CS257 Advanced Computer Architecture
Coursework Assignment
1 Introduction
The purpose of this coursework is to give you some hands-on experience in code
optimisation. By the time that you read this you will have encountered a variety
of code optimisation techniques including loop unrolling and vectorisation.
2 Submission
Your submission will consist of two parts:
- Optimised Code (70%)
A piece of C code based on the initial implementation provided. This C
code will be assessed with respect to your selection and understanding of
optimisations, functional correctness, i.e., producing the right answer, and
execution speed. - Written Report (30%)
A report (3 pages maximum, excluding references) detailing your design
and implementation decisions. Your report will be evaluated with respect
to your understanding of code optimisation techniques and the optimisations
you attempted. This means that your report should explain:
(a) which optimisations you used/didn’t use;
(b) why your chosen optimisations improve performance; and
(c) how your chosen optimisations affect floating-point correctness.
Given that you may apply many different optimisations, a sensible approach
is to build your solution incrementally, saving each partial solution and documenting
the impact of each optimisation you make. This means that it is in
your interest to attempt as many different optimisations, or combinations of
optimisations as you can.
You may discuss optimisation techniques with others but you are not allowed
to collaborate on solutions to this assignment. Please remember that the
University takes all forms of plagiarism seriously.
1 - Introduction to ACACGS
ACSCGS is a conjugate gradient proxy application for a 3D mesh. The simulation
will execute for either a fixed number of timesteps, or alternatively until
the residual value falls below a given threshold. This is done for a given mesh
size, which is passed in at runtime through command line arguments.
In this proxy application, a force is applied to each edge boundary of the
cuboid, which is then propagated throughout the mesh. As each time step
passes, the force is dissipated within the mesh, until the amount of residual is
significantly small that the simulation stops (as there is no more calculations to
perform), or a set number of time steps have passed.
In addition to providing numeric solutions, the code can also generate visuals
which depict the pressure within the mesh throughout the simulation run.
Creating the visualisations relies on two optional packages, Silo and VisIt, which
are available on the DCS systems.
Figure 1: Pressure Matrix Visualisation
2 - Compiling and Running the Code
The code includes a makefile to build the program. You can compile all of the
code using the command make. You may modify the makefile if you wish to,
but please ensure you include it with your submission. If you do not include a
makefile, then the default one will be used.
While the DCS machines do include a version of gcc, it is preferable to use a
more recent version. On the DCS systems, you can make version 9 the default
by using the‘module load gcc9’command. Once this is loaded you can simply
type‘make’to build the code, which will create an executable named‘acacgs’
in the directory.
To run the code, you need to provide the three dimensions for the mesh
as three parameters to the executable. For example to execute the provided
code on a small 10x10x10 mesh you would enter‘./acacgs 10 10 10’. On my
system the output for the code is below. This information is also stored in a
file, which is named after the wallclock date and time of when the program was
first executed (for example, 2022 01 26 12 00 00.txt).
===== Final Statistics =====
Executable name: ./acacgs
Dimensions: 10 10 10
Number of iterations: 149
Final residual: 2.226719e-92
=== Time ==
Total: 1.126600e-02 seconds
ddot Kernel: 8.390000e-04 seconds
waxpby Kernel: 1.087000e-03 seconds
sparsemv Kernel: 9.123000e-03 seconds
=== FLOP ==
Total: 9.536000e+06 floating point operations
ddot Kernel: 5.960000e+05 floating point operations
waxpby Kernel: 8.940000e+05 floating point operations
sparsemv Kernel: 8.046000e+06 floating point operations
=== MFLOP/s ==
Total: 8.464406e+02 MFLOP/s
ddot Kernel: 7.103695e+02 MFLOP/s
waxpby Kernel: 8.224471e+02 MFLOP/s
sparsemv Kernel: 8.819467e+02 MFLOP/s
Difference between computed and exact = 1.110223e-15
You will find more detailed instructions to build the code in the README.md
file, including flags to turn on verbose mode, which will output details for each
timestep in the simulation, and flags for enabling visualisation.
3
4.1 Visualisation Generation
To enable visualisation outputs, you must build your code using‘make SILO=1’.
This will then compile your code in a way which produces files suitable for
visualisation in VisIt. If you are working remotely and want to visualise the
coursework, it will be quicker and easier for you to copy the files to your local
machine, then utilise VisIt on the local machine to visualise the cuboid. Before
you make the program, make sure you load the SILO module (module load
cs257-silo).
When the program is ran with visualisations, each timestep will produce a
SILO file within a directory named after the wallclock date and time (for example: - 01 26 12 00 00). In this directory will be a collection of .silo files,
each named outputXXXX.silo, where XXXX represents the timestep it relates to.
Once the program has finished, these can be utilised in Visit. To do so,
load the VisIt module (module load cs257-visit) and open VisIt using the
command visit. From here, you will get 2 windows. The smaller, skinner one
is the control window and is used to manage everything that will be displayed.
The larger window is the display window. In the control window, select Open,
and navigate to the directory with the SILO files. You should then be able to
select these SILO files.
Now that the SILO files have been loaded, we can now draw some given
variables. To do this, click on the Add and select a mode and a variable that
should be viewed. One of the nicest ones to use is Volume and either x nodal
or p nodal. When you have finished adding elements, click on Draw. This will
generate an image in the display window, that can be dragged around so that
the cuboid can be viewed from different angles. The control window has a play
button, which will run through each timestep.
Visualisations are nice to have, but for performance purposes we turn them off
as they write a significant amount fo data to disk.
Table 1: Visualisation Data File Sizes
x y z Cells Approximate Data Size - 10 10 1000 4MB
- 25 25 15,625 39MB
- 50 50 125,000 301MB
- 100 100 1,000,000 2.4GB
- 200 200 8,000,000 19.3GB
There is the potential to go significantly over your DCS disk quota with
large meshes. I recommend that you do not exceed 30x30x30 for producing
visualisations on the DCS machines. If you are developing your solution on
your personal machine then you may wish to produce larger visualisations.
4 - Hardware Details
On a linux system, you can read the processor information using the command
“cat /proc/cpuinfo”. This will provide full details on the cpu in the machine,
including the CPU model, number of cores, the clock frequency and supported
extensions. I strongly recommend taking a look at this on your development
machine.
For the purposes of assessment, your code will be run on a DCS machine
with 6 cores. The output from /proc/cpuinfo for a single core can be seen below:
processor : 5
vendor_id : GenuineIntel
cpu family : 6
model : 158
model name : Intel(R) Core(TM) i5-8500 CPU @ 3.00GHz
stepping : 10
microcode : 0xca
cpu MHz : 799.987
cache size : 9216 KB
physical id : 0
siblings : 6
core id : 5
cpu cores : 6
apicid : 10
initial apicid : 10
fpu : yes
fpu_exception : yes
cpuid level : 22
wp : yes
flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov
pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb
rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology
nonstop_tsc aperfmperf eagerfpu pni pclmulqdq dtes64 monitor ds_cpl vmx
smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe
popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm
3dnowprefetch invpcid_single intel_pt ssbd ibrs ibpb stibp tpr_shadow vnmi
flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid
rtm mpx rdseed adx smap clflushopt xsaveopt xsavec xgetbv1 dtherm ida arat
pln pts hwp hwp_notify hwp_act_window hwp_epp md_clear spec_ctrl
intel_stibp flush_l1d
bogomips : 6000.00
clflush size : 64
cache_alignment : 64
address sizes : 39 bits physical, 48 bits virtual
power management:
Machines matching this specification are available in the cs257 queue of the
5
Batch Compute System in the Department. You will learn how to use this
system during the lab sessions, so there will be time to get used to it. - How will my code be tested for performance?
Your submission will be tested on a range of input sizes to evaluate how robust
your performance improvements are. It is recommended that you try testing
your solution on inputs that are not cubes to see if there are any weaknesses in
your optimisation strategies. The 7-pt stencil option will not be used for testing
your code.
Your code will be executed five times for each problem size on the target
hardware. The highest and lowest runtimes will be discarded, and the mean of
the three remaining values will be taken as your runtime for that problem size. - Rules
Your submitted solution must:
• Compile on the DCS workstations.
Your submitted solution must not:
• Use instruction sets not supported by the DCS machines.
• Require additional hardware e.g., GPUs
• Add relaxed math options to the compile line, e.g., -ffast-math. Note:
Manual use of approximate math functions is acceptable. - Where do I start?
This can seem like a daunting project, but we can break it down into a number
of steps. - Compile and run the code as provided. This is a quick easy check to make
sure your environment is setup correctly. - Read the code. Start in main.c and follow it through. The functions
are well documented with Doxygen comments. Don’t panic – you are not
expected to understand the physics in the code. - Measure the runtime of the code for reference purposes.
- Figure our where the most intensive sections of code are.
- Develop a small optimisation.
- Run the code and review the impact of your changes.
- Repeat steps 5 and 6 until you have exhausted your performance ideas.
6 - Instructions for Submission
Your solution should be submitted using Tabula. Please ensure that your code
works on DCS machines prior to submission.
Submission deadline: Noon, Wednesday 16th March 2022.
Files required: A single file named coursework.zip which should contain
all of your code at the top-level (i.e. no subdirectories) and the report file as a
PDF.