A null convention logic based platform for high speed low energy IP packet forwarding

Dabholkar, P 2018, A null convention logic based platform for high speed low energy IP packet forwarding, Doctor of Philosophy (PhD), Engineering, RMIT University.

Document type: Thesis
Collection: Theses

Attached Files
Name Description MIMEType Size
Dabholkar.pdf Thesis application/pdf 6.23MB
Title A null convention logic based platform for high speed low energy IP packet forwarding
Author(s) Dabholkar, P
Year 2018
Abstract By 2020, it is predicted that there will be over 5 billion people and 38.5 billion Internet-ofThings devices on the Internet. The data generated by all these users and devices will have to be transported quickly and efficiently. Routers forming the backbone of this Internet already support multiple 100 Gbps ports meaning that they would have to perform upwards of 200 Million destination addresses lookups per second in the packet forwarding block that lies in the router ‘data-path’. At the same time, there is also a huge demand to make the network infrastructure more energy efficient. The work presented in this thesis is motivated by the observation that traditional synchronous digital systems will have increasing difficulty keeping up with these conflicting demands. Further, with reducing device geometries, extremes in “process, voltage and temperature” (PVT) variability will undermine reliable synchronous operation. It is expected that asynchronous design techniques will be able to overcome many of these problems and offer a means of lowering energy while maintaining high throughput and low latency. This thesis investigates existing address lookup algorithms and investigates the possibility of combining various approaches to improve energy efficiency without affecting lookup performance. A quasi delay-insensitive asynchronous methodology - Null Convention Logic (NCL) - is then applied to this combined design. Techniques that take advantage of the characteristics of the design methodology and the lookup algorithm to further improve the area, energy and latency characteristics are also analysed. The IP address lookup scheme utilised here is a recent algorithmic approach that uses compact binary-tries and was selected for its high memory efficiency and throughput. The design is pipelined, and the prefix information is stored in large RAMs. A Boolean synchronous implementation of the algorithm is simulated to provide an initial performance benchmark. It is observed that during the address lookup process nearly 68% of the trie accesses are to nodes that contained no prefix information. Bloom filter structures that use non-cryptographic hashes and single-bit memory are introduced into the address lookup process to prevent these unnecessary accesses, thereby reducing the energy consumption. Three non-cryptographic hashing algorithms (CRC32, Jenkins and Murmur) are also analysed for their suitability in Bloom filters, and the CRC32 is found to offer the most suitable trade-off between complexity and performance. As a first step to applying the NCL design methodology, NCL implementations of the hashing algorithms are created and evaluated. A significant finding from these experiments is that, unlike Boolean systems, latency and throughput in NCL systems are only loosely coupled. An example Jenkins hash implementation with eight pipeline stages and a cycle time of 3.2 ns exhibits a total latency of 6 ns, whereas an equivalent synchronous implementation with a similar clock period exhibits a latency of 25.6 ns. Further investigations reveal that completion detection circuits within the NCL pipelines impair throughput significantly. Two enhancements to the NCL circuit library aimed particularly at optimising NCL completion detection are proposed and analysed. These are shown to enable completion detection circuits to be built with the same delay but with 30% smaller area and about 75% lower peak current compared to the conventional approach using gates from the standard NCL library. An NCL SRAM structure is also proposed to augment the conventional 6-T cell array with circuits to generate the handshaking signals for managing the NCL data flow. Additionally, a dedicated column of cells called the Null-storage column is added, which indicates if a particular address in the RAM stores no Data, i.e., it is in its Null state. This additional hardware imposes a small area overhead of about 10% but allows accesses to Null locations to be completed in 50% less time and consume 40% less energy than accesses to valid Data locations. An experimental NCL-based address lookup system is then designed that includes all of the developed NCL modules. Statistical delay models derived from circuit-level simulations of individual modules are used to emulate realistic circuit delay variability in the behavioural modules written in Verilog. Simulations of the assembled system demonstrate that unlike what was observed with the synchronous design, with NCL, the design that does not employ Bloom filters, but only the Null-storage column RAMs for prefix storage, exhibits the smallest area on the chip and also consumes the least energy per address lookup. It is concluded that to derive maximum benefit out of an asynchronous design approach; it is necessary to carefully select the architectural blocks that combine the peculiarities of the implemented algorithm with the capabilities of the NCL design methodology.
Degree Doctor of Philosophy (PhD)
Institution RMIT University
School, Department or Centre Engineering
Subjects Computer Communications Networks
Circuits and Systems
Keyword(s) Null Convention Logic
IP Address Lookup
Low energy
High speed
Version Filter Type
Access Statistics: 56 Abstract Views, 19 File Downloads  -  Detailed Statistics
Created: Mon, 01 Oct 2018, 14:49:36 EST by Adam Rivett
© 2014 RMIT Research Repository • Powered by Fez SoftwareContact us