Massively Parallel Algorithms - SS 2016
About this Course
There are big changes afoot. The era of increased performance from faster single cores and optimized single core programs has ended. Instead, highly parallel GPU cores, initially developed for shading, can now run hundreds or thousands of threads in parallel. Consequently, they are increasingly being adopted to offload and augment conventional (albeit multi-core) CPUs. And the technology is getting better, faster, and cheaper. It will probably even become a general computing processor on mobile devices, because it offers more processing power per energy amount.
The high number of parallel cores, however, poses a great challenge for software and algorithm design that must expose massive parallelism to benefit from the new hardware architecture. The main purpose of the lecture is to teach practical algorithm design for such parallel hardware.
Simulation is widely regarded as the third pillar of science (in addition to experimentation and theory). Simulation has an ever increasing demand for high-performance computing. The latter has received a boost with the advent of GPUs, and it is even becoming -- to some extent -- a commodity.
There are many scientific areas where the knowledge you will gain in this course can be very valuable and useful to you:
- Computer science (e.g., visual computing, database search)
- Computational material science (e.g., molecular dynamics simulation)
- Bio-informatics (e.g., alignment, sequencing, ...)
- Economics (e.g., simulation of financial models)
- Mathematics (e.g., solving large PDEs)
- Mechanical engineering (e.g., CFD and FEM)
- Physics (e.g., ab initio simulations)
- Logistics (e.g. simulation of traffic, assembly lines, or supply chains)
In this course, you will get hands-on experience in developing software for massively parallel computing architectures. For the first half of the lecture, there will be supervised exercises to familiarize yourself with the CUDA parallel programming model and environment. The exercises will comprise, for instance, image processing algorithms, such as you might find in Photoshop. In the second half, you can decide whether to continue with exercises, or whether to work on a small self-assignment for the rest of the cource.
- A little bit of experience with C/C++ ; note that we will need just plain old C during this course, but the concept of pointers should be familiar.
- Liking for algorithmic thinking.
Useful, but not required, is a computer/notebook with an Nvidia GPU that is CUDA capable. You can find a list of all supported GPU's here. If you don't have access to such a computer, you are welcome to work on your assignments in our lab.
Not required are
- experience with parallel programming,
- experience with computer graphics (although we will study several algorithms and applications of massively parallel algorithms in the area of visual computing)
The following table contains the topics and the accompanying slides (it will be filled step-by-step after the respective lectures).
Introduction: More Moore, stream programming model, Flynn's taxonomy,
overall speedup, Amdahl's law, Gustafson's law.
Fundamental Concepts of Parallel Programming 1: terminology, control flow, transfering data, blocks, data flow in general
|2.||Fundamental Concepts of Parallel Programming 2: multi-dimensional blocks and grids, classes in CUDA, constant memory, simplistic raytracer, warps, thread divergence, more details on the GPU architecture, warps, measuring performance, GPU/CPU synchronization, shared memory, parallel computation of the dot product, barrier synchronization, race conditions, document similarity|
|3.||Fundamental Concepts of Parallel Programming 3: utilizing lockstepped threads instad of explicit synchronization, heat transfer simulation, double buffering pattern, texture memory, edge detection (Sobel operator), GPU's memory hierarchy, parallel histogram computation, atomic operations, pipelining host-GPU calls|
|4.||Dense matrix algorithms: matrix vector product, column major storage, AoS versus SoA, coalesced memory access, auto-tuning, arithmetic intensity, matrix-matrix multiplication, blocked matrix multiplication, All Pairs Shortest Path as matrix multiplication, sparse matrices.|
|5.||Parallel prefix-sum 1: definition, simple examples, Hillis-Steele algorithm, depth & work complexity, Blelloch's algorithm, summed area tables, better percision for integral images, depth-of-field rendering (gathering & scattering), anisotropic texture filtering, face detection with Viola-Jones.|
|6.||Parallel prefix-sum 2: Brent's theorem & optimization, application to prefix-sum, theoretical speedup based on Brent, split operation, quick introduction into sequential radix sort, radix sort on the GPU, stream compaction.|
|7.||Parallel sorting: comparator networks, the 0-1 principle, odd-even mergesort, bitonic sorting, work and depth complexities, (digression: adaptive bitonic sorting)|
|8.||Applications of parallel sorting: BVH construction, space filling curves, Morton codes; deformable collision detection by sweep plane approach, PCA transformation, and clustering; faster ray-tracing by coherent ray packets.|
Dynamic Parallelism: general usage,
execution model, simple example, caveats,
Mariani-Silver algorithm for the Mandelbrot set,
the "W" pattern for multigrids.
Random Forests 1 (classification problem, simple solutions, concept of decision trees)
|10.||Random Forests 2: information gain, entropy, conditional entropy, construction of decision trees, problems of decision trees, wisdom of crowds, bootstrapping/subsampling, randomization during construction|
|11.||Random Forests 3: variants of RF's, out-of-bag error estimation, parallel construction of RFs, handwritten digit recognition, body tracking using depth images.|
|12./13.||Of Collisions and Spheres: collision detection basics, inner sphere trees, bacht neural gas, parallel bounding volume hierarchy construction, parallel distance queries for point clouds, parallel constant time collision detection, sphere packings|
|14.||Massively parallel hash tables and their applications; intersection of point clouds, geometric hashing, massively-parallel geometric hashing, cuckoo hashing, proof for fact that cuckoo hashing works with high probability|
Some very simple examples that can serve as a starting point. They contain a makefile and should compile out-of-the-box at least under Max OS X and Linux.
The following textbooks can help review the material covered in class:
- Jason Sanders, Edward Kandort: CUDA by Example. Addison-Wesley, Pearson Education.
Very easy reading, a very gentle introduction into CUDA, requires minimal C knowledge.
- Wen-Mei W. Hwu: GPU Computing Gems Jade Edition. Morgan Kaufmann.
If you are in the Uni Bremen campus LAN/WLAN, you can also read the e-book.
- David B. Kirk, Wen-Mei W. Hwu: Programming Massively Parallel Processors. Morgan Kaufmann.
- NVidia: CUDA C Programming Guide. (There is also a PDF version.)
- Russ Miller, Laurence Boxer: Algorithms, Sequenetial & Parallel. Cengage Learning.
Doesn't talk about technical details of implementing parallel algorithms, but takes more the theoretical, purely algorithmic perspective (uses PRAMs as algorithmic model).
Please note that the course is not based on one single textbook! Some topics might even not be covered in any current textbook! So, I suggest you first look at the books in the library before purchasing a copy.
If you plan on buying one of these books, you might want to consider buying a used copy -- they can often be purchased for a fraction of the price of a new one. Two good internet used book shops are Abebooks and BookButler.
Grades and Points achieved by the Assignments
For taking part in a so-called "Fachgespräch" (mini oral exam), you need a grade from the assignments >= 4.0 . You can get this by successfully completing at least 40% of all asignments.
- Reevaluating Amdahl's Law and Gustafson's Law (Source)
- Herb Sutter's Welcome to the Jungle from 2012 gives a number of inspiring insights into future trends -- and past trends that have ceased to continue. (Herb Sutter is widely respected in the C++ software engineering community.) (Source)
What Every Programmer Should Know About Memory by Ulrich Drepper, 2007
Well, I don't think that every programmer really has to know everything the paper explains, but I hope you all are interested in broadening your horizon -- who wants to know only that stuff she/he needs?
- Performance with constant memory (excerpt from CUDA by Example).
- CUDA Documents from NVidia
- Introduction to CUDA 5.0 (Source)
- Chapter 39. Parallel Prefix Sum (Scan) with CUDA by Mark Harris, Shubhabrata Sengupta, John D. Owens. In GPU Gems 3. (Source)
- A nice handout about PRAM algorithms and Brent's theorem by Siddhartha Chatterjee and Jan Prins. (Source)
- My article about Adaptive Bitonic Sorting. (A slightly updated version of my chapter in the Encyclopedia of Parallel Computing, David Padua, Ed., pages 146–157. Springer, 2011.)
- A very good explanation of BVH construction in this blog article by Tero Karras, with an application to collision detection on the first pages (Source)
- Here is a pretty comprehensive Introduction to Random Forests by Criminisi, Shotton, and Konukoglu. (Source)
- Literature for the chapter on parallel hashing:
Other Interesting Bits and Pieces
- Not exactly about massively parallel programming, but here is a clip from an interview with Linus Torvalds, where he speaks about tasteful code. And although he does not explicitely mention it, I strongly believe that tasteful code is what makes robust code. (Source: Linus Torvalds: The mind behind Linux, February 2016 at TED2016.)
- If you would like to see Hillis himself: a talk of Hillis at TED (not on his massively parallel algorithms, but on the perilous turn the internet might take) (Source)
Last modified: Mon Aug 29 16:38:15 CEST 2016