Overview:

Several studies have inferred the presence of considerable amounts of redundancy in Internet traffic content, clearly establishing the tremendous scope for performance improvements achievable through appropriate exploitation of the redundancies. However, the Internet today is memoryless. The focus of this project is to rethink this aspect of the Internet, specifically in the context of wireless data networks. Generically, equipping wireless data networks with a network memory will enable memorization of content as it flows naturally (or by design) through the network, and more importantly, using the memorized content to increase performance and lower cost of delivering any content to its intended destination.

Impact:

Using real wireless network traffic traces in a campus network spanning over 40 days, we have found that performance improvements of up to 75% can be achieved if network memory were to be supported. We also observe that neither existing coarse caching strategies nor trivial packet level caching extensions tap into these improvements possible. In this context, we study, design, and develop a network memory for wireless data networks, which we refer to as Wireless Memory. Wireless Memory solution will directly benefit network operators in both reducing operating costs and offering considerably better performance to their subscribers. In addition, users’ actually transmitted data sizes can be reduced by eliminating redundancy. Reduced data size will result in improved network performance including increased throughput and lower response time. This is particularly true for wireless networks, since the wireless media is often shared by multiple users, and data transmission is subject to various collision scenarios where smaller packet sizes are preferred.

Concept:

Wireless Memory works between wireless base station and clients. The simplest fashion in which this memory works is as follows. Assume the base station S has to deliver certain data d to a client C at time T1. That data is memorized by both S and C. Later at time T2, if S wants to send another information which contains d, S can retrieve d from C’s memory by sending d’s reference to C. The retrieval command sent from S to C is generally much smaller than the raw data. For the data that is not available in memory, S sends them as-is. Within this simple model, we explore multiple fundamental issues, such as; the granularities at which data should be memorized, how to enable dynamic learning of the receiver’s memory state by the sender and what memory operators to support in addition to store/retrieve.

The basic elements of Wireless-Memory are illustrated in Figure 1. Wireless-Memory works at the packet-level and has two types of operations:

1. Memory Referencing sequentially breaks the payload into sequence of data segments, performs memory lookup to determine the presence of segment in memory, and composes a new packet containing new segments and codes of memorized segments,
2. Memory De-referencing contains the complementary components of packet de-composition which separates raw segments and codes, looks up the memory to find segments encoded by the codes and assembles them to recreate the original data packet.

Results:

We evaluate Wireless Memory using real wireless network traces collected in 3 buildings of a major university campus. The Wi-Fi networks sniffed are two 802.11g networks. Figure 2 shows the improvement in aggregate throughput of the top 8 users in one building for a span of 7 days. The baseline scenario, with no memory, has an aggregate throughput of 6.24 Mbps. Wireless Memory provides an improvement from 16% up to 93%. The average improvement is about 40%.

Figure 1: Basic elements of Wireless Memory

Figure 2: Throughput benefits

People:

• Shruti Sanadhya (Studen)
• Raghupathy Sivakumar (Professor)