UCL logo

UCL Discovery

UCL home » Library Services » Electronic resources » UCL Discovery

Fault Tolerant Integer Data Computations: Algorithms and Applications

Anarado, IJ-F; (2017) Fault Tolerant Integer Data Computations: Algorithms and Applications. Doctoral thesis , UCL (University College London). Green open access

[img]
Preview
Text
Anarado_IjeomaAnarado_PhdThesis_final.pdf

Download (3MB) | Preview

Abstract

As computing units move to higher transistor integration densities and computing clusters become highly heterogeneous, studies begin to predict that, rather than being exceptions, data corruptions in memory and processor failures are likely to become more prevalent. It has therefore become imperative to improve the reliability of systems in the face of increasing soft error probabilities in memory and computing logic units of silicon CMOS integrated chips. This thesis introduces a new class of algorithms for fault tolerance in compute-intensive linear and sesquilinear (“one-and-half-linear”) data computations on integer data inputs within high-performance computing systems. The key difference between the proposed algorithms and existing fault tolerance methods is the elimination of the traditional requirement for additional hardware resources for system reliability. The first contribution of this thesis is in the detection of hardware-induced errors in integer matrix products. The proposed method of numerical packing for detecting a single error within a quadruple of matrix outputs is described in Chapter 2. The chapter includes analytic calculations of the proposed method’s computational complexity and reliability. Experimental results show that the proposed algorithm incurs comparable execution time overhead to existing algorithms for the detection and correction of a limited number of errors within generic matrix multiplication (GEMM) outputs. On the other hand, numerical packing becomes substantially more efficient in the mitigation of multiple errors. The achieved execution time gain of numerical packing is further analyzed with respect to its energy saving equivalent, thus paving the way for a new class of silent data corruption (SDC) mitigation method for integer matrix products that are fast, energy efficient, and highly reliable. A further advancement of the proposed numerical packing approach for the mitigation of core/processor failures in computing clusters (a.k.a., failstop failures) is described in Chapter 3 . The key advantage of this new packing approach is the ability to tolerate processor failures for all classes of sum-of-product computations. Because multimedia applications running on cloud computing platforms are now required to mitigate an increasing number of failures and outages at runtime, we analyze the efficiency of numerical packing within an image retrieval framework deployed over a cluster of AWS EC2 spot (i.e., low-cost albeit terminable) instances. Our results show that more than 70% reduction of cost can be achieved in comparison to conventional failure-intolerant processing based on AWS EC2 on-demand (i.e., higher-cost albeit guaranteed) instances. Finally, beyond numerical packing, we present a second approach for reliability in the case of linear and sesquilinear integer data computations by generalizing the recently-proposed concept of numerical entanglement. The proposed approach is capable of recovering from multiple fail-stop failures in a parallel/distributed computing environment. We present theoretical analysis of the computational and bit-width requirements of the proposed method in comparison to existing methods of checksum generation and processing. Our experiments with integer matrix products show that the proposed approach incurs 1.72% − 37.23% reduction in processing throughput in comparison to failure-intolerant processing while allowing for the mitigation of multiple fail-stop failures without the use of additional computing resources.

Type: Thesis (Doctoral)
Title: Fault Tolerant Integer Data Computations: Algorithms and Applications
Event: UNIVERSITY COLLEGE LONDON
Open access status: An open access version is available from UCL Discovery
Language: English
UCL classification: UCL > Provost and Vice Provost Offices > UCL BEAMS
UCL > Provost and Vice Provost Offices > UCL BEAMS > Faculty of Engineering Science
URI: http://discovery.ucl.ac.uk/id/eprint/1538661
Downloads since deposit
185Downloads
Download activity - last month
Download activity - last 12 months
Downloads by country - last 12 months

Archive Staff Only

View Item View Item