$29
Project 1 (Part 2): IaaS [CSE 546] Cloud Computing
Summary
In Part 2 of Project 1, we will complete the development of the elastic face recognition application using the IaaS resources from AWS. We will implement the App Tier and Data Tier of our multi-tiered cloud application, use a machine learning model to perform face recognition and implement autoscaling to allow the App Tier to dynamically scale on demand.
Description
We will strictly follow the architecture below to develop the application.
Web Tier
1. You MUST use the same web tier EC2 instance from Part 1 of the project. The web tier Must listen for requests on port number 8000.
2. It should take images received from users as input and forward it to the App Tier for model inference. It should also return the recognition result from the App Tier as output to the users. The input from each request is a .jpg file, and the output in each response is the recognition result. Find more details below.
Input:
◦ The key to the HTTP payload MUST be defined as “inputFile” and should be used as the same. In the case of a Python backend, it denotes a standard Python file object.
Copyright © Ming Zhao, PhD
• For example, the user uploads an image named “test_00.jpg”.
• Use the provided workload generator to generate requests to your Web Tier.
Output:
• The Web Tier will handle HTTP POST requests to the root endpoint ("/").
• The output MUST be in plain text, and the format <filename>:<classification_results>
• For the above example request, the output should be “test_00:Paul” in plain text.
• You need to implement the handling of concurrent requests in your Web Tier.
To facilitate the testing, a standard face dataset and the expected recognition output of each image are provided to you at
Input: visa-lab/CSE546-Cloud-Computing/face_images_1000.zip
Output: visa-lab/CSE546-Cloud-Computing/classification_face_images_1000.csv
3. The Web Tier uses only one instance. You MUST name your web-tier instance “web-instance”
4. The Web Tier sends requests to the App Tier and receives results using two SQS Queues. You MUST follow the following naming convention in naming your S3 Buckets:
a. Request Queue: <ASU ID>-req-queue
b. Response Queue: <ASU ID>-resp-queue
For example, if your ASU ID is “12345678910”, your queue names will be 12345678910-req-queue and 12345678910-resp-queue
5. The Web Tier also runs the autoscaling controller, which determines how to scale the App Tier. You are not allowed to use AWS features for autoscaling. You MUST implement your own auto scaling algorithm.
App Tier
1. The App Tier will use the provided deep learning model for model inference. The deep learning model code is available here. The first step is to create an AMI, which will be used to launch App Tier instances.
a. Launch a base EC2 instance. You can use the AWS Linux or Ubuntu AMI.
b. Install the required package on the instance using the following command.
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
c. Copy the provided deep learning model code and model weights folder to the EC2 instance using scp.
d. Refer to the README.md on how to use the deep learning model code.
Copyright © Ming Zhao, PhD
e. Create an AMI using this EC2 instance, following the instructions here. Now, you can use this AMI to create your App Tier instances. You MUST name your app-tier instances “app-tier-instance-<instance#>”
2. The App Tier should automatically scale out when the request demand increases and automatically scale in when the demand drops. The number of App Tier instances should be 0 when there are no requests being processed or waiting to be processed. The number of App Tier instances can scale to at most 20 because we have limited resources from the free tier.
Data Tier
1. All the inputs (images) and outputs (recognition results) should be stored in separate buckets on S3 for persistence.
2. S3 stores all the objects as key-value pairs. For the Input bucket, each object's key is the input file's name, e.g., test_00.jpg, and the value is the image file. For the Output bucket, the key is the image name (e.g., test_00), and the value is the classification result (e.g., “Paul”).
3. You MUST follow the following naming convention in naming your S3 Buckets
a. Input Bucket: <ASU ID>-in-bucket
b. Output Bucket: <ASU ID>-out-bucket
For example, if your ASU ID is “12345678910”, your bucket names will be 12345678910-in-bucket and 12345678910-out-bucket
Testing & Grading
• Use the provided workload generator to test your app thoroughly. Check the following:
◦ The output of the workload generator is correct.
◦ The contents in the S3 buckets are correct;
◦ While processing the workload, the number of EC2 instances is correct.
◦ All the requests are processed within a reasonable amount of time. As a reference point, for a workload of 100 concurrent requests using the TAs’ implementation of the project, completed within 80 seconds (average of 5 iterations).
• Test your application using the provided workload generator and grading script. If they fail to execute, you will receive 0 points. Refer to the Readme on how to use the grading script.
Copyright © Ming Zhao, PhD
• We will use the same Grading IAM user credentials we collected in Part 1 of the project. You MUST update the Grading IAM user with the following permissions:
◦ AmazonEC2ReadOnlyAccess
◦ AmazonS3FullAccess
◦ AmazonSQSFullAccess
• Make sure you test everything using the Grading IAM Credentials before making a submission.
• To facilitate testing, you MUST use only the resources from the US-East-1 region.
• The grading will be done using the workload generator and grading script and following the rubrics below:
#
Test Objective
Test Criteria
Test Command
Fail
Pass
Total
Sequence
Points
1) Single Web Tier instance
Run p2_grader.py &
There is no EC2
The EC2 instance with the name
instance with name
exists and the state of the
use selection
"web-instance" exists; and is in "running"
1
Validate EC2
web-instance is "running"
choice:1 for
"web-instance" (0)
state (2.5)
5
initial state
2) There are 0 running App
validating the
There are app tier
There are no app tier instance in
instances in
Tier instances with name
instances
"running" state (2.5)
"running" state (0)
The input and the output
Run p2_grader.py
The input and the
Validate S3
script & use selection
output buckets either
The input and the output buckets both
2
buckets must exist and
2.5
initial state
choice:2 for
do not exists or are
exists and are empty (2.5)
should be empty
validating S3 Buckets
not empty(0)
The SQS Request and
Run p2_grader.py
The request and the
Validate SQS
script & use selection
The request and the response queue
3
Response Queues must
choice:3 for
response queues
2.5
initial state
exist and have no
either do not exists
both exist and are empty (2.5)
validating SQS
messages in the queue
or are not empty(0)
Queues
1) Start the
p2_grader.py before
the workload
generator and use
Validate the
selection choice: 4
Total number of
1) Run
Total number of requests completed
completeness
requests completed
The HTTP Response code
workload_generator.
successfully stats from workload
4
of the
must be 200 for all the
py with 10 requests
successfully stats
generator is 10 (5)
5
10-request
from workload
requests
2) After the workload
+0.5 for every request completed
workload
generator is not 10
generator is over,
successfully.
(0)
check the stats "Total
number of requests
completed
successfully"
Copyright © Ming Zhao, PhD
Validate the
After the workload
Total number of
The Total number of correct predictions
correctness of
generator is over,
correct predictions
All the classification
stats from workload generator.py is 10
5
the
results must be correct
check the stats "Total
stats from
(5)
classification
number of correct
workload_generator.
results
predictions"
py is not 10 (0)
+0.5 for every correct classification.
The S3 input and Output
Run p2_grader.py &
The S3 Input and
Validate S3
use selection
Output Buckets do
The S3 Input and Output Buckets have 10
Buckets should each have
5
buckets
choice:2 for
not have 10 objects.
objects (5)
10 objects
validating S3 Buckets
(5)
The # of app tier
1) The # of app tier
instances do not
gradually scale from
instances should gradually
The # of app tier instances gradually
Validate the prints
0 to 10 and reduce
scale from 0 to 10 and
from p2_grader.py
back to 0 (0)
scale from 0 to 10 and reduce back to 0
Validate
reduce back to 0
(10)
when the selection
The # of SQS
15
autoscaling
2) The # of SQS messages
The # of SQS messages in Request Queue
choice: 4 was made
messages in the
in Request Queue should
from Step-1
Request Queue do
gradually increases from 0 to 10 and
increase from 0 and then
reduce back to 0 (5)
not gradually
reduce back to 0
increase from 0 and
reduce back to 0 (0)
The Total
The Total
test
Check the "Total Test
test duration
The Total test
The total test
duration
Duration" stats from
for 10
duration for 10
Validate total
Total end to end latency
duration for 10
for 10
the
requests is
requests is
15
latency
must be reasonable
requests is greater
requests
workload_generator.
between 3
between 2 and
than 4 min (0)
is less
py
and 4 min
3 min (10)
than 2
(5)
min (15)
1) Start the
p2_grader.py before
the workload
generator and use
Validate the
selection choice: 4
1) Run
Total number of
Total number of requests completed
completeness
The HTTP Response code
workload_generator.
requests completed
successfully stats from workload
of the
must be 200 for all the
5
5
py with 50 requests
successfully stats
generator is 50 (5)
50-request
requests
2) After the workload
from workload
+0.1 for every request completed
workload
generator is over,
generator is not 50(0)
successfully.
check the stats "Total
number of requests
completed
successfully"
Validate the
All the classification
After the workload
The Total number of
Total number of correct predictions stats
5
Copyright © Ming Zhao, PhD
correctness of
results must be correct
generator is over,
correct predictions
from workload generator is 50 (5)
the
check the stats "Total
stats from workload
+0.1 for every correct classification.
classification
number of correct
generator is not 50
result
predictions"
(0)
The S3 input and Output
Run p2_grader.py &
The S3 Input and
Validate S3
use selection
Output Buckets do
The S3 Input and Output Buckets have 50
Buckets should each have
5
buckets
choice:2 for
not have 50 objects.
objects (5)
10 objects
validating S3 Buckets
(5)
The # of app tier
1) The # of app tier
instances do not
gradually scale from
instances should gradually
The # of app tier instances gradually
Validate the prints
0 to 20 and reduce
scale from 0 to 20 and
from p2_grader.py
back to 0 (0)
scale from 0 to 20 and reduce back to 0
Validate
reduce back to 0
(10)
when the selection
The # of SQS
15
autoscaling
2) The # of SQS messages
The # of SQS messages in Request Queue
choice: 4 was made
messages in Request
in Request Queue should
from Step-1
Queue do not
gradually increase from 0 to 20 and
increase from 0 and then
reduce back to 0 (5)
gradually increase
reduce back to 0
from 0 and then
reduce back to 0 (0)
Total test
Total test
Check the "Total Test
Total test
duration
duration for
Duration" stats from
Total test duration for
duration for 50
for 50
Validate total
Total end to end latency
the
50 requests is greater
50 requests
requests is
requests
15
latency
must be reasonable
workload_generator.
than 5 min (0)
is between 4
between 3 and
is less
py
and 5 min
4 min (10)
than 3
(5)
min (15)
Submission
The submission requires four components; all must be completed by 03/15/2024 at 11:59:59
pm.
1. Upload your source code to Canvas as a single zip file named by your full name: <lastname><firstname>.zip. Submit only your source code. Do not include any code not developed by you. Do not include any binary files.
2. Share your web tier Elastic IP and AWS Grading Credentials using the Google Form
3. Keep your web tier instance running until you are informed that grading is done.
4. Before you submit, make sure you empty the Input and Output Buckets.
Failure to follow the above instructions to prepare your submission will cause a penalty to your grade.
Copyright © Ming Zhao, PhD
Policies
1) Late submissions will absolutely not be graded (unless you have verifiable proof of emergency). It is much better to submit partial work on time and get partial credit than to submit late for no credit.
2) You need to work independently on this project. We encourage high-level group discussions to help others understand the concepts and principles. However, code-level discussion is prohibited, and plagiarism will directly lead to failure in this course. We will use anti-plagiarism tools to detect violations of this policy.
