Current Practices for Multiple Interface Hosts
draft-ietf-mif-current-practices-00

Status of this Memo

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

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.

This Internet-Draft will expire on April 22, 2010.

Copyright Notice

Copyright (c) 2009 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 in effect on the date of publication of this document (http://trustee.ietf.org/license-info). Please review these documents carefully, as they describe your rights and restrictions with respect to this document.

Abstract

An increasing number of hosts are operating in multiple-interface environments, where different network interfaces are providing unequal levels of service or connectivity. This document summarizes current practices in this area, and describes in detail how some common operating systems cope with these challenges.

Table of Contents

1.  Introduction
2.  Summary of Current Approaches
    2.1.  Centralized Connection Management
    2.2.  Per Application Connection Settings
    2.3.  Stack-Level Solutions to Specific Problems
        2.3.1.  DNS Resolution Issues
        2.3.2.  Routing
        2.3.3.  Address Selection Policy
3.  Current Practices in Some Operating Systems
    3.1.  Mobile Handset Operating Systems
        3.1.1.  Nokia S60 3rd Edition, Feature Pack 2
        3.1.2.  Microsoft Windows Mobile 2003 Second Edition
        3.1.3.  BlackBerry
        3.1.4.  Google Android
        3.1.5.  Arena Connection Manager
    3.2.  Desktop Operating Systems
        3.2.1.  Microsoft Windows
            3.2.1.1.  Routing
            3.2.1.2.  Outbound and Inbound Addresses
            3.2.1.3.  DNS Configuration
        3.2.2.  Linux and BSD-based Operating Systems
        3.2.3.  Apple Mac OS X
4.  Acknowledgements
5.  IANA Considerations
6.  Security Considerations
7.  Change Log
8.  Contributors
9.  References
    9.1.  Normative References
    9.2.  Informative References
§  Author's Address

1.  Introduction

Multiple-interface hosts face several challenges not faced by single-interface hosts, some of which are described in the MIF problem statement, [I‑D.ietf‑mif‑problem‑statement] (Blanchet, M. and P. Seite, “Multiple Interfaces Problem Statement,” March 2010.). This document summarizes how current implementations deal with the problems identified in the MIF problem statement.

Publicly-available information about the multiple-interface solutions implemented in some widely used operating systems, including both mobile handset and desktop operating systems, is collected in this document, including: Nokia S60 [S60] (Nokia Corporation, “S60 Platform: IP Bearer Management,” 2007.), Microsoft Windows Mobile [WINDOWSMOBILE] (Microsoft Corporation, “SDK Documentation for Windows Mobile-Based Smartphones: Connection Manager,” 2005.), Blackberry [BLACKBERRY] (Research In Motion Limited, “BlackBerry Java Development Environment - Fundamentals Guide: Wireless gateways,” 2009.), Google Android [ANDROID] (Google Inc., “Android developers: package android.net,” 2009.), Microsoft Windows, Apple Mac OS X, Linux and BSD-based operating systems.

2.  Summary of Current Approaches

This section summarizes current approaches that are used to resolve the multi-interface issues described in the Multiple Interface Problem Statement [I‑D.ietf‑mif‑problem‑statement] (Blanchet, M. and P. Seite, “Multiple Interfaces Problem Statement,” March 2010.). These approaches can be broken down into three major categories:

• Centralized connection management
• Per-application connection settings
• Stack-level solutions to specific problems

2.1.  Centralized Connection Management

It is a common practice for mobile handset operating systems to use a centralized connection manager that performs network interface selection based on application or user input. The information used by the connection manager may be programmed into an application or provisioned on a handset-wide basis. When information is not available to make an interface selection, the connection manager will query the user to choose between available choices.

Routing tables are not typically used for network interface selection when a connection manager is in use, as the criteria for network selection is not strictly IP-based but is also dependent on other properties of the interface (cost, type, etc.). Furthermore, multiple overlapping private IPv4 address spaces are often exposed to a multiple-interface host, making it difficult to make interface selection decisions based on prefix matching.

2.2.  Per Application Connection Settings

In mobile handsets, applications are often involved in choosing what interface and related configuration information should be used. In some cases, the application selects the interface directly, and in other cases the application provides more abstract information to a connection manager that makes the final interface choice.

2.3.  Stack-Level Solutions to Specific Problems

In most desktop operating systems, multiple interface problems are dealt with in the stack and related components, based on system- level configuration information, without the benefit of input from applications or users. These solutions tend to map well to the problems listed in the problem statement:

• DNS resolution issues
• Routing
• Address selection policy

The configuration information for desktop systems comes from one of three sources: DHCP, proprietary configuration systems or manual configuration. While these systems universally accept IP address assignment on a per-interface basis, they differ in what set of information can be assigned on a per-interface basis and what can be configured only on a per-system basis.

When choosing between multiple sets of information provided, these systems will typically give preference to information received on the "primary" interface. The mechanism for designating the "primary" interface differs by system.

There is very little commonality in how desktop operating systems handle multiple sets of configuration information, with notable variations between different versions of the same operating system and/or within different software packages built for the same operating system. Although these systems differ widely, it is not clear that any of them provide a completely satisfactory user experience in multiple-interface environments.

The following sections discuss some of the solutions used in each of the areas raised in the MIF problem statement.

2.3.1.  DNS Resolution Issues

There is very little commonality in how desktop operating systems handle the DNS server list. Some systems support per-interface DNS server lists, while others only support a single system-wide list.

On hosts with per-interface DNS server lists, different mechanisms are used to determine which DNS server is contacted for a given query. In most cases, the first DNS server listed on the "primary" interface is queried first, with back off to other servers if an answer is not received.

Systems that support a single system-wide list differ in how they select which DNS server to use in cases where they receive more than one DNS server list to configure (e.g. from DHCP on multiple interfaces). Some accept the information received on the "primary" interface, while others use either the first or last set DNS server list configured.

2.3.2.  Routing

Routing information is also handled differently on different desktop operating systems. While all systems maintain some sort of routing cache, to handle redirects and/or statically configured routes, most packets are routed based on configured default gateway information.

Some systems do allow the configuration of different default router lists for different interfaces. These systems will always choose the default gateway on the interface with the lowest routing metric, with different behavior when two or more interfaces have the same routing metric.

Most systems do not allow the configuration of more than one default router list, choosing instead to use the first or last default router list configured and/or the router list configured on the "primary" interface.

2.3.3.  Address Selection Policy

There is somewhat more commonality in how desktop hosts handle address selection. Applications typically provide the destination address for an outgoing packet, and the IP stack is responsible for picking the source address.

IPv6 specifies a specific source address selection mechanism in [RFC3484] (Draves, R., “Default Address Selection for Internet Protocol version 6 (IPv6),” February 2003.), and several systems implement this mechanism with similar support for IPv4. However, many systems do not provide any mechanism to update this default policy, and there is no standard way to do so.

In some cases, the routing decision (including which interface to use) is made before source address selection is performed, and a source address is chosen from the outbound interface. In other cases, source address selection is performed before, or independently from outbound interface selection.

3.  Current Practices in Some Operating Systems

The following sections briefly describe the current multiple-interface host implementations on some widely-used operating systems. Please refer to the References section for pointers to original documentation on most of these systems, including further details.

3.1.  Mobile Handset Operating Systems

Cellular devices typically run a variety of applications in parallel, each with different requirements for IP connectivity. A typical scenario is shown in figure 1, where a cellular device is utilizing WLAN access for web browsing and GPRS access for transferring multimedia messages (MMS). Another typical scenario would be a real-time VoIP session over one network interface in parallel with best effort web browsing on another network interface. Yet another typical scenario would be global Internet access through one network interface and local (e.g. corporate VPN) network access through another.

     Web server                                       MMS Gateway
          |                                                |
         -+--Internet----            ----Operator network--+-
                 |                          |
             +-------+                  +-------+
             |WLAN AP|                  | GGSN  |
             +-------+                  +-------+
                 |        +--------+        |
                 +--------|Cellular|--------+
                          |device  |
                          +--------+

A cellular device with two network interfaces

Different network access technologies require different settings. For example, WLAN requires Service Set Identifier (SSID) and the GPRS network requires the Access Point Name (APN) of the Gateway GPRS Support Node (GGSN), among other parameters. It is common that different accesses lead to different destination networks (e.g. to "Internet", "Intranet", cellular network services, etc.).

3.1.1.  Nokia S60 3rd Edition, Feature Pack 2

S60 uses the concept of an Internet Access Point (IAP) [S60] (Nokia Corporation, “S60 Platform: IP Bearer Management,” 2007.) that contains all information required for opening a network connection using a specific access technology. A device may have several IAPs configured for different network technologies and settings (multiple WLAN SSIDs, GPRS APNs, dial-up numbers, and so forth). There may also be 'virtual' IAPs that define parameters needed for tunnel establishment (e.g. for VPN).

For each application, a correct IAP needs to be selected at the point when the application requires network connectivity. This is essential, as the wrong IAP may not be able to support the application or reach the desired destination. For example, MMS application must use the correct IAP in order to reach the MMS Gateway, which typically is not accessible from the public Internet. As another example, an application might need to use the IAP associated with its corporate VPN in order to reach internal corporate servers. Binding applications to IAPs avoids several problems, such as choosing the correct DNS server in the presence of split DNS (as an application will use the DNS server list from its bound IAP), and overlapping private IPv4 address spaces used for different interfaces (as each application will use the default routes from its bound IAP).

If multiple applications utilize the same IAP, the underlying network connection can typically be shared. This is often the case when multiple Internet-using applications are running in parallel.

The IAP for an application can be selected in multiple ways:

• Statically: e.g. from a configuration interface, via client provisioning/device management system, or at build-time.
• Manually by the user: e.g. each time an application starts the user may be asked to select the IAP to use. This may be needed, for example, if a user sometimes wishes to access his corporate intranet and other times would prefer to access the Internet directly.
• Automatically by the system: after the destination network has been selected statically or dynamically.

The static approach is fine for certain applications, like MMS, for which configuration can be provisioned by the network operator and does not change often. Manual selection works, but may be seen as troublesome by the user. An automatic selection mechanism needs to have some way of knowing which destination network the user, or an application, is trying access.

S60 3rd Edition, Feature Pack 2, introduces a concept of Service Network Access Points (SNAPs) that group together IAPs that lead to the same destination. This enables static or manual selection of the destination network for an application and leaves the problem of selecting the best of the available IAPs within a SNAP to the operating system.

When SNAPs are used, it is possibly for the operating system to notify applications when a preferred IAP, leading to the same destination, becomes available (for example, when a user comes within range of his home WLAN access point), or when the currently used IAP is no longer available and applications have to reconnect via another IAP (for example, when a user goes out of range of his home WLAN and must move to the cellular network).

Please see the source documentation for more details and screenshots: [S60] (Nokia Corporation, “S60 Platform: IP Bearer Management,” 2007.).

3.1.2.  Microsoft Windows Mobile 2003 Second Edition

A Connection Manager architecture is described in [WINDOWSMOBILE] (Microsoft Corporation, “SDK Documentation for Windows Mobile-Based Smartphones: Connection Manager,” 2005.). This architecture centralizes and automates network connection establishment and management, and makes it possible to automatically select a connection, to dial-in automatically or by user initiation, and to optimize connection and shared resource usage. Connection Manager periodically re-evaluates the validity of the connection selection. The Connection Manager uses various attributes such as cost, security, bandwidth, error rate, and latency in its decision making.

The Connection Manager selects the best possible connection for the application based on the destination network the application wishes to reach. The selection is made between available physical and virtual connections (e.g. VPN, GPRS, WLAN, and wired Ethernet) that are known to provide connectivity to the destination network, and the selection is based on the costs associated with each connection. Different applications are bundled to use the same network connection when possible, but in conflict situations when a connection cannot be shared, higher priority applications take precedence, and the lower priority applications lose connectivity until the conflict situation clears.

During operation, Connection Manager opens new connections as needed, and also disconnects unused or idle connections.

To optimize resource use, such as battery power and bandwidth, Connection Manager enables applications to synchronize network connection usage by allowing applications to register their requirements for periodic connectivity. An application is notified when a suitable connection becomes available for its use.

3.1.3.  BlackBerry

In BlackBerry devices [BLACKBERRY] (Research In Motion Limited, “BlackBerry Java Development Environment - Fundamentals Guide: Wireless gateways,” 2009.) Java applications can use one of two wireless gateways to proxy the connection to the Internet or to a corporate network. The application can be designed to always use the default Internet gateway, or to use a more preferred enterprise gateway when available. The intent is to hide connectivity issues from users.

DISCUSS: How does the Blackberry decides when a WLAN interface, a cellular interface or some other physical interface is used?

3.1.4.  Google Android

The Android reference documentation describes the android.net package [ANDROID] (Google Inc., “Android developers: package android.net,” 2009.) and the ConnectivityManager class that applications can use to request a route to a specified destination address via a specified network interface (Mobile or Wifi). Applications also ask Connection Manager for permission to start using a network feature. The Connectivity Manager monitors changes in network connectivity and attempts to failover to another network if connectivity to an active network is lost. When there are changes in network connectivity, applications are notified. Applications are also able to ask for information about all network interfaces, including their availability, type and other information.

DISCUSS: Are applications bound to use one network type at a time, or can one application use multiple network features in parallel?

3.1.5.  Arena Connection Manager

The Arena Connection Manager is described in [I‑D.zhang‑mif‑connection‑manager‑arena] (Zhang, Y., Sun, T., and H. Chen, “Multi-interface Network Connection Manager in Arena Platform,” February 2009.) and [I‑D.yang‑mif‑connection‑manager‑impl‑req] (Yang, J., Sun, T., and S. Fan, “Multi-interface Connection Manager Implementation and Requirements,” March 2009.). The arena connection manager provides a means for applications to register their connectivity requirement with the Connection Manager. The Connection Manager can then choose an interface that matches the application's needs while considering other factors such as availability, cost and stability. Also, the Connection Manager can handle multi-interface issues such as connection sharing.

3.2.  Desktop Operating Systems

Multi-interface issues also occur in desktop environments in those cases where a desktop host has multiple (logical or physical) interfaces connected to networks with different reachability properties, such as one interface connected to the global Internet, while another interface is connected to a corporate VPN.

3.2.1.  Microsoft Windows

The multi-interface functionality currently implemented in Microsoft Windows operation systems is described in more detail in [I‑D.montenegro‑mif‑multihoming] (Montenegro, G., Thaler, D., and S. Seshadri, “Multiple Interfaces on Windows,” March 2009.).

3.2.1.1.  Routing

It is possible, although not often desirable, to configure default routers on more than one Windows interface. In this configuration, Windows will use the default route on the interface with the lowest routing metric (i.e. the fastest interface). If multiple interfaces share the same metric, the behavior will differ based on the version of Windows in use. Prior to Windows Vista, the packet would be routed out of the first interface that was bound to the TCP/IP stack, the preferred interface. In Windows vista, host-to-router load sharing [RFC4311] (Hinden, R. and D. Thaler, “IPv6 Host-to-Router Load Sharing,” November 2005.) is used for both IPv4 and IPv6.

3.2.1.2.  Outbound and Inbound Addresses

If the source address of the outgoing packet has not been determined by the application, Windows will choose from the addresses assigned to its interfaces. Windows implements [RFC3484] (Draves, R., “Default Address Selection for Internet Protocol version 6 (IPv6),” February 2003.) for source address selection in IPv6 and, in Windows Vista, for IPv4. Prior to Windows Vista, IPv4 simply chose the first address on the outgoing interface.

For incoming packets, Windows will check if the destination address matches one of the addresses assigned to its interfaces. Windows has implemented the weak host model [RFC1122] (Braden, R., “Requirements for Internet Hosts - Communication Layers,” October 1989.) on IPv4 in Windows 2000, Windows XP and Windows Server 2003. The strong host model became the default for IPv4 in Windows Vista and Windows server 2008, however the weak host model is available via per-interface configuration. IPv6 has always implemented the strong host model.

3.2.1.3.  DNS Configuration

Host-wide DNS configuration is input via static configuration or, in sites that use Active Directory, Microsoft's Group Policy. The host-wide configuration consists of a DNS suffix to be used for the local host, as well as a list of domain names that can be appended to names being queried. Before Windows Vista and Windows Server 2008, there was also a host-wide DNS server list that took precedent over per-interface DNS configuration.

Interface specific DNS configuration can be input via static configuration or via DHCP. It includes:

• An interface-specific suffix list.
• A list of DNS server IP addresses.

In the list of DNS server addresses, the first server is considered the "primary" server, with all other servers being secondary.

When a DNS query is performed in Windows, the query is first sent to the primary DNS server on the preferred interface. If no response is received in one second, the query is sent to the primary DNS servers on all interfaces under consideration. If no response is received for 2 more seconds, the DNS server sends the query to all of the DNS servers on the DNS server lists for all interfaces under consideration. If the host still doesn't receive a response after 4 seconds, it will send to all of the servers again and wait 8 seconds for a response.

3.2.2.  Linux and BSD-based Operating Systems

Most BSD and Linux distributions rely on their DHCP client to handle the configuration of interface-specific information (such as an IP address and netmask), and a set of system-wide configuration information, (such a DNS server list, an NTP server list and default routes). Users of these operating systems have the choice of using any DHCP client available for their platform, with an operating system default. This section discusses the behavior of several DHCP clients that may be used with Linux and BSD distributions.

The Internet Systems Consortium (ISC) DHCP Client [ISCDHCP] (Internet Software Consortium, “ISC DHCP,” 2009.) and its derivative for OpenBSD [OPENBSDDHCLIENT] (OpenBSD, “OpenBSD dhclient,” 2009.) can be configured with specific instructions for each interface. However, each time new configuration data is received by the host from a DHCP server, regardless of which interface it is received on, the DHCP client rewrites the global configuration data, such as the default routes and the DNS server list (in /etc/resolv.conf) with the most recent information received. Therefore, the last configured interface always become the primary one. The ISC DHCPv6 client behaves similarly.

The Phystech dhcpcd client [PHYSTECHDHCPC] (Phystech, “dhcpcd,” 2009.) behaves similarly to the ISC client. It replaces the DNS server list in /etc/resolv.conf and the default routes each time new DHCP information is received on any interface. However, the -R flag can be used to instruct the client to not replace the DNS servers in /etc/resolv.conf. However, this flag is a global flag for the DHCP server, and is therefore applicable to all interfaces. When dhcpd is called with the -R flag, the DNS servers are never replaced.

The pump client [PUMP] (RedHat, “PUMP,” 2009.) also behaves similarly to the ISC client. It replaces the DNS servers in /etc/resolv.conf and the default routes each time new DHCP information is received on any interface. However, the nodns and nogateway options can be specified on a per interface basis, enabling the user to define which interface should be used to obtain the global configuration information.

The udhcp client [UDHCP] (Busybox, “uDHCP,” 2009.) is often used in embedded platforms based on busybox. The udhcp client behaves similarly to the ISC client. It rewrites default routes and the DNS server list each time new DHCP information is received.

Redhat-based distributions, such as Redhat, Centos and Fedora have a per-interface configuration option (PEERDNS) that indicates that the DNS server list should not be updated based on configuration received on that interface.

The most configurable DHCP clients can be set to define a primary interface to use only that interface for the global configuration data. However, this is limited, since a mobile host might not always have the same set of interfaces available. Connection managers may help in this situation.

Some distributions also have a connection manager. However, most connection managers serve as a GUI to the DHCP client, therefore not changing the functionality described above . TODO: Verify all connection managers.

TODO: DHCPv6 clients

3.2.3.  Apple Mac OS X

This section is based on testing Mac OS X (version 10.5.6).

When using multiple interfaces on Mac OS X, global configuration data such as default routes and the DNS server list are taken from the DHCP data received on the primary interface. Therefore, the order in which the interfaces receive their configuration data is not relevant. For example, if the primary interface receives its configuration data first, then the second interface receives its configuration data, the interface-specific information for the second interface will be configured, but the global configuration information such as the DNS server list and default routes is not overwritten.

4.  Acknowledgements

Authors of the document would like to thank following people for their input and feedback: Hui Deng, Jari Arkko.

This document was prepared using xml2rfc template and related web-tool.

5.  IANA Considerations

This memo includes no request to IANA.

6.  Security Considerations

This document describes current operating system implementations and how they handle the issues raised in the MIF problem statement. While it is possible that the currently implemented mechanisms described in this document may affect the security of the systems described, this document merely reports on current practice. It does not attempt to analyze the security properties (or any other architectural properties) of the currently implemented mechanisms.

7.  Change Log

The following changes were made between draft-mrw-mif-current-practices and draft-ietf-mif-current-practices-00.txt.

• Updated operating system info based on new information supplied.
• Renamed "Current Solutions" to "Summary of Current Solutions" and moved above OS specifics. Modified this section substantially to match problems in problem statement.
• Removed "Current Problems" section, as it duplicated information in the problem statement.
• Broke OS specifics into separate sections for mobile handsets and desktop operating systems.

The following changes were made between versions -00 and -01:

• Number of authors passed five, so moved to Contributors section.
• Added information on Microsoft Windows.
• Added information on Arena Connection Manager.

8.  Contributors

The following people contributed most of the per-Operating System information found in this document:

• Marc Blanchet, Viagenie
• Hua Chen, Leadcoretech, Ltd.
• Shunan Fan, Huawei Technology
• Gabriel Montenegro, Microsoft Corporation
• Teemu Savolainen, Nokia
• Shyam Seshadri, Microsoft Corporation
• Tao Sun, China Mobile
• Dave Thaler, Microsoft Corporation
• Jian Yang, Huawei Technology
• Yan Zhang, Leadcoretech Ltd.

9.  References

9.1. Normative References

9.2. Informative References

Author's Address