Why does Riverbed's Universal Data Store architecture matter?

In a typical WAN optimization deployment, employees located at different remote branch offices need fast accelerated access to centrally-stored data on servers located the main data center.  When deploying WAN optimization into such an environment, Riverbed's universal data store architecture is essential for scalability and storage efficiency, especially for large deployments involving many branch offices. 

Riverbed's Universal Data Store

A universal data store maintains byte-level data in a single shared dictionary that is used by all peer WAN optimization devices.  A common file, web object, or any other data that is fetched to multiple remote sites is only represented a single time in the central WAN optimization device.  This property ensures optimal usage of the available storage capacity in the central Riverbed Steelhead device.

However, most non-Riverbed vendors use a per-peer data store architecture.  The properties of a per-peer data store are illustrated below:

Per-Peer Data Store (Cisco WAAS, Blue Coat, Juniper, most others)

A per-peer data store uses a separate dictionary to store and map byte-level data for each remote peer device.  Common files, emails, and other data that are sent to multiple remote peer devices must be stored multiple times in the central WAN optimization device (once for each peer).  This approach results in storage inefficiency in the central device in the main data center.  Consider the example of where a single email with a 100MB attachment is distributed to employees located at 30 different remote sites.  Such an email consumes at about 3GB of storage capacity in a WAN optimization device that uses a per-peer data store architecture; only 100MB of storage is consumed when using Riverbed's Steelhead solution.

Why this all matters:

When used to support networks with more than a few remote sites, a per-peer data store results in fragmentation of the data store capacity of the central WAN optimization device.  Poor performance results when the effective available data store capacity is inadequate.  A phenomenon known as "data store wrapping" occurs, where stored byte-level data is overwritten and erased prematurely in order to make available scarce storage capacity for new data actively being sent over the WAN.  The result is disk churn, where the WAN optimization device is constantly writing data to disk, and yet much of that newly written data is erased before it can be used even once.  While the WAN optimization device is constantly busy processing and writing data, nevertheless it delivers poor performance because this data store wrapping issue prematurely erases the disk-resident data needed to optimize newly-arriving data.

In small networks with simple topologies, the performance issues described above may not become apparent.  This includes in lab tests and limited-scale POC's, which many customers use to evaluate the efficacy of a given WAN optimization solution.  Unfortunately, some customers assume that the performance they observe in a limited-scale POC will be comparable to what they will obtain after full-scale deployment.  Tragically, these customers discover this problem only after months of frustration in an ultimately-failed deployment effort, when their chosen per-peer data store product did not deliver hoped-for results.