代做Cross-Platform留学生、代写OptimisedCode、代写C/C++程序、代做C/C++设计

日期：2018-11-04 09:41

Scalable Cross-Platform Software Design: Optimisation Coursework#1

25% of the module mark: Marking Scheme on Moodle

Part 1: Code Optimisation.

Summary: On Moodle there is a very poorly written computational program, ToBeOptimisedCode.cpp. It is a “made

up“ code that doesn’t actually do anything useful, but it has many features common to a range of large-scale

numerical simulation codes in engineering. This is a rather poorly written code for a number of reasons!

I compiled it with g++ -o task ToBeOptimisedCode.cpp and ran it with time./task on Jenna node comp00. The run

time data on the screen was,

real 0m14.322s

user 0m13.998s

sys 0m0.022s

Objective: Your task is to improve the C coding of the algorithm to minimise the run time; you must not change the

code’s function in any way. Changing language, direct use of machine code or parallelisation is not allowed.

Please make sure you get exactly the same results: run the original code as above and make a copy of the file

produced, data_out. Then for any new version check for identical results by doing,

diff data_out original_ data_out

You should get no differences at all.

Specific Assessment Requirements:

Submission through the Moodle page by the deadline shown there. Submit a single zip file containing all files required

to simply load and run your project. A brief pdf document is required that briefly describes:

a) A copy of the original listing annotated with brief comments to identify poor programming style and where you

think there is scope to improve the code: comments should cover both “coding bad practice” and “speedup”

possibilities.

b) Your optimised code (suitably documented).

c) A brief (1/2 page) pdf summarising the results of your optimisation including discussion of modifications that

you tried that did not yield benefits and any resources (e.g. profiling) that you drew on.

Part 2: Characterisation of the Jenna platform.

Setup a very simple serial C program to add two double precision arrays.

a[i]=b[i]+c[i+offset] for all 0<i<n-offset.

The arrays are to be dynamically allocated. The arrays b and c are to be initialised so that b[i]=c[i]=i

Time the execution of the loop above (i.e. excluding the array setup and initialisation).

To get reasonable timing accuracy it may be necessary to wrap an outer loop around the above loop so as to repeat

the above sum-loop many times.

To do the timings you might use gettimeofday, but there are alternatives – if you trust them and they are sufficiently

accurate.

#include <sys/time.h>

struct timeval start_time,end_time;

gettimeofday(&start_time,NULL);

// Your code to time

gettimeofday(&end_time,NULL);

cout<<"\n\ngettimeofday wall time="<<

end_time.tv_sec - start_time.tv_sec+(end_time.tv_usec-start_time.tv_usec)/1e6;

Specific Assessment Requirements:

a) With offset=0, investigate and plot the timings for different n.

b) Based upon the previous results, select a few "interesting" values of n and repeat for different offset values.

c) Now repeat (a) and (b) using OpenMP to parallelise the loop.

d) You should expect that static scheduling will give the best performance for (c). Please design, implement nd

characterise an alternative loop-based test where dynamic scheduling might prove advantageous.

Required results: Submit both your test codes and the results obtained and in a brief report (< 2pages excluding

graphs) summarise, discuss and explain what you observe.