COMP90025 Parallel and multicore computing

COMP90025 Parallel and multicore computing Project 1: Spell checker — identifying words Project 1 requires developing and submitting a C/C++ OpenMP program and a written report. Project 1 is worth 20% of the total assessment and must be done individually. The due date for the submission is nominally 11.59pm on 11 September, the Wednesday of Week 8. Submissions received after this date will be deemed late. A penalty of 10% of marks per working day late (rounded up) will be applied unless approved prior to submission. The weekend does not count as working days. A submission any time on Thursday will attract 10% loss of marks and so on. Please let the subject coordinator know if there is anything unclear or of concern about these aspects. 1 Background We have all come to rely on spelling checkers, in everything from word processors to web forms and text messages. This year’s projects will combine to form a simple parallel spell checker. In Project 1, you will write OpenMP (shared memory) code to collect a list of distinct words from a document, ready to check the spelling of each. In Project 2, you will write MPI (message passing) code to take such a list of words and check them against a dictionary of known words. 2 Assessment Tasks

1. Write an OpenMP (or pthreads) C/C++ function void WordList(uint8_t *doc, size_t doc_len, int non_ASCII, uint16_t ***words, int *count) to do the following: (a) take a pointer doc to a UTF-8 string and its length, doc len; (b) parse the string into a list of words, where a word is defined as a maximal string of characters that are Unicode letter type, as determined by iswalnum (or isalnum for ASCII) in the current locale (not necessarily the standard C locale). That is, it is a string of Unicode letters that ? is either preceded by a non-letter or starts at the start of the string; and ? is either followed by a non-letter or ends at the end of the string; The function should malloc two regions of memory: one containing all of the (‘\0’- terminated) words contiguously, and the other containing a list of pointers to the starts of words within the first region. A pointer to the second should be returned in words. The memory should be able to be freed by free (**words); free (*words); (c) set *count to the number of distinct words. The array should be sorted according to the current locale. The C function strcoll compares according to the current locale. We will test with locales LC ALL=C, en AU. In bash, the locale can be specified on the command line, like 1 LC ALL=en AU run find words text1-ASCII.txt LC ALL=en AU run find words text1-en.txt or it can be specified using export export LC ALL=en AU run find words text1-ASCII.txt run find words text1-en.txt There should be no spaces around the “=”. If nonASCII is zero, then you can assume that the UTF-8 input is all ASCII, and optimize for that case. If nonASCII is 1, then you can assume that non-ASCII characters are rare. If nonASCII is 2, then non-ASCII characters may be common. This is not required, unless you are aiming to get the fastest possible code. You must write your own sorting code. (Choose a simple algorithm first, and replace it if you get time.) Do not use any container class libraries. You can assume that all UTF-16 characters fit into a single uint16_t. The driver program to call your code, and a skeleton Makefile, are available at https: //canvas.lms.unimelb.edu.au/files/20401320/download?download_frd=1 and on Spartan at /data/projects/punim0520/2024/Project1.
2. Write a minor report (3000 words (+/- 30%) not including figures, tables, diagrams, pseu- docode or references). The lengths given below are guidelines to give a balanced report; if you have a good reason to write more or less, then you may. Use the following sections and details: (a) Introduction (400 words): define the problem as above in your own words and discuss the parallel technique 代 写COMP90025 Parallel and multicore computing that you have implemented. Present the technique using parallel pseudo-code. Cite any relevant literature that you have made use of, either as a basis for your code or for comparison. This can include algorithms you chose not to use along with why you didn’t use them. If you use an AI assistant like ChatGPT, then clearly identify which text was based on the AI output, and state in an appendix what prompt you used to generate that text. (b) Methodology (500 words): discuss the experiments that you will use to measure the per- formance of your program, with mathematical definitions of the performance measures and/or explanations using diagrams, etc. (c) Experiments (500 words): show the results of your experiments, using appropriate charts, tables and diagrams that are captioned with numbers and referred to from the text. The text should be only enough to explain the presented results so it is clear what is being presented, not to analyse result. (d) Discussion and Conclusion (1600 words): analyze your experimental results, and dis- cuss how they provide evidence either that your parallel techniques were successful or otherwise how they were not successful or, as may be the case, how the results are inconclusive. Provide and justify, using theoretical reasoning and/or experimental evi- dence, a prediction on the performance you would expect using your parallel technique if the number of threads were to increase to a much larger number; taking architectural aspects and technology design trends into account as best as you can – this may require some speculation. For each test case, there will be a (generous) time limit, and code that fails to complete in that time will fail the test. The time limit will be much larger than the time taken by the sequential skeleton, so it will only catch buggy implementations. 2 (e) References: cite literature that you have cited in preparing your report. Use, for example, the ACM Style guide at authors.acm.org/proceedings… tion-information/preparing-your-article-with-latex for all aspects of formatting your report, i.e., for font size, layout, margins, title, authorship, etc. 3 Assessment Criteria Your code will be tested on Spartan. Although you are welcome to develop on your own system if that is convenient, it must compile and run on Spartan to get marks for “speed” and “correctness” (see below). Assessment is divided between your written report and the degree of understanding you can demonstrate through your selection and practical implementation of a parallel solution to the problem. Your ability to implement your proposed parallel solution, and the depth of understand- ing that you show in terms of practical aspects, is called “software component”. In assessing the software component of your project the assessor may look at your source code that you submit and may compile and run it to verify results in your report. Programs that fail to compile, fail to provide correct solutions to any problem instances, and/or fail to provide results as reported may attract significant loss of marks. The remaining aspects of the project are called “written report”. In assessing the written report, the assessor will focus solely on the content of the written report, assessing a range of aspects from presentation to critical thinking and demonstration of a well designed and executed methodology for obtaining results and drawing conclusions. The assessment of software component and written report is weighted 40/60, i.e., 40% of the project marks are focussed on the software component (mostly for speed and correctness) and 60% of the project marks are focussed on the written report. Assessing a written report requires significant qualitative assessment. The guidelines applied for qualitative assessment of the written report are provided below. 4 Quality Assessment Guidelines A general rubric that we are using in this subject is provided below. It is not criteria with each criterion worth some defined points. Rather it is a statement of quality expectations for each grade. Your feedback for your written assessment should make it clear, with respect to the quality expectations below, why your submission has received a certain grade, with exact marks being the judgement of the assessor. Bear in mind that for some assignments, independent discussions (paragraphs) in the assignment may be attracting independent assessment which later sums to total the final mark for the assignment. As well, please bear in mind that assessors are acting more as a consistent reference point than as an absolute authority. Therefore while you may disagree with the view point of the assessor in the feedback that is given, for reasons of consistency of assessment it is usually the case that such disagreement does not lead to changes in marks. Quality expectations: Marks will be awarded according to the following rubric. Speed (20%) H1 Speed-up likely to be O(p/ log(p)) for p processors for sufficiently large prob- lems. Probably faster with two processors than the best sequential algorithm. Among the fastest in the class. H2 Speed-up likely to be O(√n) for sufficiently large problems Faster than most submissions H3 Some speedup for large problems and large numbers of processors P Performance is similar to a sequential algorithm. N Markedly slower than a good sequential implementation. 3 Evaluation set-up (30%) H1 Thorough and well-focused testing for correctness. Thorough and well-focused tests of scalability with problem size. Thorough and well-focused tests of scal- ability with number of processors. Good choice of figures and tables. Reports on times of appropriate sub-tasks. Good choice of other tests in addition to scaling H2 Makes a reasonable effort at all of the above, except perhaps “other tests”. Test probably less well focused (testing things that don’t matter, and/or missing things that do). Reports on times of some sub-tasks. H3 At least one class of evaluation is absent or of poor quality P Basic testing and evaluation. Probably one or two well-chosen graphs, or multiple less-suitable graphs N Major problems with evaluation methodology: insufficient, unclear or mislead- ing Discussion (30%) H1 Excellent. Clearly organised, at the level of subsections, paragraphs and sen- tences. Insightful. Almost no errors in grammar, spelling, word choice etc.. You Show potential to do a research higher degree. H2 Very good. Clear room for improvement in one of the areas above. H3 Good. Discussion addresses the questions, but may be unclear, missing impor- tant points and/or poorly expressed. P Acceptable, but with much room for improvement. Discussion may not address important aspects of the project, or may have so many errors that it is difficult to follow. N Correctness (20%) Your code is expected to be correct. Unlike other in categories, you can get negative marks for having wildly incorrect results. 20 Correct results 15 One error, or multiple errors with a clear explanation in the text 0–14 Multiple errors with inadequate explanation -10– -0 Major errors -20– -10 May crash, give completely ridiculous results or fail to compile When considering writing style, The “Five C’s of Writing” is adapted here as a guideline for writing/assessing a discussion: Clarity - is the discussion clear in what it is trying to communicate? When sentences are vague or their meaning is left open to interpretation then they are not clear and the discussion is therefore not clear. Consistency - does the discussion use consistent terminology and language? If different terms/ language are/is used to talk about the same thing throughout the discussion then it is not consistent. Correctness - is the discussion (arguably) correct? If the claims in the discussion are not logically sound/reasonable or are not backed up by supporting evidence (citations), or otherwise are not commonly accepted, then they are not assumed to be correct. Conciseness - is the discussion expressed in the most concise way possible? If the content/ meaning/understanding of the discussion can be effectively communicated with less words, then the discussion is not as concise as it could be. Each sentence in the discussion should be concisely expressed. Completeness - is the discussion completely covering what is required? If something is missing from the discussion that would have significant impact on the content/meaning/understanding 4 of what it conveys, then the discussion is incomplete. 5 Submission Your submission must be via Canvas (the assignment submission option will be made available closer to the deadline) and be either a single ZIP or TAR archive that contains the following files: ? find words.c or find words.cc: Ensure that your solution is a single file, and include com- ments at the top of your program containing your name and student number, and examples of how to test the correctness and performance of your program, using additional test case input files if you require. ? Makefile: A simple makefile to compile your code, and link it to the driver run find words.c provided. ? find words.slurm: A SLURM script to run your code on Spartan. ? Report.pdf: The only acceptable format is PDF. ? test case inputs: Any additional files, providing test case inputs, that are needed to follow the instructions you provide in your solution’s comments section. (Note that we will also test your code on our own test cases to check for correctness.) 6 Other information
3. AI software such as ChatGPT can generate code, but it will not earn you marks. You are allowed to use tools like ChatGPT, but if you do then you must strictly adhere to the following rules. (a) Have a file called AI.txt (b) That file must state the query you gave to the AI, and the response it gave (c) You will only be marked on the differences between your final submission and the AI output. If the AI has built you something that gains you points for Task 1, then you will not get points for Task 1; the AI will get all those points. If the AI has built you something that gains no marks by itself, but you only need to modify five lines to get something that works, then you will get credit for identifying and modifying those five lines. (d) If you ask a generic question like “How do I control the number of threads in OpenMP?” or “What does the error ‘implicit declaration of function rpc close server’ mean?” then you will not lose any marks for using its answer, but please report it in your AI.txt file. If these rules seem too strict, then do not use generative AI tools.
4. There are some sample datasets in /data/projects/punim0520/2024/Project1, which you can use for debugging. Most are too small to use to test timing, but you can extract long sequences from abstractsi 1-1166.txt.gz, which are the abstracts from articles in medical journals. You do not need to copy these to your home directory. You can either specify the full path on the command line, or create a “symbolic link”, which makes it appear that the file is in your directory, with the command ln -s ?full path? . 5 where the final “.” means “the current directory”. It can be replaced by a filename. If you run with a large input file, please remember to delete the output file once you are finished with it.
5. If you want to use prefix sum (I would), you can use the OpenMP 5.0 “scan” directive. A simple example is given at developers.redhat.com/blog/2021/0… tures-in-openmp-5-0-and-5-1 Implement all other parallel algorithms yourself.
6. You can ignore Unicode codepoints that require more than 16 bits in UTF-16.
7. When describing speed-up, remember that this is defined relative to the speed of the best possible sequential algorithm, not your chosen algorithm run with a single thread.
8. It is good to show a breakdown of the time spent in various components.
9. Your submission will be evaluated on ASCII text, French Unicode text (mostly ASCII but with some non-ASCII characters) and Hindi text (mostly non-ASCII). How can you use that knowledge to help you write faster code?

WX：codinghelp