This software simulates the execution of a FaaS system deployed over Edge and Cloud nodes.

Requirements

The simulator is written in Python. You can run it within a virtual environment installing the dependencies listed in requirements.txt. E.g.:

python -m venv env
source env/bin/activate
pip install -r requirements.txt 

You may also need to install the GLPK solver.

Usage

The simulator is started as follows:

python main.py [<config file>]

If no configuration file is specified, ./config.ini is used.

Configuration

Two key files are responsible for the configuration:

• the main .ini configuration file
• a YAML specification file of available QoS classes, functions, nodes and workload

Note that the main configuration file contains a reference to the associated specification file (spec-file option, as shown below).

Example .ini configuration:

[simulation]
stats-print-interval = 1200
stats-print-file = stats.txt
close-door-time = 1000
plot-rt = false
seed = 123
spec-file = spec.yml

[policy]
name = basic
update-interval = 120
arrival-alpha = 0.3

[container]
init-time = 1.1

An example spec.yml file is provided as spec.yml.example in this repository.

Arrival traces

Each function is associated with an arrival process. Incoming requests are randomly associated with a QoS class based on the arrival-weight attribute of each class (i.e., the higher the weight, the higher the probability of belonging to the class).

You can either specify:

• a constant average rate for the function: arrivals will follow a Poisson process.
• an inter-arrival times (IATs) trace file, using the trace attribute. The trace is a text file containing IATs, one per line.

Policies

The following policies are currently implemented:

• basic: heuristic policy that schedules all the requests for local execution when possible
• greedy: greedy heuristic that schedules requests so as to minimize the expected response time
• probabilistic: probabilistic heuristic, whose probabilities are periodically updated solving a LP problem
• random

Configuring the Probabilistic policy

You may want to adjust the following parameters when using the probabilistic policy:

• update-interval: how often (in seconds) the probabilities should be updated via optimization
• arrival-alpha: the arrival rate of each function-class pair is tracked using an Exponential Moving Average, whose 0 < alpha <= 1 parameter is set by this configuration option (e.g., a value closer to 1.0 means we weigh recent observations as more important).

Utilities

Specific scripts can be used to run single experiments, e.g.:

python experiment_policies.py

Stateful

Each node stores key-value pairs. The kv_store field of the FaaS Node associates each key with the size (in bytes) of the corresponding value.

The KeyLocator class and the key_locator object defined in stateful.py provide each node the ability to discover which node stores a certain key.

You can specify the list of keys accessed by a function in the spec file. Each key can be associated with a probability (default 1.0) of being accessed at each function invocation. Example:

functions:
  - name: f1
    memory: 1024
    duration_mean: 0.240
    duration_scv: 1.0
    keys:
      - key: a
        probability: 0.3
      - key: b

Credits

The project includes the BuTools package. Original source