This article was first published in the Summer 2016 issue of Usenix ;login:
Traditional IP routing systems route packets by comparing the destinaton address against a predefined list of routes to each available subnet; but when multiple potential routes exist between two hosts on a network, the preferred route may be dependent on context that cannot be inferred from the destination alone. The Linux kernel, together with the iproute2 suite, supports the definition of multiple routing tables and a routing policy database to select the preferred routing table dynamically. This additional expressiveness can be used to avoid multiple routing pitfalls, including asymmetric routes and performance bottlenecks from suboptimal route selection.
The CU-Boulder Research Computing environment spans three datacenters, each with its own set of special-purpose networks. A traditionally-routed host simultaneously connected to two or more of these networks compounds network complexity by making only one interface (the default gateway) generaly available across network routes. Some cases can be addressed by defining static routes; but even this leads to asymmetric routing that is at best confusing and at worst a performance bottleneck.
Over the past few months we've been transitioning our hosts from a single-table routing configuration to a policy-driven, multi-table routing configuration. The end result is full bidirectional connectivity between any two interfaces in the network, irrespective of underlying topology or a host's default route. This has reduced the apparent complexity in our network by allowing the host and network to Do the Right Thing™ automatically, unconstrained by an otherwise static route map.
Linux policy routing has become an essential addition to host configuration in the University of Colorado Boulder "Science Network." It's so useful, in fact, that I'm surprised a basic routing policy isn't provided by default for multi-homed servers.
The problem with traditional routing
The simplest Linux host routing scenario is a system with a single network interface.
# ip addr show 1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00 inet 127.0.0.1/8 scope host lo valid_lft forever preferred_lft forever 2: ens192: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast state UP qlen 1000 link/ether 00:50:56:88:56:1f brd ff:ff:ff:ff:ff:ff inet 10.225.160.38/24 brd 10.225.160.255 scope global dynamic ens192 valid_lft 60184sec preferred_lft 60184sec
Such a typically-configured network with a single uplink has a single default route in addition to its link-local route.
Traffic to hosts on
10.225.160.0/24 is delivered directly,
while traffic to any other network is forwarded to
A dual-homed host adds a second network interface and a second link-local route; but the original default route remains. (Figure 1.)
# ip addr show 1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00 inet 127.0.0.1/8 scope host lo valid_lft forever preferred_lft forever 2: ens192: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast state UP qlen 1000 link/ether 00:50:56:88:56:1f brd ff:ff:ff:ff:ff:ff inet 10.225.160.38/24 brd 10.225.160.255 scope global dynamic ens192 valid_lft 86174sec preferred_lft 86174sec 3: ens224: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc pfifo_fast state UP qlen 1000 link/ether 00:50:56:88:44:18 brd ff:ff:ff:ff:ff:ff inet 10.225.176.38/24 brd 10.225.176.255 scope global dynamic ens224 valid_lft 69193sec preferred_lft 69193sec # ip route list default via 10.225.160.1 dev ens192 10.225.160.0/24 dev ens192 proto kernel scope link src 10.225.160.38 10.225.176.0/24 dev ens224 proto kernel scope link src 10.225.176.38
The new link-local route provides access to hosts on
10.225.176.0/24, and is sufficient for a private network
connecting a small cluster of hosts. In fact, this is the
configuration that we started with in our Research Computing
.160.0/24 is a low-performance "management"
.176.0/24 is a high-performance "data" network.
In a more complex network, however, link-local routes quickly become insufficient. In the CU Science Network, for example, each datacenter is considered a discrete network zone with its own set of "management" and "data" networks. For hosts in different network zones to communicate, a static route must be defined in each direction to direct performance-sensitive traffic across the high-performance network route. (Figure 2.)
Though managing these static routes can be tedious, they do sufficiently define connectivity between the relevant network pairs: "data" interfaces route traffic to each other via high-performance networks, while "management" interfaces route traffic to each other via low-performance networks. Other networks (e.g., the Internet) can only communicate with the hosts on their default routes; but this limitation may be acceptable for some scenarios.
Even this approach is insufficient, however, to allow traffic between
"management" and "data" interfaces. This is particularly problematic
when a client host is not equipped with a symmetric set of network
interfaces. (Figure 3.) Such a client may only have a "management"
interface, but should still communicate with the server's
high-performance interface for certain types of traffic. (For example,
a dual-homed NFS server should direct all NFS traffic over its
high-performance "data" network, even when being accessed by a client
that itself only has a low-performance "management" interface.) By
default, the Linux rp_filter blocks this traffic, as the server's
response to the client targets a different route than the incomming
request; but even if
rp_filter is disabled, this asymmetric
route limits the server's aggregate network bandwidth to that of its
The server's default route could be moved to the "data" interface--in some scenarios, this may even be preferable--but this only displaces the issue: clients may then be unable to communicate with the server on its "management" interface, which may be preferred for certain types of traffic. (In Research Computing, for example, we prefer that administrative access and monitoring not compete with IPC and file system traffic.)
Routing policy rules
Traditional IP routing systems route incoming packets based solely on the the intended destination; but the Linux iproute2 stack supports route selection based on additional packet metadata, including the packet source. Multiple discrete routing tables, similar to the virtual routing and forwarding (VRF) support found in dedicated routing appliances, define contextual routes, and a routing policy selects the appropriate routing table dynamically based on a list of rules.
In this example there are three different routing contexts to consider. The first of these--the "main" routing table--defines the routes to use when the server initiates communication.
A separate routing table defines routes to use when responding to traffic on the "management" interface. Since this table is concerned only with the default route's interface in isolation, it simply reiterates the default route.
Similarly, the last routing table defines routes to use when responding to traffic on the "data" interface. This table defines a different default route: all such traffic should route via the "data" interface.
With these three routing tables defined, the last step is to define routing policy to select the correct routing table based on the packet to be routed. Responses from the "management" address should use table 1, and responses from the "data" address should use table 2. All other traffic, including server-initiated traffic that has no outbound address assigned yet, uses the "main" table automatically.
With this routing policy in place, a single-homed client (or, in fact, any client on the network) may communicate with both the server's "data" and "management" interfaces independently and successfully, and the bidirectional traffic routes consistently via the appropriate network. (Figure 4.)
Persisting the configuration
This custom routing policy can be persisted in the Red Hat "ifcfg"
network configuration system by creating interface-specific
# cat /etc/sysconfig/network-scripts/route-ens192 default via 10.225.160.1 dev ens192 default via 10.225.160.1 dev ens192 table mgt # cat /etc/sysconfig/network-scripts/route-ens224 10.225.144.0/24 via 10.225.176.1 dev ens224 default via 10.225.176.1 dev ens224 table data # cat /etc/sysconfig/network-scripts/rule-ens192 from 10.225.160.38 table mgt # cat /etc/sysconfig/network-scripts/rule-ens224 from 10.225.176.38 table data
The symbolic names
data used in these examples
are translated to routing table numbers as defined in the
Once the configuration is in place, activate it by restarting the
network service. (e.g.,
systemctl restart network) You may
also be able to achieve the same effect using
ifup on individual interfaces.
Red Hat's support for routing rule configuration has a confusing
regression that merits specific mention. Red Hat (and its derivatives)
has historically used a "
network" initscript and subscripts to
configure and manage network interfaces, and these scripts support the
rule- configuration files. Red Hat Enterprise
Linux 6 introduced NetworkManager, a persistent daemon with additional
functionality; however, NetworkManager did not support
files until version 1.0, released as part of RHEL 7.1. If you're
currently using NetworkManager, but wish to define routing policy in
rule- files, you'll need to either disable NetworkManager
entirely or exempt specific interfaces from NetworkManager by
NM_CONTROLLED=no in the relevant
In a Debian-based distribution, these routes and rules can be
post-up directives in
We're still in the process of deploying this policy-based routing configuration in our Research Computing environment; and, as we do, we discover more cases where previously complex network requirements and special-cases are abstracted away by this relatively uniform configuration. We're simultaneously evaluating other potential changes, including the possibility of running a dynamic routing protocol (such as OSPF) on our multi-homed hosts, or of configuring every network connection as a simultaneous default route for fail-over. In any case, this experience has encouraged us to take a second look at our network configuration to re-evaluate what we had previously thought were inherent limitations of the stack itself.