TCP Tuning for HTTP
draft-stenberg-httpbis-tcp-02

Abstract

This document records current best practice for using all versions of HTTP over TCP.

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 February 4, 2017.

Copyright Notice

Copyright (c) 2016 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.

Table of Contents

• 1. Introduction
• 1.1. Notational Conventions
• 2. Socket planning
• 2.1. Number of open files
• 2.2. Number of concurrent network messages
• 2.3. Number of incoming TCP SYNs allowed to backlog
• 2.4. Use the whole port range for local ports
• 2.5. Lower the TCP FIN timeout
• 2.6. Reuse sockets in TIME_WAIT state
• 2.7. TCP socket buffer sizes
• 2.8. TCP Window Scaling
• 2.9. Timers and timeouts
• 3. TCP handshake
• 3.1. TCP Fast Open
• 3.2. Initial Window
• 3.3. TCP SYN flood handling
• 4. TCP transfers
• 4.1. Packet scheduling and flow control
• 4.2. Explicit Congestion Control
• 4.3. Nagle’s Algorithm
• 4.4. Delayed ACKs
• 4.5. Keep-alive
• 5. Re-using connections
• 5.1. Slow Start after Idle
• 5.2. TCP-Bound Authentications
• 6. Closing connections
• 6.1. Half-close
• 6.2. Abort
• 6.3. Close Idle Connections
• 6.4. Tail Loss Probes
• 7. IANA Considerations
• 8. Security Considerations
• 9. References
• 9.1. Normative References
• 9.2. Informative References
• Appendix A. Acknowledgments
• Appendix B. Operating System Settings for Linux
• Author's Address

1. Introduction

HTTP version 1.1 [RFC7230] as well as HTTP version 2 [RFC7540] are defined to use TCP [RFC0793], and their performance can depend greatly upon how TCP is configured. This document records the best current practice for using HTTP over TCP, with a focus on improving end-user perceived performance.

These practices are generally applicable to HTTP/1 as well as HTTP/2, although some may note particular impact or nuance regarding a particular protocol version.

There are countless scenarios, roles and setups where HTTP is being using so there can be no single specific “Right Answer” to most TCP questions. This document intends only to cover the most important areas of concern and suggest possible actions.

1.1. Notational Conventions

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 [RFC2119].

2. Socket planning

Your HTTP server or intermediary may need configuration changes to some system tunables and timeout periods to perform optimally. Actual values will depend on how you are scaling the platform, horizontally or vertically, and other connection semantics. Changing system limits and altering thresholds will change the behavior of your web service and its dependencies. These dependencies are usually common to other services running on the same system, so good planning and testing is advised.

This is a list of values to consider and some general advice on how those values can be modified on Linux systems.

2.1. Number of open files

A modern HTTP server will serve a large number of TCP connections and in most systems each open socket equals an open file. Make sure that limit isn’t a bottle neck.

2.2. Number of concurrent network messages

Raise the number of packets allowed to get queued when a particular interface receives packets faster than the kernel can process them.

2.3. Number of incoming TCP SYNs allowed to backlog

The number of new connection requests that are allowed to queue up in the kernel. These can be connections that are in SYN RECEIVED or ESTABLISHED states. Historically, operating systems used a single backlog queue for both of these states. Newer implemntations use two separate queues: one for connections in SYN RECEIVED and one for those which are ESTABLISHED state (better known as the accept queue).

2.4. Use the whole port range for local ports

To make sure the TCP stack can take full advantage of the entire set of possible sockets, give it a larger range of local port numbers to use.

2.5. Lower the TCP FIN timeout

High connection completion rates will consume ephemeral ports quickly. Lower the time during which connections are in FIN-WAIT-2/TIME_WAIT states so that they can be purged faster and thus maintain a maximal number of available sockets. The primitives for the assignment of these values were described in [RFC0793], however significantly lower values are commonly used.

2.6. Reuse sockets in TIME_WAIT state

When running backend servers on a managed, low latency network you might allow the reuse of sockets in TIME_WAIT state for new connections when a protocol complete termination has occurred. There is no RFC that covers this behaviour.

2.7. TCP socket buffer sizes

Systems meant to handle and serve a huge number of TCP connections at high speeds can require significant amounts of memory for TCP socket buffers to maintain performance. On some systems you can tell the TCP stack what default buffer sizes to use and how much they are allowed to dynamically grow and shrink. Window Scaling is typically linked to socket buffer sizes.

The minimum and default values for socket buffers tend to require less proactive amendment than the maximum value. When deriving maximum values for use, you should consider the BDP (Bandwidth Delay Product) of the target environment and client paths. Consider also that ‘read’ and ‘write’ values do not require to be synchronised, as the BDP for a load balancer or middle-box might be very different when acting as a sender or receiver due to the network charateristics in either context. e.g. A cache with fast, low latency network to an origin serving high latency clients.

Allowing needlessly high values beyond the expected limitations of the platform will not improve performance however can cause buffer induced delays within the path or excessive retransmissions during congestion events. Extensions such as ECN coupled with AQM can help mitigate this undesirable behaviour [RFC7141].

2.8. TCP Window Scaling

Window Scaling is provided as a function of the congestion control algorithm used on a platform. Initial and maximal values can usually be configured.

The window size used at connection startup is a calculated value using the MSS discovered during the 3WHS and the Initial Window (IW) in both send and receive contexts (initcwnd and initrwnd). [RFC7323] covers Window Scaling in greater detail.

You may have to increase the largest allowed window size from the system default to increase the throuput for high latency clients. Window scaling must be accommodated within the maximal values, however it is not uncommon to see the maximum definable higher than the scalable limit; these values can be statically defined within socket parameters (SO_RCVBUF,SO_SNDBUF).

Changes to the size of the window or incomplete window usage are common secondary symptoms of ‘slow transfer rates’ on a loss free path. Locating the root cause of these symptoms, usually on the client or server system, is an important step commonly overlooked in favour of blaming the path.

2.9. Timers and timeouts

On a modern shared platform it can be common to plan for both long and short lived connections on the same implementation. However, the delivery of static assets and a ‘web push’ or ‘long poll’ service provide very different quality of service promises.

Fail ‘fast’: TCP resources can be highly contended. For fault tolerance reasons a server needs to be able to determine within a reasonable time frame whether a connection is still active or required. e.g. If static assets typically return in 100s of milliseconds, and users ‘switch off’ after <10s keeping timeouts of >30s make little sense and defining a ‘quality of service’ appropriate to the target platform is encouraged. On a shared platform with mixed session lifetimes, applications that require longer render times have various options to ensure the underlying service and upstream servers in the path can identify the session as not failed: HTTP continuations, Redirects, 202s or sending data.

Clients and servers typically have many timeout options, a few notable options are: Connect(client), time to request(server), time to first byte(client), between bytes(server/client), total connection time(server/client). Some implementations merge these values into a single ‘timeout’ definition even when statistics are reported individually. All should be considered as the defaults in many implementations are highly underiable, even infinite timeouts have been observed.

3. TCP handshake

3.1. TCP Fast Open

TCP Fast Open (a.k.a. TFO, [RFC7413]) allows data to be sent on the TCP handshake, thereby allowing a request to be sent without any delay if a connection is not open.

TFO requires both client and server support, and additionally requires application knowledge, because the data sent on the SYN needs to be idempotent. Therefore, TFO can only be used on idempotent, safe HTTP methods (e.g., GET and HEAD), or with intervening negotiation (e.g, using TLS). It should be noted that TFO requires a secret to be defined on the server to mitigate security vulnerabilities it introduces. TFO therefore requires more server side deployment planning than other enhancements.

Support for TFO is growing in client platforms, especially mobile, due to the significant performance advantage it gives.

3.2. Initial Window

[RFC6928] proposes a new IW of 10*MSS, and is now fairly widely deployed server-side. Many implementations allow you to tune both initcwnd and initrwnd values. Some implementations allow these values to be applied to specific routes which allows a greater degree of control over known paths.

There has been experimentation with larger initial windows in combination with packet pacing, however IW10 has been reported to perform fairly well even in both general and high volume use cases.

3.3. TCP SYN flood handling

TCP SYN Flood mitigations [RFC4987] are necessary and there will be thresholds to tweak.

4. TCP transfers

4.1. Packet scheduling and flow control

TBD cubic codel pacing

4.2. Explicit Congestion Control

Apple deploying in iOS and OSX.

4.3. Nagle’s Algorithm

Nagle’s Algorithm [RFC0896] is the mechanism that makes the TCP stack hold (small) outgoing packets for a short period of time so that it can potentially merge that packet with the next outgoing one. It is optimized for throughput at the expense of latency.

HTTP/2 in particular requires that the client can send a packet back fast even during transfers that are perceived as single direction transfers. Even small delays in those sends can cause a significant performance loss.

HTTP/1.1 is also affected, especially when sending off a full request in a single write() system call.

In POSIX systems you switch it off like this:

int one = 1;
setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &one, sizeof(one));

4.4. Delayed ACKs

Delayed ACK [RFC1122] is a mechanism enabled in most TCP stacks that causes the stack to delay sending acknowledgement packets in response to data. The ACK is delayed up until a certain threshold, or until the peer has some data to send, in which case the ACK will be sent along with that data. Depending on the traffic flow and TCP stack this delay can be as long as 500ms.

This interacts poorly with peers that have Nagle’s Algorithm enabled. Because Nagle’s Algorithm delays sending until either one MSS of data is provided or until an ACK is received for all sent data, delaying ACKs can force Nagle’s Algorithm to buffer packets when it doesn’t need to (that is, when the other peer has already processed the outstanding data).

Delayed ACKs can be useful in situations where it is reasonable to assume that a data packet will almost immediately (within 500ms) cause data to be sent in the other direction. In general in both HTTP/1.1 and HTTP/2 this is unlikely: therefore, disabling Delayed ACKs can provide an improvement in latency.

However, the TLS handshake is a clear exception to this case. For the duration of the TLS handshake it is likely to be useful to keep Delayed ACKs enabled.

Additionally, for low-latency servers that can guarantee responses to requests within 500ms, on long-running connections (such as HTTP/2), and when requests are small enough to fit within a small packet, leaving delayed ACKs turned on may provide minor performance benefits.

Effective use of switching off delayed ACKs requires extensive profiling.

4.5. Keep-alive

TCP keep-alive is likely disabled - at least on mobile clients for energy saving purposes. App-level keep-alive is then required for long-lived requests to detect failed peers or connections reset by stateful firewalls etc.

5. Re-using connections

5.1. Slow Start after Idle

Slow-start is one of the algorithms that TCP uses to control congestion inside the network. It is also known as the exponential growth phase. Each TCP connection will start off in slow-start but will also go back to slow-start after a certain amount of idle time.

5.2. TCP-Bound Authentications

There are several HTTP authentication mechanisms in use today that are used or can be used to authenticate a connection rather than a single HTTP request. Two popular ones are NTLM and Negotiate.

If such an authentication has been negotiated on a TCP connection, that connection can remain authenticated throughout the rest of its lifetime. This discrepancy with how other HTTP authentications work makes it important to handle these connections with care.

6. Closing connections

6.1. Half-close

The client or server is free to half-close after a request or response has been completed; or when there is no pending stream in HTTP/2.

Half-closing is sometimes the only way for a server to make sure it closes down connections cleanly so that it doesn’t accept more requests while still allowing clients to receive the ongoing responses.

6.2. Abort

No client abort for HTTP/1.1 after the request body has been sent. Delayed full close is expected following an error response to avoid RST on the client.

6.3. Close Idle Connections

Keeping open connections around for subsequent connection reuse is key for many HTTP clients’ performance. The value of an existing connection quickly degrades and after only a few minutes the chance that a connection will successfully get reused by a web browser is slim.

6.4. Tail Loss Probes

draft

7. IANA Considerations

This document does not require action from IANA.

8. Security Considerations

TBD

9. References

9.1. Normative References

9.2. Informative References

Appendix A. Acknowledgments

This specification builds upon previous work and help from Mark Nottingham, Craig Taylor

Appendix B. Operating System Settings for Linux

Here are some sample operating system settings for the Linux operating system, along with the section it refers to.

Section 2.1

fs.file-max = <number of files>

Section 2.2

net.core.netdev_max_backlog = <number of packets>

Section 2.3

net.core.somaxconn = <number>

Section 2.4

net.ipv4.ip_local_port_range = 1024 65535

Section 2.5

net.ipv4.tcp_fin_timeout = <number of seconds>

Section 2.6

net.ipv4.tcp_tw_reuse = 1

Section 2.7

net.ipv4.tcp_wmem = <minimum size> <default size> <max size in bytes>

Section 2.7

net.ipv4.tcp_rmem = <minimum size> <default size> <max size in bytes>

Section 2.8

net.core.rmem_max = <number of bytes>

Section 2.8

net.core.wmem_max = <number of bytes>

Section 5.1

net.ipv4.tcp_slow_start_after_idle = 0

Section 4.3 Turning off Nagle’s Algorithm:

int one = 1;
setsockopt(fd, IPPROTO_TCP, TCP_NODELAY, &one, sizeof(one));

Section 4.4

On recent Linux kernels (since Linux 2.4.4), Delayed ACKs can be disabled like this:

int one = 1;
setsockopt(fd, IPPROTO_TCP, TCP_QUICKACK, &one, sizeof(one));

Unlike disabling Nagle’s Algorithm, disabling Delayed ACKs on Linux is not a one-time operation: processing within the TCP stack can cause Delayed ACKs to be re-enabled. As a result, to use TCP_QUICKACK effectively requires setting and unsetting the socket option during the life of the connection.

Author's Address