ubc-cirrus-lab/unfaasener

A lightweight framework that enables serverless users to reduce their bills by harvesting non-serverless compute resources such as their VMs, on-premise servers, or personal computers.

UBC CIRRUS Lab Group Website

The website template is from https://html5up.net/.

The news preparation script is from https://github.com/mshahrad/mshahrad.github.io

unfaasener's Issues

Adding sample videos for VideoAnalytics benchmark

• In a sub directory, with links to the source.
• 5 samples
• ideally different features (length, quality, frame variability, etc.)

Error in vmModule (when having invocation rate of greater than 1)

Traceback (most recent call last):

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/api/client.py", line 268, in _raise_for_status

response.raise_for_status()

File "/home/ghazal/.local/lib/python3.10/site-packages/requests/models.py", line 1021, in raise_for_status

raise HTTPError(http_error_msg, response=self)

requests.exceptions.HTTPError: 500 Server Error: Internal Server Error for url: http+docker://localhost/v1.41/exec/31e0cc791ad6c49458f8186917ffc52c0f21ac104a1e1d3053a431b11742ee28/start

The above exception was the direct cause of the following exception:

Traceback (most recent call last):

File "/home/ghazal/de-serverlessization/vm-agent/execution-agent/vmModule.py", line 263, in

streaming_pull_future.result()

File "/usr/lib/python3.10/concurrent/futures/_base.py", line 439, in result

return self.__get_result()

File "/usr/lib/python3.10/concurrent/futures/_base.py", line 391, in __get_result

raise self._exception

File "/home/ghazal/de-serverlessization/vm-agent/execution-agent/vmModule.py", line 259, in

streaming_pull_future.result()

File "/usr/lib/python3.10/concurrent/futures/_base.py", line 446, in result

return self.__get_result()

File "/usr/lib/python3.10/concurrent/futures/_base.py", line 391, in __get_result

raise self._exception

File "/home/ghazal/.local/lib/python3.10/site-packages/google/cloud/pubsub_v1/subscriber/_protocol/streaming_pull_manager.py", line 126, in _wrap_callback_errors

callback(message)

File "/home/ghazal/de-serverlessization/vm-agent/execution-agent/vmModule.py", line 205, in callback

cont.exec_run("python3 /app/main.py '"+  str(jsonfile).replace('\'','"') + "' " + reqID,detach=False )

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/models/containers.py", line 198, in exec_run

exec_output = self.client.api.exec_start(

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/utils/decorators.py", line 19, in wrapped

return f(self, resource_id, *args, **kwargs)

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/api/exec_api.py", line 167, in exec_start

return self._read_from_socket(res, stream, tty=tty, demux=demux)

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/api/client.py", line 409, in _read_from_socket

socket = self._get_raw_response_socket(response)

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/api/client.py", line 318, in _get_raw_response_socket

self._raise_for_status(response)

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/api/client.py", line 270, in _raise_for_status

raise create_api_error_from_http_exception(e) from e

File "/home/ghazal/.local/lib/python3.10/site-packages/docker/errors.py", line 39, in create_api_error_from_http_exception

raise cls(e, response=response, explanation=explanation) from e

docker.errors.APIError: 500 Server Error for http+docker://localhost/v1.41/exec/31e0cc791ad6c49458f8186917ffc52c0f21ac104a1e1d3053a431b11742ee28/start: Internal Server Error ("Container 8957a6657d56082b51d2c51968472492824111eb48d644ddd00ff157e9137d19 is not running: Exited (137) Less than a second ago")

Automation of pub/sub channel creation for VMs

When testing with the second VM @mfouadga2 noticed that it's not straightforward to add new pub/sub for VMs. We need a methodology for this.

retrieving execution times from the VMs

upgrading the vm resource predictor

Based on the simulation results, let's go for the Exponential Moving Average policy. Can be found here: https://github.com/ubc-cirrus-lab/de-serverlessization-simuls/blob/main/vm-resource-prediction/Predictor.py

Set memory allocations of functions with AWS power tuning

VM Agent instead of CPU Agent

@mfouadga2 should we call it VM agent instead of CPU agent? (https://github.com/ubc-cirrus-lab/de-serverlessization/tree/main/cpu-agent). After all, the agent should be able to profile and make predictions for CPU as well as memory.

VM logs are partially inserted into datastore (vmModule)

considering invocation rates for offloading decisions

Fix the name of the merging point function in serverless

The merging function is referenced twice in the code
once:
Text2SpeechCensoringWorkflow_MergedFunction
and the other time
Text2SpeechCensoringWorkflow_MergingPoint

Test the dependence of regressiontuning on sample count (n)

containerizing functions on the VMs

Enforcing CPU limits for execution containers

• see if this is viable
• similar to this: https://cloud.google.com/functions/docs/configuring/memory

Adapting the offloading solutions to invocation rates in real time

Either by the Header function or ...

Exploring conservative resource predictions

Maybe start with a histogram based approach for a moving window of time

Excluding offloaded functions resource usage from monitoring/prediction

From the solver's viewpoint, available resources on VMs includes anything that is not used by host's original workload. In this design, 1) the resources used by offloaded serverless functions should not be counted by the monitoring/prediction agent, and 2) offloaded serverless functions should be capped to not use more resources than what is predicted to be available. I think our current implementation violates this model. Creating this issue for us to think about this more carefully and to fix it.

restructuring the vm agent

• monitoring
• invoker
• reporter

implementing RegressionTuning for Google Cloud Workflows

Exploring docker SDK

Exploring whether docker SDK would be faster/more efficient that the syscalls

Consolidating installation of python dependencies of vm agent into the original requirements file

Just noticed that in the vm agent makefile there are pip install commands. Currently, we have a setup script which takes care of installing all dependencies listed in the requirements file. I suggest consolidating all requirement installations there for easier management.

Allowing for concurrency > 1 for the same function

right now maximum one instance of each function is allowed to run. we need to remove this limit.
this will probably fix #41 too

Porting the benchmarks to Google Cloud Workflows

@mohannashahrad has volunteered to port our benchmarks to Google Cloud Workflows. She should have access to the necessary cloud resources.

@mohannashahrad please use this issue to provide updates on this front in the coming weeks. This is not on our critical path and you can take your time. If you run into problems about the benchmarks you can let myself or @GhazalSdn know.

Thanks for your help.

Build a make script for all parts

System modifications after arrival rate incorporation

• Updating the header function - @GhazalSdn
• Updating the scheduler - @GhazalSdn
• Updating the agent to 1) implement the new trigger (serverless offloads going beyond dedicated capacity) and 2) pass the trigger type - @mfouadga2

Porting the Montage benchmark

Mentioned in #13

Modify the serverless logging parser

• Include the last n observations
• Store the observations in a proper datastore
• Keep the timestamp for the last observation, and be able to append new logs to the previously stored logs

Add more sample images to the ImageProcessing benchmark

And include them in a sub-directory under the benchmark

adding GitHub testing workflows

I'll take care of it in the next few days

adding the notion extra latency tolerance to the solver

basic keep-alive for VM containers

Removing previous serverless records in the dataframe

This task should be done asynchronously, in order to remove old records based on their timestamps.

Looking for a new benchmark

Parametrizing the fanout of video analytics

Instead of hardcoded fan out of 10

measuring exec times on VM-hosted functions

Docker utilization reading issue

Seems that the process ID used in dockerprocstatparser is hard coded. Please look into it and automate it, as reading wrong docker util data messes up the entire VM prediction/execution flow.

Exec time buffer

A buffer needed to keep the record of recent execution times on the VM side before they are submitted. The submission will be taken care of by #11

Readme for Benchmarks

We need to add a README.md file under benchmarks. This README should include a table with each benchmark application in a row and should have the following columns:

• Application name
• Brief explanation of the application
• A diagram showing the function dependency graph (workflow)
• Language of the application
• Link to the original source (if this is an application developed by others)

Error in the middle of test for vmModule

At some point after executing the tests, I got this error for the vmModule, but after that it was working.

Exception in thread Thread-2 (threaded_function):

Traceback (most recent call last):

File "/usr/lib/python3.10/threading.py", line 1009, in _bootstrap_inner

self.run()

File "/usr/lib/python3.10/threading.py", line 946, in run

self._target(*self._args, **self._kwargs)

File "/home/ghazal/de-serverlessization/vm-agent/execution-agent/vmModule.py", line 81, in threaded_function

for key in lastexectimestamps:

RuntimeError: dictionary changed size during iteration

Porting benchmarks to AWS Step Function

Check out Illegal Recognizer (IR) benchmark

Making the merging nodes explicit in the DAG

We decided to make the merging nodes explicit. @GhazalSdn has already started working on this. Creating an issue to keep track of it until completion.

Removing system log files from the repo

Seems like system logs are committed to the repo. I saw a few examples such as https://github.com/ubc-cirrus-lab/de-serverlessization/blob/main/log_parser/get_workflow_logs/logs/logParser.log and https://github.com/ubc-cirrus-lab/de-serverlessization/blob/main/scheduler/logs/scheduler.log

I suggest adding *.log to the out gitignore file.

Multi-VM support

• Reporting available resources (instead of writing to local file)
• Leader election for VM hosting the scheduler
• Triggering the solver (scheduler) on remote VM

Extending the solver to support multi-VM

Supporting multiple workflows from the same users

taking care of the serverless glue part

Reusing existing containers

So that we don't go the long path (json -> image -> create -> run), every single time.

Remove unnecessary pub/sub fields in routing epilogue

Fields such as publishTime, etc. are not needed for pub/sub communication. Please remove any unnecessary field to ensure low overhead for the routing epilogue.

Consolidating test scripts and data

Adding a tests directory to contain all the test scripts and data.

worst-case, best-case, and median offloading recommendations