HPC GPU Cluster (Hybrid Computing - Heterogeneous Programming)
App-kernels-overview
The computational power of GPUs has widely
attracted the scientific community and GPUs provide unprecedented computational
power to solve the data intensive applications.
In the recent years, much attention has been gained for general purpose CPU (GPGPU)
processing. The word "general purpose" in the context of High Performance Computing
(HPC) usually means "data intensive applications in scientific and engineering fields.
Several categories of applications exist that demand very high levels of processing power.
The challenge is to develop
application software that transparently scales its parallelism to leverage
the increasing number of processor cores.
The evolution of the CUDA and OpenCL programming model has made it possible for the modern GPUs
to use massive multithreading for gaining huge application performance. Depending on
how well the algorithm lends itself to parallelization, the GPU implementations provide
correspondingly greater performance as compared to the CPU implementations.
A brief summary of application kernels which could benefit significantly by the use of GPUs
are discussed.
|
HPC GPU Cluster (Hybrid computing systems) consists of many core host CPUs
and 2 or 4 or 8 GPU accelerator devices or FPGA devices.
The CUDA enabled NVIDIA GPUs for High Performance Computing (HPC)
is commonly used in hybrid computing platforms and these accelerators
tailored for HPC GPU cluster. The Tesla (Fermi) GPUs for HPC are available
either as standard add-on boards, or in high-density self-contained IU rack
mount cases containing one or two or four GPU devices with independent power
and cooling, for attachment to rack-mounted HPC nodes.
In hyPACK-2013 workshop, a prototype HPC GPU cluster (CUDA /OpenCL enabled NVIDIA GPUs
& AMD-ATI OpenCL Prog. env) is used to solve
application kernels, that are based on heterogeneous
programming model.
In this workshop, programming and performance issues for applications on Heterogeneous on
HPC GPU Clusters will be discussed.
In laboratory session, a prototype Hybrid Heterogneous HPC GPU Cluster is made available,
which can address some of the Heterogeneous computing workloads.
The HPC GPU Cluster can be made "adaptive" to the
application it is running, assigning the most effective resources in real-time as
per application demands, without requiring modifications to the application. One of the objectives of HPC GPU Cluster (hybrid computing system) is
to allocate resources of CPUs & GPUs in an optimal way to solve applications of different
characteristics.
|
-
Computational Fluid Dynamics (CFD) : CFD Simulation involves solution of
Initial Value problems or Boundary Value Problems in which the solution
of governing partial differential equations (PDEs) by Finite Difference
Method (FDM) and Finite Element Method (FEM) is obtained. The generation of
structured /unstructured grids in complex three-dimensional regions, followed by
solution of matrix system of linear equations involves data-parallel computations.
These involve solution of matrix system of linear equations in which matrix is banded
or sparse and the respective equations are solved on a HPC GPU cluster.
-
Monte Carlo (MC) methods : These are stochastic techniques which are based on the use of
random numbers and probability statistics to investigate problems.
In a Monte Carlo simulation, a random value is selected for each of the tasks, based
on the range of estimates. The model is calculated based on this random value.
The result of the model is recorded and the process is repeated. A typical Monte Carlo simulation calculates the model hundreds or thousands of times, each time using different randomly-selected values.
When the simulation is complete, we have a large number of results from the model,
each based on random input values. These results are used to describe the likelihood,
or probability, of reaching various results in the model.
Monte Carlo estimation of Pi value will be done using a circle inscribed in
a unit square.
Given that the circle and the square have a ratio of areas that is Pi/4 /4,
the value of Pi
can be approximated using a Monte Carlo method.
N-body Simulation :
An N-body simulation
numerically approximates the evolution of a system of bodies in which each body
continuously interacts with every other body. N-body simulation arise in many
computational science problems such as astrophysical simulation, protein folding
which involve calculation of electrostatic and van der Walls Forces, turbulent flow
simulation, and global illumination computation in computer graphics.
Many Computational physics simulations involve computing the interactions of a
large number of particles or objects and force exists among these. If the force
between the particles is completely described by adding the forces between all pairs
of particles, and the force between each pair acts along the line between them,
that is called an N-body central force problem (N-body problem).
N-body problem
can be described with N items (the particles) but requires O(N2) computation
(all the pairs of particles). This is a good example of data-parallelism in
which performance and scalability can be achieved for large problems.
Important
step in implementing an N-body code is to decide how the particles are distributed
among the processes. To compute the forces on the particles, it is necessary for
each process to access all the particles on the other processes. One of the approaches
is for all processes to exchange all the particles and then compute with them.
The all-pairs approach to N-body simulation is a brute-force technique that evaluates
all pair-wise interactions among the N-bodies and simulation of large systems often
poses scalability in view of O(N2) computational complexity.
-
Image processing Kernels : Image Processing is gaining larger importance in a variety of application areas and it
deals with the manipulation and analysis of pictorial information. Image processing can
be a time consuming task and, parallel algorithms can be designed. Fast processing response
is a major requirement in many image processing applications. Even when the size of the
image is very large, typical vision systems involve real-time processing where a sequence
of image frames must be processed in a very short time.
Laplacian Edge Detection :
A common method for image processing
is pixel classification. Pixel classification defines a pixel's class based on one of its
features, in the case of edge detection (Laplacian Edge Detection); the feature examined is
its intensity versus the intensity of its neighbor pixels.
Face recognition: :
The face recognition using machines is an active research topic and is widely used nowadays
in disciplines like image processing, pattern recognition, and computer vision. The main
interest is to acquire facial information from images and it provide clue to understand several
commercial systems such as security and surveillance. A face image analysis includes face
detection, recognition, tracking and rendering. As the basis for all other related image
analysis of human faces, face detection and tracking are of great importance. The major
challenges on the issues of facial recognition which identifies a relationship between two
basics variables of the process are reliability/accuracy of the technique being used and
computational cost of this technique. It is important to develop image analysis algorithms
that can meet real-time constraints along with data capturing devices and the vast amounts
of data that they generate.
Image inpainting :
Image inpainting refers to the process of reconstructing the original image which has been
damaged due to factors such as ageing, wear and tear and occlusion. The challenge lies in
the fact that the observer seeing the inpainted image should not be able to guess that the
image had been tampered with. Commonly used inpainting technique is an exemplar-based techniques
in which searching the best exemplar or the best patch in the undamaged portion of the image
that will be used for filling the damaged portions of the image.
-
Computational Biology : Computational Biology applications such as
bioinformatics, and molecular modelling high performance is critical.
Biological sequence comparison is a very important operation
in Bioinformatics.
There are several methods to compare biological sequences, and some
of these have quadratic time and space complexity.
To-day, a huge amount of new
DNA sequences will need to be compared, in order to infer functional/structural characteristics.
Time spent in each comparison, as well as the accuracy of the result obtained, will be
challenging in genome projects.
Sequence comparison is, a very basic and important
operation in Bioinformatics. As a result of this step,
one or more sequence alignments can be produced.
A sequence alignment has a similarity score associated to it
that is obtained by placing one sequence above the other,
making clear the correspondence between the characters
and possibly introducing gaps into them. The most
common types of sequence alignment are global and local.
To solve a global alignment problem is to find the best
match between the entire sequences.
On the other hand,
local alignment algorithms must find the best match between
parts of the sequences.
Smith-Waterman (SW) is an exact algorithm based
on the longest common subsequence (LCS) concept that
uses dynamic programming to find local alignments between
two sequences of size m and n in
O(mn) space
and time. In order to accelerate these methods, many
GPU algorithms were proposed in the literature.
Smith Watermann (SW) algorithm : Parallelisation on HPC GPU Cluster
String Search algorithms
Pattern Matching or String Matching is very important subject which has application on
broad domain of text processing. The core part of the algorithm are also used in many
implementation of practical software existing under many application domain like "Speech
Processing", "DNA analysis" etc. Moreover, they emphasize programming methods that serve as
paradigms in other fields of computer science (system or software design). Theoretically, pattern
matching pose as a challenging problem under theoretical computer science.
Sequential pattern matching algorithms refers to the class of algorithm where a single
stream of execution do pattern matching activity at a time.
Brute force Pattern Matching algorithm is considered to
be most simple to understand and implement. This algorithm is also called as naive approach
of pattern matching.
characters of P and T left to right until either two unequal characters are found or until P is
exhausted, in which case an the occurrence of P is reported in side T on this particular location. On
the other hand, if match is not found then P is shifted one place right, and the same loop of
comparison is continued. This process continues up to end of sequence T.
Boyer Moore algorithm is considered to be the most efficient pattern matching algorithm.
The algorithm scan the characters of the pattern from right to left beginning with the right-most
one.
In case of mismatch, it may use two pre-compiled functions to shift the window to the right.
-
Video Processing :
The graphics performance of specialized software, e.g. scientific software,
image manipulation, video decoders/encoders, games that make GPU performance
is pretty important. In video processing, the GPU processor takes a video
stream and perform image processing operations (effects) on its content
using graphic APIs to generate a video stream that can be presented to the
display in real time. The scalability of video processing can be achieved
when multiple video streams are enhanced to combine multiple video streams
together to produce a single output stream.
On GPUs, hardware decoding with OpenVideo Decode API have been used. Many
decoding libraries exist for the various platforms supporting video
decode using the power of CPU Multi-core processors. Video processing
effects can be carried out on GPU. The CUDA enabled NVIDIA GPU or OpenCL
can be used to decode video on the GPU. Working in OpenCL and output video
display using OpenGL is one programming paradigm.
To decide a frame of video, we must fist initialize the decoding framework and request
that it open the file. The decode framework will then decode frame of video
in the designated format. The process of data is done using OpenCL.
Important steps of video Processing will be discussed in laboratory sessions.
-
Intrusion Detection System:
The attack attempts at the edge of the Internet are increasing nowadays. To prevent such attacks,
many high-performance intrusion detection systems (IDSes) have been employed. The demand for
a high-speed intrusion detection system (IDS) is increasing as high-bandwidth networks become
commonplace.
Securing the internal networks has become a common and crucial task that constantly deals with
flash crowds and external attacks. Detecting malicious attack patterns in the high-speed networks
poses number of performance challenges. The detection engine is required to monitor the network
traffic at line rate to identify potential intrusion attempts, for which it should execute efficient
pattern matching algorithms to detect thousands of known attack patterns in real time.
Today's high-performance IDS engines often meet these challenges with dedicated network processors,
special pattern matching memory , or regular expression matching on FPGAs.
|
|
|
|
