UNIVERSITY OF EXETER
COLLEGE OF ENGINEERING, MATHEMATICS AND PHYSICAL SCIENCES
COMPUTER SCIENCE
High Performance Computing
Copyright By PowCoder代写 加微信 powcoder
Continuous Assessment 2
Date Set: 10 March 2022 Date Due: 30 March 2022 Return Date: 4 May 2022
This CA comprises 60% of the overall module assessment.
This is an individual exercise and your attention is drawn to the College and University guidelines on collaboration and plagiarism, which are available from the College website.
This is the coursework for this module and tests your understanding of how to find numerical solutions to partial di↵erential equations, and parallel programming.
This continuous assessment consists of two parts. The first part comprises 40% of the overall module assessment and the second part comprises 20% of the overall module assessment.
1 The Manager-worker version of the Mandelbrot calcula- tion
The Mandelbrot set is the set of complex numbers which contains the points for which the iteration
zi+1 = zi2 + C (1)
remains finite, where C is a complex number and the initial value z0 = C. This apparently simple relation leads a fractal structure (see figure 1 which shows the Mandlebrot set plotted on the complex plane).
Figure 1: The Mandelbrot set plotted on the complex plane. The horizontal axis is the real axis and the vertical axis is the imaginary axis.
Workshop 8 looked at three di↵erent versions of the Mandelbrot calculation which had been parallelised using MPI. One of these programs used the manager-worker pattern to distribute work between MPI processes.
In the program the number of iterations at each point in the complex plane is stored as a two dimensional array called nIter. An element of the array is accessed using nIter[i][j] where the index i refers to the real axis and the index j refers to the imaginary axis. In the manager-worker version of the program the manger process tells the worker processes which value of the i index to work on. The worker processes loop over j to calculate a column of values in the complex plane (this is done in the calc_vals function). When they have completed one column they send a message to the manager process to request another i value to work on. Workers keep requesting work until all the values of i have been handed out. Each worker process stores the results it has calculated, and the results are collated at the end of the calculation using a call to MPI_Reduce (this is done in the do_communication function).
Using MPI_Reduce to collate the results sends more data than is required because each process which participates in the call sends the whole computational domain. The values in the domain
ECMM461: CA2
will includes zeros where a process has not calculated any values. The amount of data sent can be minimised if each worker process sends its results back to the the manger process each time it requesta a new i value. This means that each process only sends values it has calculated and there is no need to call MPI_Reduce to collate the results at the end of the calculation.
In this assignment you will modify the communication pattern in the manger-worker version of the Mandelbrot program so that worker processes send their results to the manger process after each column has been calculated instead of calling MPI_Reduce at the end of the calculation. You will then measure the impact of this optimisation on the parallel scaling of the program.
2 Optimising the communication pattern
Task 1: Modify the communication pattern in the manger-worker version of the Mandelbrot pro- gram so that worker processes send their results to the manger process after each column has been calculated instead of calling MPI_Reduce at the end of the calculation.
You have been provided with the original version of the manager-worker Mandelbrot program and for task 1 you need to optimise the communication pattern by making sure that the worker processes only send data that they have calculated. The following steps can be used to implement this optimisation:
1. Create bu↵er space for sending and receiving data. The bu↵er space needs to be the size of one column in the complex plane plus two extra values (see next point).
2. Modify the point-to-point communication calls so that the worker processes send their results back to the manger process after calculating each column (each i value). This can be done as part of the request for work: the worker process sends its rank, the i value of the completed column and the data in the column. These are all int values which can be packed into a single bu↵er.
3. You will need to handle the case of the initial handout where there is no data to be sent back to the manager process. This can be done by setting the i value of the column to a missing data value (e.g. a negative value) so that the manager process knows to discard the data.
4. The manager process receives the message, unpacks the column of values and stores it in the correct column of the nIter array. The manager process then hands out the next i value to the requesting worker process as before. When the calculation is complete the manager process will hold all the values in the nIter array.
5. You should now remove or comment out the call to do_communication as this is no longer required.
6. Write out the results from the manager process instead of the rank 1 process.
When you have modified the program you should check that the results match the results from the original program exactly. You should also check that all MPI processes exit cleanly after making a call to MPI_Finalize.
You may find the following advice useful:
• You do not have to use Isca to develop the MPI version of the program. This part of the assignment can be carried out using another computer with a C compiler and MPI library (for example the Lovelace lab PCs) if you prefer.
• The diff –brief command can be used to check whether the output from the modified version matches the output from the original version.
• If the output is not as expected it may help to initialise the nIter array to a missing data value (e.g. negative integer). This will make it clear where the array is not being filled with values.
• You may wish to decrease the problem size whilst testing the program but remember to change back to the original size for the performance tests in the next section.
ECMM461: CA2
3 Measuring the performance impact
Task 2: Measure the impact of the optimisation on the strong scaling of the calculation when run on Isca.
For this scaling test you should switch o↵ I/O (by setting doIO=false) and use N_RE=12000 and N_IM=8000. This part of the assignment needs to be carried out using the Isca HPC system. The workshop 8 tar file include job scripts which you can use to run these tests on Isca.
To carry out the scaling test follow these steps:
1. Run the original program using 2, 4, 8, 12 and 16 MPI process on one node, and 32 MPI processes on 2 nodes.
2. Run the modified program using 2, 4, 8, 12 and 16 MPI process on one node, and 32 MPI processes on 2 nodes.
3. Using your results calculate parallel speed-up data for the original and modified versions of the program. You may assume that the serial run time is 90.985 seconds (i.e. the value used in workshop 8).
4. Plot the parallel speed-up against the number of MPI processes for both the original and modified versions of the program.
4 Deliverables
There are two deliverables for this assignment:
1. The source code for your modified program. The submitted program should implement the point-to-point communication pattern described in these instructions and should produce output which identical to the original program. When the program is run all MPI processes should exit cleanly after making a call to MPI_Finalize.
2. A short report (guideline length 2–3 pages) which describes your modifications to the program and presents the results of your scaling study. The report should be presented in a consistent font no smaller than 11 point, with margins not less than 2cm, single column, and single space between lines. The report should include the following elements:
(a) A title
(b) A short introduction describing what the report is about
(c) A description of your modifications to the program. You may include snippets of source code if required.
(d) A plot showing the results from your scaling test. The plot should show parallel-speed up against number of MPI processes for both the original program and your modified version. The axes should be labelled and the two lines should be clearly distinguishable.
(e) A paragraph describing and assessing your results. You should say whether your modi- fications have improved the parallel performance of the program and quantify the size of any improvement.
The deliverables should uploaded to BART a single zip or tar file containing the source code for the modified version of the program, and the report. The deadline for submission is 12 noon 30 March 2022.
5 Mark scheme
A total of 100 marks are available for this assignment:
ECMM461: CA2
1. Task 1: Optimising the communication pattern (50 marks)
This task will be assessed based on the modified version of the program. Marks will be
awarded according to the following criteria:
• The program implements the required communication pattern correctly and the results are identical to the original program. When the program is run all MPI processes exit cleanly after making a call to MPI_Finalize: 50 marks
• The program implements the required communication pattern but a minor error causes the results to di↵er slightly from the original program. When the program is run all MPI processes exit cleanly after making a call to MPI_Finalize: 30 marks
• The program implements the required communication pattern but there is one major error. Either the results di↵er significantly from the original program or when the pro- gram is run all MPI processes do not exit cleanly after making a call to MPI_Finalize: 20 marks
• The program implements the required communication pattern but there are two major errors. The results di↵er significantly from the original program and when the program is run all MPI processes do not exit cleanly after making a call to MPI_Finalize: 10 marks
2. Task 2: Measuring the performance impact (50 marks)
This task will be assessed based on the on the report according to the following criteria:
(a) The report includes a title and a short introduction describing what the report is about:
(b) The report correctly describes all the modifications to the program: 10 marks
(c) The report includes a plot showing the results from the scaling test presenting data for both the original program and the modified version. The axes are labelled and the two lines are clearly distinguishable: 20 marks
(d) The report include a paragraph describing and assessing the results including a quanti- tive statement about whether the modifications have improved the parallel performance of the program: 10 marks
ECMM461: CA2
1 Introduction and Background
The 2020’s are expected to see the arrival of the first “exascale” supercomputers, capable of more than 1018 floating point operations per second as measured by the LINPACK benchmark. This development is considered to be an important milestone in High Performance Computing.
There are significant challenges associated with constructing an exascale supercomputer and making e↵ective use of its capability. Section 21.4 of Sterling, Anderson and Brodowicz (the HPC textbook for the module) discusses exascale computing and the challenges that it presents. The challenges associated with weather and climate models in particular are discussed in the paper “Crossing the Chasm: How to develop weather and climate models for next generation computers?” E↵orts are currently underway to address exascale computing challenges, including the Excalibur programme in the UK.
2 The Assignment
For this assessment you will write a 1000 word essay on one of the challenges associated with exascale computing. Your essay should describe the challenge and argue why it is vital to solve this challenge. You should also describe how the challenge is being addressed and evaluate the advantages and disadvantages of the potential solutions.
3 Guidance
The University Library has a Computer Science Subject Guide with links to a number of key resources (including the “Web of Science” bibliographic index) to help find relevant literature.
This assessment is subject to standard University rules on academic conduct and honesty. You can find the regulations here: http://as.exeter.ac.uk/academic-policy-standards/tqa-manual/ aph/managingacademicmisconduct/. The Essay Marking Criteria are given at the end of this doc- ument. In writing your essay, you should consider the following points:
• Quality of answer: Essay brief interpreted correctly and answered well.
• Structure and quality of writing: Clear and articulate. Specialist terminology defined where
appropriate. Proper use of sections, paragraphs and signposting.
• Use of literature: An appropriate range of relevant content from academic journals or other respected sources.
• Quality of argument: Claims justified by evidence. Conclusions follow from main body of essay.
• Presentation: Presented in a consistent font no smaller than 11 point, with margins not less than 2cm, single column, and single space between lines. Tables and figures of good quality with appropriate captions. Correct grammar and no spelling mistakes.
• Referencing: References should be consistently cited in an approved style, with all external material explicitly and unambiguously acknowledged. A list of approved referencing styles can be found at https://libguides.exeter.ac.uk/c.php?g=654150&p=4795420.
There are many resources available to help you write a good essay. Academic support on essay writing is available from ASSET. They provide several relevant online courses, including “Essay writing”, “Academic honesty and plagiarism”, and “Academic reading”.
ECMM461: CA2
4 Deliverables
The deliverable for this assessment is a 1000 word essay which is due at 12 noon on 30 March 2022.
• Your essay should be 1000 words (± 10%) in length. References are not included in the word limit.
• You may include diagrams, figures and/or tables if appropriate.
• Your essay needs to be submitted to e-BART
• This essay is worth 20% of the module grade for ECMM461.
ECMM461: CA2
程序代写 CS代考 加微信: powcoder QQ: 1823890830 Email: powcoder@163.com