COMP 322: Fundamentals of Parallel Programming
Vivek Sarkar is Professor and Chair of Computer Science at Rice University, where he holds the E.D. Butcher Chair in Engineering. He conducts research in multiple aspects of parallel software including programming languages, program analysis, compiler optimizations and runtimes for parallel and high performance computer systems. He currently leads the Habanero Extreme Scale Software Research group at Rice University, and serves as Associate Director of the NSF Expeditions Center for Domain-Specific Computing. He teaches the undergraduate class on Fundamentals of Parallel Programming in the spring semesters, and an advanced graduate class in the fall semesters. His professional biography and curriculum vitae can be found here.
The prerequisite course requirements are COMP 182 and COMP 215. COMP 322 should be accessible to anyone familiar with the foundations of sequential algorithms and data structures, and with basic Java programming. COMP 221 is also recommended as a co-requisite.
The goal of COMP 322 is to introduce you to the fundamentals of parallel programming and parallel algorithms, using a pedagogic approach that exposes you to the intellectual challenges in parallel software without enmeshing you in the jargon and lower-level details of today's parallel systems. A strong grasp of the course fundamentals will enable you to quickly pick up any specific parallel programming model that you may encounter in the future, and also prepare you for studying advanced topics related to parallelism and concurrency in more advanced courses such as COMP 422.
To ensure that students gain a strong knowledge of parallel programming foundations, the classes and homeworks will place equal emphasis on both theory and practice. The programming component of the course will mostly use the Habanero-Java Library (HJ-lib) pedagogic extension to the Java language developed in the Habanero Multicore Software Research project at Rice University. The course will also introduce you to real-world parallel programming models including Java Concurrency, MapReduce, MPI, OpenCL and CUDA. An important goal is that, at the end of COMP 322, you should feel comfortable programming in any parallel language for which you are familiar with the underlying sequential language (Java or C). Any parallel programming primitives that you encounter in the future should be easily recognizable based on the fundamentals studied in COMP 322.
COMP 322 provides the student with a comprehensive introduction to the building blocks of parallel software, which includes the following concepts:
- Primitive constructs for task creation & termination, synchronization, task and data distribution
- Abstract models: parallel computations, computation graphs, Flynn's taxonomy (instruction vs. data parallelism), PRAM model
- Parallel algorithms for data structures that include arrays, lists, strings, trees, graphs, and key-value pairs
- Common parallel programming patterns including task parallelism, pipeline parallelism, data parallelism, divide-and-conquer parallelism, map-reduce, concurrent event processing including graphical user interfaces.
These concepts will be introduced in three modules:
1. Deterministic Shared-Memory Parallelism: creation and coordination of parallelism (async, finish), abstract performance metrics (work, critical paths), Amdahl's Law, weak vs. strong scaling, data races and determinism, data race avoidance (immutability, futures, accumulators, dataflow), deadlock avoidance, abstract vs. real performance (granularity, scalability), collective & point-to-point synchronization (phasers, barriers), parallel algorithms, systolic arrays
2. Nondeterministic Shared-Memory Parallelism and Concurrency: critical sections, atomicity, isolation, high level data races, nondeterminism, linearizability, liveness/progress guarantees, actors, request-response parallelism, Java Concurrency, locks, condition variables, semaphores, memory consistency models.
3. Distributed-Memory Parallelism and Locality: memory hierarchies, cache affinity, data movement, message-passing (MPI), communication overheads (bandwidth, latency), MapReduce, accelerators, GPGPUs, CUDA, OpenCL, energy efficiency, resilience.
There are no required textbooks for the class. Instead, lecture handouts are provided for each module as follows:
- Module 1 handout (Deterministic Shared-Memory Parallelism)
- Module 2 handout (Nondeterministic Shared-Memory Parallelism and Concurrency)
- Module 3 handout (Distributed-Memory Parallelism and Locality)
You are expected to read the relevant sections in each lecture handout before coming to the lecture. We will also provide a number of references in the slides and handouts.
There are also a few optional textbooks that we will draw from quite heavily. You are encouraged to get copies of any or all of these books. They will serve as useful references both during and after this course:
Grading, Honor Code Policy, Processes and Procedures:
Grading will be based on your performance on:
- Six homeworks (weighted 40% in all),
- Two exams (weighted 20% each),
- Weekly lecture & lab quizzes (weighted 10% in all), and
- Class participation (weighted 10% in all).
The purpose of the homework's is to train you to solve problems and to help deepen your understanding of concepts introduced in class. Homeworks are due on the dates and times specified in the course schedule. Please turn in all your homeworks using the CLEAR turn-in system. Homework is worth full credit when turned in on time. A 10% penalty per day will be levied on late homeworks, up to a maximum of 6 days. No submissions will be accepted more than 6 days after the due date.
You will be expected to follow the Honor Code in all homeworks, quizzes and exams. All submitted homeworks are expected to be the result of your individual effort. You are free to discuss course material and approaches to homework problems with your other classmates, the teaching assistants and the professor, but you should never misrepresent someone else’s work as your own. If you use any material from external sources, you must provide proper attribution (as shown here).
Exams 1 and 2 and all quizzes are pledged under the Honor Code. They test your individual understanding and knowledge of the material. Collaboration on quizzes and exams is strictly forbidden. Quizzes are open-book and exams are closed book. Finally, it is also your responsibility to protect your homeworks, quizzes and exams from unauthorized access. Graded homeworks will be returned to you via email, and exams as marked-up hardcopies. If you believe we have made an error in grading your homework or exam, please bring the matter to our attention within one week.
Accommodations for Students with Special Needs:
Students with disabilities are encouraged to contact me during the first two weeks of class regarding any special needs. Students with disabilities should also contact Disabled Student Services in the Ley Student Center and the Rice Disability Support Services