Scalable Clustering

Scalable Clustering

The Cisco caching solution was designed to enable network administrators to easily cluster cache engines to scale high traffic loads. This design approach allows customers to linearly scale performance and cache storage as cache engines are added. For example, a single Cisco Cache Engine 590 can support a 45-Mbps WAN link and 144 GB of cache storage; adding a second Cisco Cache Engine 590 provides support for a 90-Mbps WAN link and 288 GB of cache storage. Up to 32 cache engines can be clustered together.

This linear scalability is achieved because of the manner in which WCCP-enabled routers redirect traffic to cache engines. WCCP-enabled routers perform a hashing function on the incoming request's destination IP address, mapping the request into one of 256 discrete buckets. Statistically, this hashing function distributes incoming requests evenly across all buckets. In addition, these buckets are evenly allocated among all cache engines in a cluster. WCCP-enabled routers ensure that a certain cache engine deterministically fulfills requests for a certain destination IP address on the Internet. Empirically, this distribution algorithm has consistently demonstrated even load distribution across a cache engine cluster. Most of the popular Web sites have multiple IP addresses, thus preventing uneven load distribution.

When the customer adds a new cache engine to the cluster, the WCCP-enabled router detects the presence of the new cache engine and reallocates the 256 buckets to accommodate the additional cache engine. For example, the simplest installation using one router and one cache engine assigns all 256 buckets to the single cache engine. If a customer adds another cache engine, the WCCP-enabled router redirects packets to the two cache engines evenly—128 buckets are allocated to each cache engine. If the customer adds a third cache engine, the WCCP-enabled router assigns 85 or 86 buckets to each of the three cache engines.

Customers can hot-insert cache engines into a fully operating cache cluster. In this situation, the WCCP-enabled router automatically reallocates the buckets evenly among all cache cluster members, including the new cache engine. Because a new cache engine will not have any content, it will incur frequent cache misses until enough content has been populated in its local storage. To alleviate this cold startup problem, the new cache engine, for an initial period, sends a message to the other cache cluster members to see if they have the requested content. If they have the content, they will send it to the new cache engine. Once the new cache engine determines it has retrieved enough content from its peers (based on configurable numbers), it will handle cache misses by directly requesting the content from the end server rather than from its peers.

Fault Tolerance and Fail Safety

If any cache engine in a cache cluster fails, the cluster automatically heals itself. The WCCP-enabled router redistributes the failed cache engine's load evenly among the remaining cache engines. The cache cluster continues operation using one less cache engine, but operation is otherwise unaffected. The Cisco network caching solution enables an WCCP-enabled, Multigroup Hot-Standby Router Protocol (MHSRP) router pair to share a cache engine cluster, creating a fully redundant caching system.

This is referred to as WCCP multihoming. If the WCCP-enabled router fails, existing Cisco IOS fault tolerance and fail-safe mechanisms are applied. For example, a hot-standby router could dynamically take over operations, redirecting Web requests to the cache cluster. If an entire cache cluster fails, the WCCP-enabled router automatically stops redirecting traffic to the cache cluster, sending clients' Web requests to the actual destination Web site in the traditional fashion.

This loss of the entire cache cluster can appear to users as an increase in download time forWeb content, but has no other significant effect. This designed-in, failsafe response is made possible because the cache cluster is not directly in line with clients' other network traffic.