Title

Active Storage Networks: Topology, Routing and Application

Date of Completion

January 2011

Keywords

Engineering, Electronics and Electrical|Computer Science

Degree

Ph.D.

Abstract

High performance computing systems are often inhibited by the performance of their storage systems and their ability to deliver data. Active Storage Networks (ASN) provide an opportunity to optimize storage system and computational performance by offloading some computation to the network switch. An ASN is based around an intelligent network switch that allows data processing to occur on data as it flows through the storage area network from storage nodes to client nodes. A key design element for an ASN is the switching topology. In this thesis, we present an ASN switching topology named 2-Dilated flattened butterfly (2DFB) which is a nonblocking, low latency, low cost network compared to other nonblocking interconnecting networks. We have implemented this network topology using the NetFPGA as the basic building block of the switching network. The ASN 2DFB architecture has been used in a variety of applications including data sort, data search, data clustering, and min-max. ^ We have also developed an adaptive load balanced routing scheme (ALDFB) which exploits the topological properties of 2DFB network. ALDFB always gives priority to forwarding packets through the minimal path and therefore, for local and benign traffic the performance of this routing scheme is equal to that of the minimal routing. In adversarial traffic, ALDFB provides better load balance by one non minimal forwarding in each dimension. ALDFB provides high throughput on adversarial traffic patterns and provides better latency on benign traffic patterns. We have compared the performance of ALDFB on a 2DFB network with non-minimal global adaptive routing (UGAL), Minimal Adaptive and Adaptive Clos routing algorithm for different traffic patterns. We observed that a 2DFB network with ALDFB routing provides high throughput with reduced latency compared to other routing schemes for all the traffic patterns. ^ Finally, we show how a 2DFB-based ASN can be used to improve parallel file system performance. We have done simulations of striping files and file writes using file locking protocols in a parallel file system. In an ASN we offload some operations from the end-terminals to the ASN switch. In the case of file striping, the splitting of files and the parity calculations are done on the run inside ASN switch. In the file locking case we offload the file locking protocol to the ASN switch. In both cases we observe a significant reduction in traffic through the network and this helps an ASN based parallel file system to offer significant performance improvement. ^