Page 236 - Invited Paper Session (IPS) - Volume 1
P. 236
IPS 151 Rub´en C.
Improvements on a parallel/distributed
algorithm for bootstrapping
Rub´en Carvajal-Schiaffino
Departamento de Matemática y Ciencia de la Computación
Universidad de Santiago de Chile
Abstract
Bootstrapping is a statistical technique that is very expensive from a
computational point of view. For this reason, parallel versions have been
developed, typically implemented on multicore architectures. However this
approach can be limited by the amount of cores and the bottleneck produced
by the simultaneous access to the memory. We present a comparison between
parallel implementations in shared memory, private memory and one based
on the use of massive parallel computing by means of the use of Graphics
Processing Unit (GPU).
Keywords
Bootstrap; Parallel Algorithms
1. Introduction
In the resolution of any problem it is necessary to consider what the
resources are available and therefore it is essential to choose the most
appropriate tools. In the case of the implementation of algorithms, the
resources are the processing time and the storage space. Considering these
two resources, the tools are: the programming language in which the
algorithm and the computation model are implemented. The programming
languages can be classified according to their execution mode: compiled (C,
FORTRAN), interpreted (R, MATHLAB) or running with virtual machine (JAVA).
Regarding the computer model this can be sequential, parallel or distributed.
Since the bootstrap method requires a large number of operations, it is
best to choose a language compiled as C. Regarding the computational model
for bootstrap in addition to using the sequential model, it is possible to
perform a parallel implementation. Parallel computing models can be
classified into those that use shared memory where several processors can
access a common memory such as pthreads[1] or GPU’s[4] and those that use
private memory where the processors collaborate with each other through
message exchange [2,3].
In this paper we present a comparison between the sequential
implementation of the boot-strap method to calculate the Pearson correlation
coefficient, the parallel implementation using the cores of a computer and the
parallel implementation using GPUs.
225 | I S I W S C 2 0 1 9