]> NetApp, Inc.

Sonnenallee 1 Kirchheim 85551 Germany +49 89 900594712 alexander.zimmermann@netapp.com

Dept. of Elec. Engg., Texas A&M University

214 ZACH College Station TX 77843-3128 sumitha@tamu.edu

Dept. of Elec. Engg., Texas A&M University

315C WERC College Station TX 77843-3128 reddy@ee.tamu.edu

ICSI Center for Internet Research

1947 Center Street, Suite 600 Berkeley CA 94704-1198 (440) 235-1792 mallman@icir.org

Purdue University Computer Science

305 N. University St. West Lafayette IN 47907 mallman@icir.org

Transport TCP Maintenance and Minor Extensions (TCPM) WG Transmission Control Protocol (TCP), Reordering, TCP-NCR This document specifies Non-Congestion Robustness (NCR) for TCP. In the absence of explicit congestion notification from the network, TCP uses loss as an indication of congestion. One of the ways TCP detects loss is using the arrival of three duplicate acknowledgments. However, this heuristic is not always correct, notably in the case when network paths reorder segments (for whatever reason), resulting in degraded performance. TCP-NCR is designed to mitigate this degraded performance by increasing the number of duplicate acknowledgments required to trigger loss recovery, based on the current state of the connection, in an effort to better disambiguate true segment loss from segment reordering. This document specifies the changes to TCP, as well as the costs and benefits of these modifications.

One strength of TCP lies in its ability to adjust its sending rate according to the perceived congestion in the network . In the absence of explicit notification of congestion from the network, TCP uses segment loss as an indication of congestion (i.e., assuming queue overflow). TCP receivers send cumulative acknowledgments (ACKs) indicating the next sequence number expected from the sender for arriving segments . When segments arrive out of order, duplicate ACKs are generated. As specified in , a TCP sender uses the arrival of three duplicate ACKs as an indication of segment loss. The TCP sender retransmits the lost segment and reduces the load imposed on the network, assuming the segment loss was caused by resource contention within the network path. The TCP sender does not assume loss on the first or second duplicate ACK, but waits for three duplicate ACKs to account for minor packet reordering. However, the use of this constant threshold of duplicate ACKs has several problems that can be mitigated with a dynamic threshold. The following is an example of TCP's behavior: TCP A is the data sender, and TCP B is the data receiver. TCP A sends 10 segments, each consisting of a single data byte (i.e., transmits bytes 1-10 in segments 1-10). Assume segment 3 is dropped in the network. TCP B cumulatively acknowledges segments 1 and 2, making the cumulative ACK transmitted to the sender 3 (the next expected sequence number). (Note: TCP B may generate one or two ACKs, depending on whether delayed ACKs are employed.) The arrival of segments 4-10 at TCP B will each trigger the transmission of a cumulative ACK for sequence number 3. (Note: recommends that delayed ACKs not be used when the ACK is triggered by an out-of-order segment.) When TCP A receives the third duplicate ACK (or fourth ACK overall) for sequence number 3, TCP A will retransmit segment 3 and reduce the sending rate by roughly half (see for specifics on the congestion control state adjustments). Alternatively, suppose segment 3 was not dropped by the network, but rather delayed such that segment 3 arrives at TCP B after segment 10. The above scenario will play out in precisely the same manner insomuch as a retransmission of segment 3 will be triggered. In other words, TCP is not capable of disambiguating this reordering event from a segment loss, resulting in an unnecessary retransmission and rate reduction. The following is the specific motivation behind making TCP robust to reordered segments: A number of Internet measurement studies have shown that packet reordering is not a rare phenomenon . Further, the reordering can be well beyond that required for fast retransmit to be falsely triggered. show the negative performance implications that packet reordering has on current TCP. The requirement imposed by TCP for almost in-order packet delivery places a constraint on the design of future technology. Novel routing algorithms, network components, link-layer retransmission mechanisms, and applications could all be looked at with a fresh perspective if TCP were to be more robust to segment reordering. For instance, high-speed packet switches could cause resequencing of packets if TCP were more robust. There has been work proposed in the literature explicitly to ensure that packet ordering is maintained in such switches (e.g., ). Also, link-layer mechanisms that attempt to recover from packet corruption by retransmitting could be allowed to reorder packets, and thus increase the chances of local loss repair rather than rely on TCP to repair the loss (and, needlessly reduce its sending rate). Additional examples include multi-path routing, high-delay satellite links, and some of the schemes proposed for a differentiated services architecture. By making TCP more robust to non-congestion events, TCP-NCR may open the design space of the future Internet components. In this document, we specify a set of TCP sender modifications to provide Non-Congestion Robustness (NCR) to TCP. In particular, these changes are built on top of TCP with selective acknowledgments (SACKs) and the SACK-based loss recovery scheme given in , since SACK is widely deployed at this point ( indicates that 68% of web servers and 88% of web clients utilize SACK as of spring 2004). Note that the TCP-NCR algorithm provided in this document could be easily adapted to SCTP since SCTP uses congestion control algorithms similar to TCP's (and thus has the same reordering robustness issues). As noted in several places in the remainder of this document, we consider TCP-NCR experimental in that more experience with the techniques is required before TCP-NCR should be used on a large scale on the Internet. We encourage implementation and experimentation with TCP-NCR in the hopes of gaining an understanding of its suitability for wide-scale deployment. The remainder of this document is organized as follows. provides a high-level description of the TCP-NCR mechanisms. In , we specify the TCP-NCR algorithm. provides a brief overview of the benefits of TCP-NCR, while discusses the drawbacks. discusses related work. discusses security concerns.

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 . Readers should be familiar with the TCP terminology (e.g., FlightSize, Pipe) given in and .

As discussed above, in the face of packet reordering, three duplicate ACKs may not be enough to disambiguate loss from reordering. In this section we provide a non-normative sketch of TCP-NCR. The detailed algorithms for implementing Non-Congestion Robustness for TCP are presented in the next section. The general idea behind TCP-NCR is to increase the threshold used to trigger a fast retransmission from the current fixed value of three duplicate ACKs to approximately a congestion window of data having left the network (but not less than the currently standardized value of three duplicate ACKs). Since cwnd represents the amount of data a TCP flow can transmit in one round-trip time (RTT), waiting to receive notice that cwnd bytes have left the network before deciding whether the root cause is loss or reordering imposes a delay of roughly one RTT on both the retransmission and the congestion control response. The appropriate choice for a new value of the threshold is essentially a trade-off between making the best decision regarding the cause of the duplicate ACKs and responsiveness. The choice to trigger a retransmission only after a cwnd's worth of data is known to have left the network represents roughly the largest amount of time a TCP can wait before the (often costly) retransmission timeout may be triggered. Therefore, the algorithm described in this document attempts to make the best decision possible at the expense of timeliness. Simply increasing the threshold before retransmitting a segment can make TCP brittle to packet loss or ACK loss since such loss reduces the number of duplicate ACKs that will arrive at the sender from the receiver. For instance, if the cwnd is 10 segments and one segment is lost, a duplicate ACK threshold of 10 will never be met because duplicate ACKs corresponding to at most 9 segments will arrive at the sender. To offset the issue of loss, we extend TCP's Limited Transmit scheme to allow for the sending of new data during the period when the TCP sender is disambiguating loss and reordering. This new data serves to increase the likelihood that enough duplicate ACKs arrive at the sender to trigger loss recovery if it is appropriate. Note that TCP tightly couples reliability and congestion control: when a segment is declared lost, a retransmission is triggered, and a change to the sending rate is also made on the assumption that the drop is due to resource contention . Therefore, simply by changing the retransmission trigger, the congestion control response is also changed. However, we lack experience on the Internet as to whether delaying the point that a rate reduction takes place is appropriate for wide-scale deployment. Therefore, the Extended Limited Transmit mechanism proposed in this document offers two variants for experimentation. The first Extended Limited Transmit variant, Careful Limited Transmit, calls for the transmission of one previously unsent segment, in response to duplicate acknowledgments, for every two segments that are known to have left the network. This effectively halves the sending rate, since normal TCP operation calls for the sending of one segment for every segment that has left the network. Further, the halving starts immediately and is not delayed until a retransmission is triggered. In the case of packet reordering (i.e., not segment loss), the congestion control state is restored to its previous state when reordering is determined. The second variant, Aggressive Limited Transmit, calls for transmitting one previously unsent data segment, in response to duplicate acknowledgments, for every segment known to have left the network. With this variant, while waiting to disambiguate the loss from a reordering event, ACK-clocked transmission continues at roughly the same rate as before the event started. Retransmission and the sending rate reduction happen per , albeit with the delayed threshold described above. Although this approach delays legitimate rate reductions (possibly slightly and temporarily aggravating overall congestion on the network), the scheme has the advantage of not reducing the transmission rate in the face of segment reordering. Which of the two Extended Limited Transmit variants is best for use on the Internet is an open question.

The TCP-NCR modifications make two fundamental changes to the way currently operates, as follows. First, the trigger for retransmitting a segment is changed from three duplicate ACKs to indications that a congestion window's worth of data has left the network. Second, TCP-NCR decouples initial congestion control decisions from retransmission decisions, in some cases delaying congestion control changes relative to TCP's current behavior as defined in . The algorithm provides two alternatives for extending Limited Transmit. The two variants of extended Limited Transmit are: This variant calls for reducing the sending rate at approximately the same time implementations reduce the congestion window, while at the same time withholding a retransmission (and the final congestion determination) for approximately one RTT. This variant calls for maintaining the sending rate in the face of duplicate ACKs until TCP concludes that a segment is lost and needs to be retransmitted (which TCP-NCR delays by one RTT when compared with current loss recovery schemes). A TCP-NCR implementation MUST use either Careful Limited Transmit or Aggressive Limited Transmit. A constant MUST be set, depending on which variant of extended Limited Transmit is used, as follows: LT_F = 2/3 LT_F = 1/2 This constant reflects the fraction of outstanding data (including data sent during Extended Limited Transmit) that must be SACKed before a retransmission is triggered. Since Aggressive Limited Transmit sends a new segment for every segment known to have left the network, a total of roughly cwnd segments will be sent during Aggressive Limited Transmit, and therefore ideally a total of roughly 2*cwnd segments will be outstanding when a retransmission is triggered. The duplicate ACK threshold is then set to LT_F = 1/2 of 2*cwnd (or about 1 RTT worth of data). The factor is different for Careful Limited Transmit because the sender only transmits one new segment for every two segments that are SACKed and therefore will ideally have a total of 1.5*cwnd segments outstanding when the retransmission is to be triggered. Hence, the required threshold is LT_F=2/3 of 1.5*cwnd to delay the retransmission by roughly 1 RTT. There are situations whereby the sender cannot transmit new data during Extended Limited Transmit (e.g., lack of data from the application, receiver's advertised window limit). These situations can lead to the problems discussed in the last section when a TCP does not employ Extended Limited Transmit and is starved for ACKs. Therefore, TCP-NCR adapts the duplicate ACK threshold on each SACK arrival to be as robust as possible given the actual amount of data that has been transmitted, or roughly LT_F times the number of outstanding segments. The TCP-NCR modifications specified in this document lend themselves to incremental deployment. Only the TCP implementation on the sender side requires modification (assuming both hosts support SACK). The changes themselves are modest. However, as will be discussed below, availability of additional buffer space at the receiver will help maximize the benefits of using TCP-NCR but is not strictly necessary. The following algorithms depend on the notions provided by , and we assume the reader is familiar with the terminology given in . The TCP-NCR algorithm can be adapted to alternate SACK- based loss recovery schemes. outline non-SACK-based algorithms; however, we do not specify those algorithms in this document and do not recommend them due to both the complexity and security implications of having only a gross understanding of the number of outstanding segments in the network. A TCP connection using the Nagle algorithm MAY employ the TCP-NCR algorithm. If a TCP implementation does implement TCP-NCR, the implementation MUST follow the various specifications provided in Sections to . If the Nagle algorithm is not being used, there is no way to accurately calculate the number of outstanding segments in the network (and, therefore, no good way to derive an appropriate duplicate ACK threshold) without adding state to the TCP sender. A TCP connection that does not employ the Nagle algorithm SHOULD NOT use TCP-NCR. We envision that NCR could be adapted to an implementation that carefully tracks the sequence numbers transmitted in each segment. However, we leave this as future work.

When entering a period of loss/reordering detection and Extended Limited Transmit, a TCP-NCR MUST initialize several state variables. A TCP MUST enter Extended Limited Transmit upon receiving the first ACK with a SACK block after the reception of an ACK that (a) did not contain SACK information and (b) did increase the connection's cumulative ACK point. The initializations are: The TCP MUST save the current FlightSize. FlightSizePrev = FlightSize The TCP MUST set a variable for tracking the number of segments for which an ACK does not trigger a transmission during Careful Limited Transmit. Skipped = 0 (Note: Skipped is not used during Aggressive Limited Transmit.) The TCP MUST set DupThresh (from ) based on the current FlightSize. DupThresh = max (LT_F * (FlightSize / SMSS),3) Note: We keep the lower bound of DupThresh = 3 from , . In addition to the above steps, the incoming ACK MUST be processed with the E series of steps in .

Extended Limited Transmit MUST be terminated at the start of loss recovery as outlined in . The arrival of an ACK that advances the cumulative ACK point while in Extended Limited Transmit, but before loss recovery is triggered, signals that a series of duplicate ACKs was caused by reordering and not congestion. Therefore, the receipt of an ACK that extends the cumulative ACK point MUST terminate Extended Limited Transmit. As described below (in (T.4)), an ACK that extends the cumulative ACK point and *also* contains SACK information will also trigger the beginning of a new Extended Limited Transmit phase. Upon the termination of Extended Limited Transmit, and especially when using the Careful variant, TCP-NCR may be in a situation where the entire cwnd is not being utilized, and therefore TCP-NCR will be prone to transmitting a burst of segments into the network. Therefore, to mitigate this bursting when a TCP-NCR in the Extended Limited Transmit phase receives an ACK that updates the cumulative ACK point (regardless of whether the ACK contains SACK information), the following steps MUST be taken: A TCP MUST reset cwnd to: cwnd = min (FlightSize + SMSS, FlightSizePrev) This step ensures that cwnd is not grossly larger than the amount of data outstanding, a situation that would cause a line rate burst. A TCP MUST set ssthresh to: ssthresh = FlightSizePrev This step provides TCP-NCR with a sense of "history". If step (T.1) reduces cwnd below FlightSizePrev, this step ensures that TCP-NCR will slow start back to the operating point in effect before Extended Limited Transmit. A TCP is now permitted to transmit previously unsent data as allowed by cwnd, FlightSize, application data availability, and the receiver's advertised window. When an incoming ACK extends the cumulative ACK point and also contains SACK information, the initializations in steps (I.2) and (I.3) from MUST be taken (but step (I.1) MUST NOT be executed) to re-start Extended Limited Transmit. In addition, the series of steps in (the "E" steps) MUST be taken.

On each ACK containing SACK information that arrives after TCP-NCR has entered the Extended Limited Transmit phase (as outlined in ) and before Extended Limited Transmit terminates, the sender MUST use the following procedure. The SetPipe () procedure from MUST be used to set the "pipe" variable (which represents the number of bytes still considered "in the network"). Note: the current value of DupThresh MUST be used by SetPipe () to produce an accurate assessment of the amount of data still considered in the network. If the comparison in equation (1), below, holds and there are SMSS bytes of previously unsent data available for transmission, then the sender MUST transmit one segment of SMSS bytes. (pipe + Skipped) <= (FlightSizePrev - SMSS) If the comparison in equation (1) does not hold or no new data can be transmitted (due to lack of data from the application or the advertised window limit), skip to step (E.6). Pipe MUST be incremented by SMSS bytes. If using Careful Limited Transmit, Skipped MUST be incremented by SMSS bytes to ensure that the next SMSS bytes of SACKed data processed does not trigger a Limited Transmit transmission (since the goal of Careful Limited Transmit is to send upon receipt of every second duplicate ACK). A TCP MUST return to step (E.2) to ensure that as many bytes as are appropriate are transmitted. This provides robustness to ACK loss that can be (largely) compensated for using SACK information. DupThresh MUST be reset via: DupThresh = max (LT_F * (FlightSize / SMSS),3) where FlightSize is the total number of bytes that have not been cumulatively acknowledged (which is different from "pipe").

When a segment is deemed lost via the algorithms in , Extended Limited Transmit MUST be terminated, leaving the algorithms in to govern TCP's behavior. One slight change to MUST be made, however. In Section 5, step (2) of MUST be changed to: ssthresh = cwnd = (FlightSizePrev / 2) This ensures that the congestion control modifications are made with respect to the amount of data in the network before FlightSize was increased by Extended Limited Transmit. Note: Once the algorithm in takes over from Extended Limited Transmit, the DupThresh value MUST be held constant until the loss recovery phase is terminated.

The major advantages of TCP-NCR are twofold. As discussed in Section 1, TCP-NCR will open up the design space for network applications and components that are currently constrained by TCP's lack of robustness to packet reordering. The second advantage is in terms of an increase in TCP performance. presents ns-2 simulations of a pre-cursor to the TCP-NCR algorithm specified in this document, called TCP-DCR (Delayed Congestion Response). The paper shows that TCP-DCR aids performance in comparison to unmodified TCP in the presence of packet reordering. In addition, the extended version of presents results based on emulations involving Linux (kernel 2.4.24). These results show that the performance of TCP-DCR is similar to Linux's native implementation that seeks to "undo" wrong decisions according to duplicate-SACK (DSACK) feedback (similar to the schemes outlined in ), when packets are reordered by less than one RTT. The advantage of using TCP-DCR over the DSACK-based scheme is that the DSACK-based scheme tries to estimate the exact amount of reordering in the network using fairly complex algorithms, whereas TCP-DCR achieves similar results with less complicated modifications. In addition, illustrate the ability of TCP-DCR to allow for the improvement of other parts of the system. For example, these papers show that increasing TCP's robustness to packet reordering allows a novel wireless ARQ mechanism to be added at the link-layer. The added robustness of the link-layer to channel errors, in turn, increases TCP performance by not requiring TCP to retransmit packets that were dropped due to corruption (and thus also prevents TCP from needlessly reducing the sending rate when retransmitting these segments).

Although all the changes outlined above are implemented in the sender, the receiver also potentially has a part to play. In particular, TCP-NCR increases the receiver's buffering requirement by up to an extra cwnd -- in the case of the TCP sender using Aggressive Limited Transmit and actual loss occurring in the network. Therefore, to maximize the benefits from TCP-NCR, receivers should advertise a large window to absorb the extra out-of-order traffic. In the case that the additional buffer requirements are not met, the use of the above algorithm takes into account the reduced advertised window -- with a corresponding loss in robustness to packet reordering. In addition, using TCP-NCR could delay the delivery of data to the application by up to one RTT because the fast retransmission point is delayed by roughly one RTT in TCP-NCR. Applications that are sensitive to such delays should turn off the TCP-NCR option. For instance, a socket option could be introduced to allow applications to control whether NCR would be used for a particular connection. Finally, the use of TCP-NCR makes the recovery from congestion events sluggish in comparison to the standard reaction in . show (via simulation) that the delay in congestion response has minimal impact on the connection itself and the traffic sharing a bottleneck. also indicates (again, via simulation) that "slowly responsive" congestion control may be safe for deployment in the Internet. These studies suggest that schemes that slightly delay congestion control decisions may be reasonable; however, further experimentation on the Internet is required to verify these results.

Over the past few years, several solutions have been proposed to improve the performance of TCP in the face of segment reordering. These schemes generally fall into one of two categories (with some overlap): mechanisms that try to prevent spurious retransmits from happening and mechanisms that try to detect spurious retransmits and "undo" the needless congestion control state changes that have been taken. attempt to prevent segment reordering from triggering spurious retransmits by using various algorithms to approximate the duplicate ACK threshold required to disambiguate loss and reordering over a given network path at a given time. TCP-NCR similarly tries to prevent spurious retransmits. However, TCP-NCR takes a simplified approach compared to those in , in that TCP-NCR simply delays retransmission by an amount based on the current cwnd (in comparison to standard TCP), while the other schemes use relatively complex algorithms in an attempt to derive a more precise value for DupThresh that depends on the current patterns of packet reordering. While TCP-NCR offers simplicity, the other schemes may offer more precision such that applications would not be forced to wait as long for their retransmissions. Future work could be undertaken to achieve robustness without needless delay. On the other hand, several schemes have been developed to detect and mitigate needless retransmissions after the fact. present algorithms to detect spurious retransmits and mitigate the changes these events made to the congestion control state. TCP-NCR could be used in conjunction with these algorithms, with TCP-NCR attempting to prevent spurious retransmits and some other scheme kicking in if the prevention failed. In addition, note that TCP-NCR is concentrated on preventing spurious fast retransmits; some of the above algorithms also attempt to detect and mitigate spurious timeout-based retransmits.

This memo includes no request to IANA.

General attacks against the congestion control of TCP are described in . SACK-based loss recovery for TCP mitigates some of the duplicate ACK attacks against TCP's congestion control. This document builds upon that work, and the Extended Limited Transmit algorithms specified in this document have been designed to thwart the ACK division problems that are described in .

Feedback from Lars Eggert, Ted Faber, Wesley Eddy, Gorry Fairhurst, Sally Floyd, Sara Landstrom, Nauzad Sadry, Pasi Sarolahti, Joe Touch, Nitin Vaidya, and the TCPM working group have contributed significantly to this document. Our thanks to all!

&rfc0793; &rfc2018; &rfc2119; &rfc3042; &rfc5681; &rfc6675; &rfc0896; &rfc1122; &rfc2883; &rfc2960; &rfc3465; &rfc3522; &rfc3708; &rfc4015; &rfc5682;

This appendix should be removed by the RFC Editor before publishing this document as an RFC.

Only small editorial changes are made comparing to RFC4653. Technical changes will be made in upcomming versions of this document.