Parallel Computing for Data Science: With Examples in R, C++ and CUDA

Introduction to Parallel Processing in R
Recurring Theme: The Principle of Pretty Good Parallelism
A Note on Machines
Recurring Theme: Hedging One's Bets
Extended Example: Mutual Web Outlinks

"Why Is My Program So Slow?": Obstacles to Speed
Obstacles to Speed
Performance and Hardware Structures
Memory Basics
Network Basics
Latency and Bandwidth
Thread Scheduling
How Many Processes/Threads?
Example: Mutual Outlink Problem
"Big O" Notation
Data Serialization
"Embarrassingly Parallel" Applications

Principles of Parallel Loop Scheduling
General Notions of Loop Scheduling
Chunking in Snow
A Note on Code Complexity
Example: All Possible Regressions
The partools Package
Example: All Possible Regressions, Improved Version
Introducing Another Tool: multicore
Issues with Chunk Size
Example: Parallel Distance Computation
The foreach Package
Stride
Another Scheduling Approach: Random Task Permutation
Debugging snow and multicore Code

The Shared Memory Paradigm: A Gentle Introduction through R
So, What Is Actually Shared?
Clarity and Conciseness of Shared-Memory Programming
High-Level Introduction to Shared-Memory Programming: Rdsm Package
Example: Matrix Multiplication
Shared Memory Can Bring a Performance Advantage
Locks and Barriers
Example: Finding the Maximal Burst in a Time Series
Example: Transformation of an Adjacency Matrix
Example: k-Means Clustering

The Shared Memory Paradigm: C Level
OpenMP
Example: Finding the Maximal Burst in a Time Series
OpenMP Loop Scheduling Options
Example: Transformation an Adjacency Matrix
Example: Transforming an Adjacency Matrix, R-Callable Code
Speedup in C
Run Time vs. Development Time
Further Cache/Virtual Memory Issues
Reduction Operations in OpenMP
Debugging
Intel Thread Building Blocks (TBB)
Lockfree Synchronization

The Shared Memory Paradigm: GPUs
Overview
Another Note on Code Complexity
Goal of This Chapter
Introduction to NVIDIA GPUs and CUDA
Example: Mutual Inlinks Problem
Synchronization on GPUs
R and GPUs
The Intel Xeon Phi Chip

Thrust and Rth
Hedging One's Bets
Thrust Overview
Rth
Skipping the C++
Example: Finding Quantiles
Introduction to Rth

The Message Passing Paradigm
Message Passing Overview
The Cluster Model
Performance Issues
Rmpi
Example: Pipelined Method for Finding Primes
Memory Allocation Issues
Message-Passing Performance Subtleties

MapReduce Computation
Apache Hadoop
Other MapReduce Systems
R Interfaces to MapReduce Systems
An Alternative: "Snowdoop"

Parallel Sorting and Merging
The Elusive Goal of Optimality
Sorting Algorithms
Example: Bucket Sort in R
Example: Quicksort in OpenMP
Sorting in Rth
Some Timing Comparisons
Sorting on Distributed Data

Parallel Prefix Scan
General Formulation
Applications
General Strategies for Parallel Scan Computation
Implementations of Parallel Prefix Scan
Parallel cumsum() with OpenMP
Example: Moving Average

Parallel Matrix Operations
Tiled Matrices
Example: Snowdoop Approach to Matrix Operations
Parallel Matrix Multiplication
BLAS Libraries
Example: A Look at the Performance of OpenBLAS
Example: Graph Connectedness
Solving Systems of Linear Equations
Sparse Matrices

Inherently Statistical Approaches: Subset Methods
Chunk Averaging
Bag of Little Bootstraps
Subsetting Variables

Appendix A: Review of Matrix Algebra
Appendix B: R Quick Start
Appendix C: Introduction to C for R Programmers