TCP and SCTP RTO Restart
draft-hurtig-tcpm-rtorestart-03

Abstract

This document describes a modified algorithm for managing the TCP and SCTP retransmission timers that provides faster loss recovery when a connection's amount of outstanding data is small. The modification allows the transport to restart its retransmission timer more aggressively in situations where fast retransmit cannot be used. This enables faster loss detection and recovery for connections that are short-lived or application-limited.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at http:/⁠/⁠datatracker.ietf.org/⁠drafts/⁠current/⁠.

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."

This Internet-Draft will expire on April 02, 2013.

Copyright Notice

Copyright (c) 2012 IETF Trust and the persons identified as the document authors. All rights reserved.

This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http:/⁠/⁠trustee.ietf.org/⁠license-⁠info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.

1. Introduction

TCP uses two mechanisms to detect segment loss. First, if a segment is not acknowledged within a certain amount of time, a retransmission timeout (RTO) occurs, and the segment is retransmitted [RFC6298]. While the RTO is based on measured round-trip times (RTTs) between the sender and receiver, it also has a conservative lower bound of 1 second to ensure that delayed segments are not mistaken as lost. Second, when a sender receives duplicate acknowledgments, the fast retransmit algorithm infers segment loss and triggers a retransmission. Duplicate acknowledgments are generated by a receiver when out-of-order segments arrive. As both segment loss and segment reordering cause out-of-order arrival, fast retransmit waits for three duplicate acknowledgments before considering the segment as lost. In some situations, however, the number of outstanding segments is not enough to trigger three duplicate acknowledgments, and the sender must rely on lengthy RTOs for loss recovery.

The amount of outstanding segments can be small for several reasons:

(1)

The connection is limited by the congestion control when the path has a low total capacity (bandwidth-delay product) or the connection's share of the capacity is small. It is also limited by the congestion control in the first RTTs of a connection or after an RTO when the available capacity is probed using slow-start.

(2)

The connection is limited by the receiver's available buffer space.

(3)

The connection is limited by the application if the available capacity of the path is not fully utilized (e.g. interactive applications), or at the end of a transfer, which is frequent if the total amount of data is small (e.g. web traffic).

The first two situations can occur for any flow, as external factors at the network and/or host level cause them. The third situation primarily affects flows that are short or have a low transmission rate. Typical examples of applications that produce short flows are web servers. [RJ10] shows that 70% of all web objects, found at the top 500 sites, are too small for fast retransmit to work. [BPS98] shows that about 56% of all retransmissions sent by a busy web server are sent after RTO expiry. While the experiments were not conducted using SACK [RFC2018], only 4% of the RTO-based retransmissions could have been avoided. Applications have a low transmission rate when data is sent in response to actions, or as a reaction to real life events. Typical examples of such applications are stock trading systems, remote computer operations and online games. What is special about this class of applications is that they are time-dependant, and extra latency can reduce the application service level [P09]. Although such applications may represent a small amount of data sent on the network, a considerable number of flows have such properties and the importance of low latency is high.

The RTO restart approach outlined in this document makes the RTO slightly more aggressive when the number of outstanding segments is small, in an attempt to enable faster loss recovery for all segments while being robust to reordering. While it still conforms to the requirement in [RFC6298] that segments must not be retransmitted earlier than RTO seconds after their original transmission, it could increase the chance for a spurious timeout, which could degrade performance when the congestion window (cwnd) is large -- for example, when an application sends enough data to reach a cwnd covering 100 segments and then stops. The likelihood and potential impact of this problem as well as possible mitigation strategies are currently under investigation.

While this document focuses on TCP, the described changes are also valid for the Stream Control Transmission Protocol (SCTP) [RFC4960] which has similar loss recovery and congestion control algorithms.

1.1. Requirements Language

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].

2. RTO Restart Overview

The RTO management algorithm described in [RFC6298] recommends that the retransmission timer is restarted when an acknowledgment (ACK) that acknowledges new data is received and there is still outstanding data. The restart is conducted to guarantee that unacknowledged segments will be retransmitted after approximately RTO seconds. However, by restarting the timer on each incoming acknowledgment, retransmissions are not typically triggered RTO seconds after their previous transmission but rather RTO seconds after the last ACK arrived. The duration of this extra delay depends on several factors but is in most cases approximately one RTT. Hence, in most situations the time before a retransmission is triggered is equal to "RTO + RTT".

The extra delay can be significant, especially for applications that use a lower RTOmin than the standard of 1 second and/or in environments with high RTTs, e.g. mobile networks. The restart approach is illustrated in Figure 1 where a TCP sender transmits three segments to a receiver. The arrival of the first and second segment triggers a delayed ACK [RFC1122], which restarts the RTO timer at the sender. The RTO restart is performed approximately one RTT after the transmission of the third segment. Thus, if the third segment is lost, as indicated in Figure 1, the effective loss detection time is "RTO + RTT" seconds. In some situations, the effective loss detection time becomes even longer. Consider a scenario where only two segments are outstanding. If the second segment is lost, the time to expire the delayed ACK timer will also be included in the effective loss detection time.

		  
Sender                               Receiver
              ...
DATA [SEG 1] ----------------------> (ack delayed)
DATA [SEG 2] ----------------------> (send ack)
DATA [SEG 3] ----X         /-------- ACK
(restart RTO)  <----------/
              ...
(RTO expiry)
DATA [SEG 3] ---------------------->

		

Figure 1: RTO restart example

During normal TCP bulk transfer the current RTO restart approach is not a problem. Actually, as long as enough segments arrive at a receiver to enable fast retransmit, RTO-based loss recovery should be avoided. RTOs should only be used as a last resort, as they drastically lower the congestion window compared to fast retransmit, and the current approach can therefore be beneficial -- it is described in [EL04] to act as a "safety margin" that compensates for some of the problems that the authors have identified with the standard RTO calculation. Notably, the authors of [EL04] also state that "this safety margin does not exist for highly interactive applications where often only a single packet is in flight."

There are only a few situations where timeouts are appropriate, or the only choice. For example, if the network is severely congested and no segments arrive, RTO-based recovery should be used. In this situation, the time to recover from the loss(es) will not be the performance bottleneck. Furthermore, for connections that do not utilize enough capacity to enable fast retransmit, RTO is the only choice. The time needed for loss detection in such scenarios can become a serious performance bottleneck.

3. RTO Restart Algorithm

To enable faster loss recovery for connections that are unable to use fast retransmit, an alternative RTO restart can be used. By resetting the timer to "RTO - T_earliest", where T_earliest is the time elapsed since the earliest outstanding segment was transmitted, retransmissions will always occur after exactly RTO seconds. This approach makes the RTO more aggressive than the standardized approach in [RFC6298] but still conforms to the requirement in [RFC6298] that segments must not be retransmitted earlier than RTO seconds after their original transmission.

This document specifies the following update of step 5.3 in Section 5 of [RFC6298] (and a similar update in Section 6.3.2 of [RFC4960] for SCTP):

When an ACK is received that acknowledges new data:

(1)

Set T_earliest = 0.

(2)

If the following two conditions hold:

(a)

The number of outstanding segments is less than four.

(b)

There is no unsent data ready for transmission or the receiver's advertised window does not permit transmission.

set T_earliest to the time elapsed since the earliest outstanding segment was sent.

(3)

Restart the retransmission timer so that it will expire after "RTO - T_earliest" seconds (for the current value of RTO).

The update requires TCP implementations to track the time elapsed since the transmission of the earliest outstanding segment (T_earliest). As the alternative restart is used only when the number of outstanding segments is less than four only four segments need to be tracked. Furthermore, some implementations of TCP (e.g. Linux TCP) already track the transmission times of all segments.

4. Discussion

The currently standardized algorithm has been shown to add at least one RTT to the loss recovery process in TCP [LS00] and SCTP [HB08][PBP09]. Applications that have strict timing requirements (e.g. telephony signaling and gaming) rather than throughput requirements may want to use a lower RTOmin than the standard of 1 second [RFC4166]. For such applications the modified restart approach could be important as the RTT and also the delayed ACK timer of receivers will be large components of the effective loss recovery time. Measurements in [HB08] have shown that the total transfer time of a lost segment (including the original transmission time and the loss recovery time) can be reduced with up to 35% using the suggested approach. These results match those presented in [PGH06][PBP09], where the modified restart approach is shown to significantly reduce retransmission latency.

There are several proposals that address the problem of not having enough ACKs for loss recovery. In what follows, we explain why the mechanism described here is complementary to these approaches:

The limited transmit mechanism [RFC3042] allows a TCP sender to transmit a previously unsent segment for each of the first two duplicate acknowledgments. By transmitting new segments, the sender attempts to generate additional duplicate acknowledgments to enable fast retransmit. However, limited transmit does not help if no previously unsent data is ready for transmission or if the receiver is out of buffer space. [RFC5827] specifies an early retransmit algorithm to enable fast loss recovery in such situations. By dynamically lowering the amount of duplicate acknowledgments needed for fast retransmit (dupthresh), based on the number of outstanding segments, a smaller number of duplicate acknowledgments are needed to trigger a retransmission. In some situations, however, the algorithm is of no use or might not work properly. First, if a single segment is outstanding, and lost, it is impossible to use early retransmit. Second, if ACKs are lost, the early retransmit cannot help. Third, if the network path reorders segments, the algorithm might cause more unnecessary retransmissions than fast retransmit.

TCP-NCR [RFC4653] sets the dupthresh to three or more, to better disambiguate reordered and lost segments. In addition, early retransmit lowers the dupthresh when the amount of outstanding data is small, to enable faster loss recovery. The reasons why the RTO restart procedure described in this document does not take dynamic dupthresh considerations into account are twofold. First, if a larger dupthresh is used, the RTO restart approach could be used when the congestion window, and the amount of outstanding data, is larger. However, in such situations the actual amount of outstanding data can significantly impact the RTT of the connection, making it potentially dangerous to be more aggressive. Second, if a smaller dupthresh is used, the amount of outstanding data needed for a restart is smaller. However, as the congestion window is already small, it does not matter if a retransmission is due to a fast retransmit or an RTO. The resulting congestion window will still be very small, and the only difference is how quickly TCP infers segment loss.

Tail Loss Probe [TLP] is a proposal to send up to two "probe segments" when a timer fires which is set to a value smaller than the RTO. A "probe segment" is a new segment if new data is available, else a retransmission. The intention is to compensate for sluggish RTO behavior in situations where the RTO greatly exceeds the RTT, which, according to measurements reported in [TLP], is not uncommon. The Probe timeout (PTO) is at least 2 RTTs, and only scheduled in case the RTO is farther than the PTO. A spurious PTO is less risky than a spurious RTO, as it would not have the same negative effects (clearing the scoreboard and restarting with slow-start). In contrast, RTO restart is trying to make the RTO more appropriate in cases where there is no need to be overly cautious.

TLP could kick in in situations where RTO restart does not apply, and it could overrule (yielding a similar general behavior, but with a lower timeout) RTO restart in cases where the number of outstanding segments is smaller than 4 and no new segments are available for transmission. The shorter RTO from RTO restart also reduces the probability that TLP is activated because PTO might be farther than RTO. This could make RTO restart more aggressive than the algorithm in [TLP] when: [RFC5681], 3 packets can immediately be retransmitted, whereas TLP only allows up to two consecutive PTOs.

(1)

no data has been sent in an interval exceeding the RTO

(2)

the number of outstanding segments is 3

(3)

(defined in [RFC5681]) is at least 3

because, under these conditions, in accordance with

5. IANA Considerations

This memo includes no request to IANA.

6. Security Considerations

This document discusses a change in how to set the retransmission timer's value when restarted. This change does not raise any new security issues with TCP or SCTP.

7. References

7.1. Normative References

7.2. Informative References

Authors' Addresses

Table of Contents

• 1. Introduction
• 1.1. Requirements Language
• 2. RTO Restart Overview
• 3. RTO Restart Algorithm
• 4. Discussion
• 5. IANA Considerations
• 6. Security Considerations
• 7. References
• 7.1. Normative References
• 7.2. Informative References
• Authors' Addresses