WCCP is de-facto semi-standard used by routers to redirect network traffic to caches. It is available on most Cisco™ routers although it does not appear to be officially supported by Cisco. The primary benefits of WCCP are
- If already have a router that supports WCCP inline you do not not have to change your network topology.
- WCCP fails open so that if the Traffic Server machine fails, it is bypassed and users continue to have Internet access.
Use of WCCP only makes sense for client side transparency1 because if the clients are explicitly proxied by Traffic Server there's no benefit to WCCP fail open, as the clients will continue to directly access the unresponsive Traffic Server host. It would be better to adjust the routing tables on the router for explicit proxying.
Because the router serves as the inline network element, Traffic Server must run on a separate host. This host can be located anywhere as long as Traffic Server is either on the same network segment or a GRE tunnel can be maintained between the Traffic Server host and the router.
This document presumes that the router is already properly configured to handle traffic between the clients and the origin servers. If you are not certain, verify it before attempting to configure Traffic Server with WCCP. This is also a good state to which to revert should the configuration go badly.
Setting WCCP is a three step process, first configuring the router, the Traffic Server host, and Traffic Server.
The router will not respond to WCCP protocol packets unless explicitly configured to do so. Via WCCP, the router can be made to perform packet interception and redirection needed by Traffic Server transparency. The WCCP protocol in effect acts as means of controlling a rough form of policy routing with positive heartbeat cutoff.
The Traffic Server host system must also be configured using
iptables to accept connections on foreign addresses. This is done roughly the same way as the standard transparency use.
Finally Traffic Server itself must be configured for transparency and use of WCCP. The former is again very similar to the standard use, while WCCP configuration is specific to WCCP and uses a separate configuration file (example), referred to by the
The primary concern for configuration in which of three basic topologies are to be used.
- Dedicated -- Traffic Server traffic goes over an interface that is not used for client nor origin server traffic.
- Shared -- Traffic Server traffic shares an interface with client or server traffic.
- Inside Shared -- Traffic Server and client traffic share an interface.
- Outside Shared -- Traffic Server and origin server traffic share an interface.
In general the dedicated topology is preferred. However, if the router has only two interfaces one of the shared topologies will be required2. Click the links above for more detailed configuration information on a specific topology.
Shared interface issues
A shared interface topology has additional issues compared to a dedicated topology that must be handled. Such a topology is required if the router has only two interfaces, and because of these additional issues it is normally only used in such cases, although nothing prevents it use even if the router has three or more interfaces.
The basic concept for a shared interface is to use a tunnel to simulate the dedicated interface case. This enables the packets to be distinguished at layer 3. For this reason, layer 2 redirection cannot be used because the WCCP configuration cannot distinguish between packets returning from the origin server and packets returning from Traffic Server as they are distinguished only by layer 2 addressing3. Fortunately the GRE tunnel used for packet forwarding and return can be used as the simulated interface for Traffic Server.
Frequently encountered problems
MTU and fragmentation
In most cases the basic configuration using a tunnel in any topology can fail due to issues with fragmentation. The socket logic is unable to know that its packets will eventually be put in to a tunnel which will by its nature have a smaller MTU than the physical interface which it uses. This can lead to pathological behavior or outright failure as the packets sent are just a little too big. It is not possible to solve easily by changing the MTU on the physical interface because the tunnel interface uses that to compute its own MTU.
1 Server side transparency should also be used, but it is not as significant. In its absense, however, origin servers may see the source address of connections suddenly change from the Traffic Server address to client addresses, which could cause problems. Further, the primary reason for not having server side transparency is to hide client addresses which is defeated if the Traffic Server host fails.
2 If your router has only one interface, it's hardly a router.
3 This is not fundamentally impossible, as the packets are distinct in layer 2. However, the WCCP implementation does not seem to be able to use this information to correctly route packets. Further testing should be done.