Resource and Awareness RE operations Essay

A. Resource and Awareness RE involves resourceintensive operations such as indexing content, looking up content fingerprints and compressing data, and reconstructing the original content from locally stored information. An ideal approach must explicitly account for the resource constraints on network elements in performing these RE functions. These constraints arise mainly from (a) throughput bounds which depend on the number of memory operations possible per second and (b) memory capacity which limits the amount of data that can be cached for RE purposes. Nave approaches that do not account for these constraints, such as the straw man framework of Anand et al, offer suboptimal performance. In contrast, using the limited resources available at each node intelligently can offer close to the best possible benefits.

B. Network Wide Analysis

The architecture should allow network operators to specify networkwide goals such as increasing overall efficiency (e.g., improving the network throughput) or to achieve specific traffic engineering goals (e.g., alleviating congested hotspots).

C. Flexibility

The architecture must be incrementally adoptable providing benefits even under partial deployment, and must supplement, not replace, current network operations such as existing routing and network management practices. We present the design, implementation, and evaluation of Smart RE, an architecture for networkwide RE that meets the above requirements.

In SmartRE, redundancy elimination is performed in a coordinated fashion by multiple devices. Smart RE uses the available resources on RE devices efficiently and naturally accommodates several networkwide objectives. In describing SmartRE, we focus largely on packetlevel RE in

ISP networks [12], where RE devices on routers cache packet payloads and strip duplicate strings from individual packets. However, we believe that our design can apply to other deployment scenarios, e.g., in multihop wireless networks and datacenters. In SmartRE, a packet can potentially be reconstructed or decoded several hops downstream from the location where it was compressed or encoded. In this respect, SmartRE represents a significant departure from packetlevel RE designs proposed in prior solutions [29, 12], where each compressed packet is reconstructed at the immediate downstream router. Further, SmartRE uses a network wide coordinated approach for intelligently allocating encoding and decoding responsibilities across network elements. In general, encoding incurs greater overhead than decoding. Thus, SmartRE allocates encoding to ingress routers to avoid overloading interior routers that operate at higher linerates and thus have stricter resource constraints. Since the number of edge routers is large, a large number of encoded packets are introduced into the network. Interior routers in SmartRE perform less expensive decoding actions. Decoding is performed in a coordinated fashion with each interior router responsible for storing and reconstructing a fraction of the encoded packets on a path. We use hashbased sampling techniques [31] to facilitate coordination across interior

routers while incurring negligible overhead When allocating encoding and decoding responsibilities across a network, SmartRE takes into account the memory capacity and packet processing throughput at each RE device along with the prevailing traffic conditions, and configures the actions of different devices so as to best meet an operatorspecified networkwide goal.