCP. RFC 3517, April 2003. Author's Addresses: Mark Allman ICSI Center for Internet Research (ICIR) 1947 Center Street, Suite 600 Berkeley, CA 94704-1198 Phone: 216-243-7361 mallman@icir.org http://www.icir.org/mallman/ Konstantin Avrachenkov INRIA Expires: June 2004 [Page 7]
draft-allman-tcp-early-rexmt-03.txt December 2003 2004 route des Lucioles, B.P.93 06902, Sophia Antipolis France Phone: 00 33 492 38 7751 Email: k.avrachenkov@sophia.inria.fr http://www.inria.fr/mistral/personnel/K.Avrachenkov/moi.html Urtzi Ayesta France Telecom R&D 905 rue Albert Einstein 06921 Sophia Antipolis France Email: Urtzi.Ayesta@francetelecom.com http://www.inria.fr/mistral/personnel/Urtzi.Ayesta/me.html Josh Blanton Ohio University 301 Stocker Center Athens, OH 45701 jblanton@irg.cs.ohiou.edu Appendix A: Research Issues in Adjusting the Duplicate ACK Threshold Decreasing the number of duplicate ACKs required to trigger Fast Retransmit, as suggested in section 2, has the drawback of making Fast Retransmit less robust in the face of minor network reordering. Two egregious examples of problems caused by reordering are given in section 2. This appendix outlines several schemes that have been suggested to mitigate the problems caused to Early Retransmit by reordering. These methods need further research before they are suggested for general use (and, current consensus is that the cases that make Early Retransmit unnecessarily retransmit a large amount of data are patalogical and therefore these mitigations are not generally required). MITIGATION A.1: Allow a connection to use Early Retransmit as long as the algorithm is not injecting a "too much" spurious data into the network. For instance, using the information provided by TCP's DSACK option [RFC2883] or SCTP's Duplicate-TSN notification, a sender can determine when segments sent via Early Retransmit are needless. Likewise, using Eifel [RFC3522] the sender can detect spurious Early Retransmits. Once spurious Early Retransmits are detected the sender can either eliminate the use of Early Retransmit or limit the use of the algorithm to ensure that an acceptably small fraction of the connection's transmissions are not spurious. Alternatively, if a sender cannot reliably determine if an Early Retransmitted segment is spurious or not the sender could simply limit Early Retransmits either to some fixed number per connection (e.g., Early Retransmit is allowed only once per connection) or to some small percentage of the total traffic being transmitted. MITIGATION A.2: Allow a connection to trigger Early Retransmit using the criteria given in section 2, in addition to a "small" timeout Expires: June 2004 [Page 8]
draft-allman-tcp-early-rexmt-03.txt December 2003 [Pax97]. For instance, a sender may have to wait for 2 duplicate ACKs and then T msec before Early Retransmitting a segment. The added time gives reordered acknowledgments time to arrive at the sender and avoid a needless retransmit. Designing a method for choosing an appropriate timeout is part of the research that would need to be involved in this scheme. Full Copyright Statement Copyright (C) The Internet Society (2003). All Rights Reserved. This document and translations of it may be copied and furnished to others, and derivative works that comment on or otherwise explain it or assist in its implementation may be prepared, copied, published and distributed, in whole or in part, without restriction of any kind, provided that the above copyright notice and this paragraph are included on all such copies and derivative works. However, this document itself may not be modified in any way, such as by removing the copyright notice or references to the Internet Society or other Internet organizations, except as needed for the purpose of developing Internet standards in which case the procedures for copyrights defined in the Internet Standards process must be followed, or as required to translate it into languages other than English. The limited permissions granted above are perpetual and will not be revoked by the Internet Society or its successors or assigns. This document and the information contained herein is provided on an "AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Expires: June 2004 [Page 9]
