DidaWiki

2h: Introduction to the course, structure of the lessons, hardware evolution and the urgency of parallel/distributed computing dictated by parallel hw (multicore CPUs, GPUs, clusters).

Assignment: look at top500.org and green500.org and try to figure out evolution of parallel machines through the “statistics” graphs in the lists.

2h: Examples of parallel execution of different kind of real life tasks:

• translating a book
• counting the euros in pockets of people in a room
• assembling an object out of pieces p1, …, pn to be assembled in order

How the tasks may be executed by a single person and by number of cooperating persons. Different way of executing tasks in parallel. Effect of “slow” persons recruited to execute a task. Figuring out general principles out of this: patterns, mechanisms, overheads, …

Blackboard

2h: Finding concurrency aspects:

• concurrent activities
• decomposition strategies
• computation grain
• concurrent activity graph
• maximum parallelism degree and minimum time

2h: Mechanisms to implement parallel computations on multicores

• C++ threads: creation, join, usage of shared memory with locks and condition variables, async tasks
  • std::thread
  • std::async
  • mini tutorial
• Cilk: cilk_spawn and cilk_sync, sample code
  • Home page
  • cilk_spwan and cilk_sync
  • mini tutorial

(apologize I miss the last period of the lesson of the afternoon, due to my mistake)

BlackBoard1BlackBoard2

Assigment: write a program using Cilk and C++ threads “translating” a book in parallel, where:

• the input book is a char[]
• the translation consist in capitalizing all letters (e.g. “Ciao, come va?” → “CIAO, COME VA?”)

2h: Time and parallelism related measures

• how to measure time (std::chrono, time (Unix) command
• Speedup, scalability and efficiency
• commands to filter benchmark output (script, grep, awk, sort, at)

Blackboard

Assignment: write a program using C++ threads counting word occurrences in a text.

2h: Access to lab machines. More on mechanisms in C++ for concurrency.

• see also this link or this link

Blackboard

2h: overheads and counter measures

Blackboard

Assignment: look for overheads in translate book and word count assignments.

2h: more on overheads (NUMA memory, false sharing). Introduction to vectorization (with sample code). Tools for checking sequential/vector code performance.

• Vectorization guide Intel
• Vtune Getting Started
• Vtune start page

Blackboard

Assignment (1): single Jacobi iteration (as in the blackboard PDF). Create randomly generated matrix A and vectors B and X and run a single iteration of the Jacobi method verifying that the code is properly vectorized and measuring the increase in speed due to vectorization (optimization of the sequential code in general)

Assignment (2): consider a large matrix (4-8K rows/columns at least) and measure the improvement w.r.t. non optimized sequential computation of a single Jacobi iteration

• when only vectorization is used
• when the problem is parallelized using C++11 threads (only)
• when you use threads (coarse grain parallelisation) and vectorization (fine grain parallelisation)

2h: Introduction to parallel patterns/skeletons

• using OpenMP to implement parallel for (example of simple data parallel pattern) (Summary of 4.0, Old Tutorial slides). OpenMP is available in g++ and in icc by specifying the flag -fopenmp.

• using GrPPI as high level programming model for skeletons (needs c++14)

Blackboard

Assignment

• read the Chapter on skeletons and the one on design patterns in the Course notes book (pdf on the web site, Chap. 3 and 6)
• try to re-write the initial assignments using OpenMP parallel for
• if you succeed with openMP try to use GrPPI for the same assignments

2h:

Sample code

• code with different implementation of a “map” …

Parallel design patterns and algorithmic skeletons

• concepts, similarities and differences
• Usage of performance models and rewriting rules (this is the reading key for the chapter 10 (sections 10.1 to 10.3)

Blackboard

Assignment:

• implement a divide and conquer computation using c++ threads. The computation should look for a word in a text. Recursively divide the text up to the point you either get something of the a given length (parameter), and in this case you look for the string sequentially. Return an integer value counting the occurrences of the string. (sub assignment: consider programming the solution in such a way the code for the divide and conquer strategy may be used to program another divide and conquer application)

2h

• data parallel patterns (map, reduce, google map reduce, stencil, stencil-reduce)
• implementation of pattern frameworks (supporting composition)
  • template based
  • data flow based

PDF

Assignments

• read Chapter 4 of the notes
• implement the word count with your own google mapreduce in C++ (use threads to compute in parallel the map phase and again threads to compute in parallel the reduce phase)

2h

• FastFlow programming framework: Introduction, core patterns.
• documentation may be found at calvados.di.unipi.it
• download code (via SVN!!) at SourceForge

PDF

Assignments

• try to write a FF pipeline (with core patterns) that implements the Sieve of Eratostene: one stage emits the stream of integers and stage i filters out all the numbers multiple of i. The last stage prints the results. Try generating 20 integers and using k<20 filter stages.

by David Del Rio Astorga: 1h

• Introduction to GrPPI (slides sent as PDF to the course mailing list).

Morning: 2h

• methodology to develop parallel code with structured programming
• final project presentation (see the exams page)

H1H2

Afternoon: 2h (By Massimo Torquati)

• high level usage of core FastFlow patterns
• Slides
• Additional code files