Run large-scale simulations with AWS Batch multi-container jobs


Voiced by Polly

Industries like automotive, robotics, and finance are more and more implementing computational workloads like simulations, machine studying (ML) mannequin coaching, and large knowledge analytics to enhance their merchandise. For instance, automakers depend on simulations to check autonomous driving options, robotics firms practice ML algorithms to boost robotic notion capabilities, and monetary companies run in-depth analyses to raised handle danger, course of transactions, and detect fraud.

Some of those workloads, together with simulations, are particularly difficult to run because of their variety of parts and intensive computational necessities. A driving simulation, as an example, entails producing 3D digital environments, car sensor knowledge, car dynamics controlling automotive conduct, and extra. A robotics simulation would possibly take a look at lots of of autonomous supply robots interacting with one another and different techniques in a large warehouse surroundings.

AWS Batch is a totally managed service that may provide help to run batch workloads throughout a spread of AWS compute choices, together with Amazon Elastic Container Service (Amazon ECS), Amazon Elastic Kubernetes Service (Amazon EKS), AWS Fargate, and Amazon EC2 Spot or On-Demand Instances. Traditionally, AWS Batch solely allowed single-container jobs and required additional steps to merge all parts right into a monolithic container. It additionally didn’t enable utilizing separate “sidecar” containers, that are auxiliary containers that complement the principle utility by offering further companies like knowledge logging. This further effort required coordination throughout a number of groups, corresponding to software program growth, IT operations, and high quality assurance (QA), as a result of any code change meant rebuilding the whole container.

Now, AWS Batch gives multi-container jobs, making it simpler and quicker to run large-scale simulations in areas like autonomous autos and robotics. These workloads are often divided between the simulation itself and the system underneath take a look at (often known as an agent) that interacts with the simulation. These two parts are sometimes developed and optimized by totally different groups. With the flexibility to run a number of containers per job, you get the superior scaling, scheduling, and value optimization supplied by AWS Batch, and you should utilize modular containers representing totally different parts like 3D environments, robotic sensors, or monitoring sidecars. In reality, prospects corresponding to IPG Automotive, MORAI, and are already utilizing AWS Batch multi-container jobs to run their simulation software program within the cloud.

Let’s see how this works in apply utilizing a simplified instance and have some enjoyable making an attempt to unravel a maze.

Building a Simulation Running on Containers
In manufacturing, you’ll in all probability use current simulation software program. For this submit, I constructed a simplified model of an agent/mannequin simulation. If you’re not serious about code particulars, you possibly can skip this part and go straight to the best way to configure AWS Batch.

For this simulation, the world to discover is a randomly generated 2D maze. The agent has the duty to discover the maze to discover a key after which attain the exit. In a manner, it’s a basic instance of pathfinding issues with three places.

Here’s a pattern map of a maze the place I highlighted the beginning (S), finish (E), and key (Okay) places.

Sample ASCII maze map.

The separation of agent and mannequin into two separate containers permits totally different groups to work on every of them individually. Each workforce can give attention to enhancing their very own half, for instance, so as to add particulars to the simulation or to seek out higher methods for a way the agent explores the maze.

Here’s the code of the maze mannequin ( I used Python for each examples. The mannequin exposes a REST API that the agent can use to maneuver across the maze and know if it has discovered the important thing and reached the exit. The maze mannequin makes use of Flask for the REST API.

import json
import random
from flask import Flask, request, Response

prepared = False

# How map knowledge is saved inside a maze
# with dimension (width x top) = (4 x 3)
#    012345678
# 0: +-+-+ +-+
# 1: | |   | |
# 2: +-+ +-+-+
# 3: | |   | |
# 4: +-+-+ +-+
# 5: | | | | |
# 6: +-+-+-+-+
# 7: Not used

class FlawedDirection(Exception):

class Maze:
    UP, RIGHT, DOWN, LEFT = 0, 1, 2, 3
    OPEN, WALL = 0, 1

    def distance(p1, p2):
        (x1, y1) = p1
        (x2, y2) = p2
        return abs(y2-y1) + abs(x2-x1)

    def random_dir():
        return random.randrange(4)

    def go_dir(x, y, d):
        if d == Maze.UP:
            return (x, y - 1)
        elif d == Maze.RIGHT:
            return (x + 1, y)
        elif d == Maze.DOWN:
            return (x, y + 1)
        elif d == Maze.LEFT:
            return (x - 1, y)
            elevate FlawedDirection(f"Direction: {d}")

    def __init__(self, width, top):
        self.width = width = top        

    def space(self):
        return self.width *

    def min_lenght(self):
        return / 5

    def min_distance(self):
        return (self.width + / 5

    def get_pos_dir(self, x, y, d):
        if d == Maze.UP:
            return self.maze[y][2 * x + 1]
        elif d == Maze.RIGHT:
            return self.maze[y][2 * x + 2]
        elif d == Maze.DOWN:
            return self.maze[y + 1][2 * x + 1]
        elif d ==  Maze.LEFT:
            return self.maze[y][2 * x]
            elevate FlawedDirection(f"Direction: {d}")

    def set_pos_dir(self, x, y, d, v):
        if d == Maze.UP:
            self.maze[y][2 * x + 1] = v
        elif d == Maze.RIGHT:
            self.maze[y][2 * x + 2] = v
        elif d == Maze.DOWN:
            self.maze[y + 1][2 * x + 1] = v
        elif d ==  Maze.LEFT:
            self.maze[y][2 * x] = v
            FlawedDirection(f"Direction: {d}  Value: {v}")

    def is_inside(self, x, y):
        return 0 <= y < and 0 <= x < self.width

    def generate(self):
        self.maze = []
        # Close all borders
        for y in vary(0, + 1):
            self.maze.append([Maze.WALL] * (2 * self.width + 1))
        # Get a random place to begin on one of many borders
        if random.random() < 0.5:
            sx = random.randrange(self.width)
            if random.random() < 0.5:
                sy = 0
                self.set_pos_dir(sx, sy, Maze.UP, Maze.OPEN)
                sy = - 1
                self.set_pos_dir(sx, sy, Maze.DOWN, Maze.OPEN)
            sy = random.randrange(
            if random.random() < 0.5:
                sx = 0
                self.set_pos_dir(sx, sy, Maze.LEFT, Maze.OPEN)
                sx = self.width - 1
                self.set_pos_dir(sx, sy, Maze.RIGHT, Maze.OPEN)
        self.begin = (sx, sy)
        been = [self.start]
        pos = -1
        solved = False
        generate_status = 0
        old_generate_status = 0                    
        whereas len(been) <
            (x, y) = been[pos]
            sd = Maze.random_dir()
            for nd in vary(4):
                d = (sd + nd) % 4
                if self.get_pos_dir(x, y, d) != Maze.WALL:
                (nx, ny) = Maze.go_dir(x, y, d)
                if (nx, ny) in been:
                if self.is_inside(nx, ny):
                    self.set_pos_dir(x, y, d, Maze.OPEN)
                    been.append((nx, ny))
                    pos = -1
                    generate_status = len(been) /
                    if generate_status - old_generate_status > 0.1:
                        old_generate_status = generate_status
                        print(f"{generate_status * 100:.2f}%")
                elif solved or len(been) < self.min_lenght():
                    self.set_pos_dir(x, y, d, Maze.OPEN)
                    self.finish = (x, y)
                    solved = True
                    pos = -1 - random.randrange(len(been))
                pos -= 1
                if pos < -len(been):
                    pos = -1
        self.key = None
        whereas(self.key == None):
            kx = random.randrange(self.width)
            ky = random.randrange(
            if (Maze.distance(self.begin, (kx,ky)) > self.min_distance()
                and Maze.distance(self.finish, (kx,ky)) > self.min_distance()):
                self.key = (kx, ky)

    def get_label(self, x, y):
        if (x, y) == self.begin:
        elif (x, y) == self.finish:
        elif (x, y) == self.key:
            c=" "
        return c

    def map(self, strikes=[]):
        map = ''
        for py in vary( * 2 + 1):
            row = ''
            for px in vary(self.width * 2 + 1):
                x = int(px / 2)
                y = int(py / 2)
                if py % 2 == 0: #Even rows
                    if px % 2 == 0:
                        v = self.get_pos_dir(x, y, self.UP)
                        if v == Maze.OPEN:
                            c=" "
                        elif v == Maze.WALL:
                else: # Odd rows
                    if px % 2 == 0:
                        v = self.get_pos_dir(x, y, self.LEFT)
                        if v == Maze.OPEN:
                            c=" "
                        elif v == Maze.WALL:
                        c = self.get_label(x, y)
                        if c == ' ' and [x, y] in strikes:
                row += c
            map += row + 'n'
        return map

app = Flask(__name__)

def hello_maze():
    return "<p>Hello, Maze!</p>"

@app.route('/maze/map', strategies=['GET', 'POST'])
def maze_map():
    if not prepared:
        return Response(standing=503, retry_after=10)
    if request.technique == 'GET':
        return '<pre>' + + '</pre>'
        strikes = request.get_json()

def maze_start():
    if not prepared:
        return Response(standing=503, retry_after=10)
    begin = { 'x': maze.begin[0], 'y': maze.begin[1] }
    return json.dumps(begin)

def maze_size():
    if not prepared:
        return Response(standing=503, retry_after=10)
    dimension = { 'width': maze.width, 'top': }
    return json.dumps(dimension)

def maze_pos(y, x):
    if not prepared:
        return Response(standing=503, retry_after=10)
    pos = {
        'right here': maze.get_label(x, y),
        'up': maze.get_pos_dir(x, y, Maze.UP),
        'down': maze.get_pos_dir(x, y, Maze.DOWN),
        'left': maze.get_pos_dir(x, y, Maze.LEFT),
        'proper': maze.get_pos_dir(x, y, Maze.RIGHT),

    return json.dumps(pos)

WIDTH = 80
maze = Maze(WIDTH, HEIGHT)
prepared = True

The solely requirement for the maze mannequin (in necessities.txt) is the Flask module.

To create a container picture working the maze mannequin, I exploit this Dockerfile.

FROM --platform=linux/amd64


COPY necessities.txt necessities.txt
RUN pip3 set up -r necessities.txt

COPY . .

CMD [ "python3", "-m" , "flask", "run", "--host=", "--port=5555"]

Here’s the code for the agent ( First, the agent asks the mannequin for the dimensions of the maze and the beginning place. Then, it applies its personal technique to discover and clear up the maze. In this implementation, the agent chooses its route at random, making an attempt to keep away from following the identical path greater than as soon as.

import random
import requests
from requests.adapters import HTTPAdapter, Retry

HOST = ''
PORT = 5555

BASE_URL = f"http://{HOST}:{PORT}/maze"

UP, RIGHT, DOWN, LEFT = 0, 1, 2, 3
OPEN, WALL = 0, 1

s = requests.Session()

retries = Retry(whole=10,

s.mount('http://', HTTPAdapter(max_retries=retries))

r = s.get(f"{BASE_URL}/dimension")
dimension = r.json()
print('SIZE', dimension)

r = s.get(f"{BASE_URL}/begin")
begin = r.json()
print('START', begin)

y = begin['y']
x = begin['x']

found_key = False
been = set((x, y))
strikes = [(x, y)]
moves_stack = [(x, y)]

whereas True:
    r = s.get(f"{BASE_URL}/pos/{y}/{x}")
    pos = r.json()
    if pos['here'] == 'Okay' and never found_key:
        print(f"({x}, {y}) key discovered")
        found_key = True
        been = set((x, y))
        moves_stack = [(x, y)]
    if pos['here'] == 'E' and found_key:
        print(f"({x}, {y}) exit")
    dirs = record(vary(4))
    for d in dirs:
        nx, ny = x, y
        if d == UP and pos['up'] == 0:
            ny -= 1
        if d == RIGHT and pos['right'] == 0:
            nx += 1
        if d == DOWN and pos['down'] == 0:
            ny += 1
        if d == LEFT and pos['left'] == 0:
            nx -= 1 

        if nx < 0 or nx >= dimension['width'] or ny < 0 or ny >= dimension['height']:

        if (nx, ny) in been:

        x, y = nx, ny
        been.add((x, y))
        strikes.append((x, y))
        moves_stack.append((x, y))
        if len(moves_stack) > 0:
            x, y = moves_stack.pop()
            print("No strikes left")

print(f"Solution size: {len(strikes)}")

r = s.submit(f'{BASE_URL}/map', json=strikes)

print(r.textual content)


The solely dependency of the agent (in necessities.txt) is the requests module.

This is the Dockerfile I exploit to create a container picture for the agent.

FROM --platform=linux/amd64


COPY necessities.txt necessities.txt
RUN pip3 set up -r necessities.txt

COPY . .

CMD [ "python3", ""]

You can simply run this simplified model of a simulation domestically, however the cloud permits you to run it at bigger scale (for instance, with a a lot larger and extra detailed maze) and to check a number of brokers to seek out one of the best technique to make use of. In a real-world situation, the enhancements to the agent would then be carried out right into a bodily gadget corresponding to a self-driving automotive or a robotic vacuum cleaner. If you wish to enhance the complexity of the simulation and scale into the tens or lots of of hundreds of dynamic entities try AWS SimSpace Weaver.

Running a simulation utilizing multi-container jobs
To run a job with AWS Batch, I must configure three assets:

  • The compute surroundings by which to run the job
  • The job queue by which to submit the job
  • The job definition describing the best way to run the job, together with the container pictures to make use of

In the AWS Batch console, I select Compute environments from the navigation pane after which Create. Now, I’ve the selection of utilizing Fargate, Amazon EC2, or Amazon EKS. Fargate permits me to carefully match the useful resource necessities that I specify within the job definitions. However, simulations often require entry to a big however static quantity of assets and use GPUs to speed up computations. For this cause, I choose Amazon EC2.

Console screenshot.

I choose the Managed orchestration kind in order that AWS Batch can scale and configure the EC2 cases for me. Then, I enter a reputation for the compute surroundings and choose the service-linked function (that AWS Batch created for me beforehand) and the occasion function that’s utilized by the ECS container agent (working on the EC2 cases) to make calls to the AWS API on my behalf. I select Next.

Console screenshot.

In the Instance configuration settings, I select the dimensions and kind of the EC2 cases. For instance, I can choose occasion varieties which have GPUs or use the Graviton processor. I would not have particular necessities and go away all of the settings to their default values. For Network configuration, the console already chosen my default VPC and the default safety group. In the ultimate step, I assessment all configurations and full the creation of the compute surroundings.

Now, I select Job queues from the navigation pane after which Create. Then, I choose the identical orchestration kind I used for the compute surroundings (Amazon EC2). In the Job queue configuration, I enter a reputation for the job queue. In the Connected compute environments dropdown, I choose the compute surroundings I simply created and full the creation of the queue.

Console screenshot.

I select Job definitions from the navigation pane after which Create. As earlier than, I choose Amazon EC2 for the orchestration kind.

To use a couple of container, I disable the Use legacy containerProperties construction possibility and transfer to the subsequent step. By default, the console creates a legacy single-container job definition if there’s already a legacy job definition within the account. That’s my case. For accounts with out legacy job definitions, the console has this selection disabled.

Console screenshot.

I enter a reputation for the job definition. Then, I’ve to consider which permissions this job requires. The container pictures I wish to use for this job are saved in Amazon ECR personal repositories. To enable AWS Batch to obtain these pictures to the compute surroundings, within the Task properties part, I choose an Execution function that offers read-only entry to the ECR repositories. I don’t must configure a Task function as a result of the simulation code will not be calling AWS APIs. For instance, if my code was importing outcomes to an Amazon Simple Storage Service (Amazon S3) bucket, I may choose right here a task giving permissions to take action.

In the subsequent step, I configure the 2 containers utilized by this job. The first one is the maze-model. I enter the title and the picture location. Here, I can specify the useful resource necessities of the container when it comes to vCPUs, reminiscence, and GPUs. This is much like configuring containers for an ECS job.

Console screenshot.

I add a second container for the agent and enter title, picture location, and useful resource necessities as earlier than. Because the agent must entry the maze as quickly because it begins, I exploit the Dependencies part so as to add a container dependency. I choose maze-model for the container title and START because the situation. If I don’t add this dependency, the agent container can fail earlier than the maze-model container is working and in a position to reply. Because each containers are flagged as important on this job definition, the general job would terminate with a failure.

Console screenshot.

I assessment all configurations and full the job definition. Now, I can begin a job.

In the Jobs part of the navigation pane, I submit a brand new job. I enter a reputation and choose the job queue and the job definition I simply created.

Console screenshot.

In the subsequent steps, I don’t must override any configuration and create the job. After a couple of minutes, the job has succeeded, and I’ve entry to the logs of the 2 containers.

Console screenshot.

The agent solved the maze, and I can get all the main points from the logs. Here’s the output of the job to see how the agent began, picked up the important thing, after which discovered the exit.

SIZE {'width': 80, 'top': 20}
START {'x': 0, 'y': 18}
(32, 2) key discovered
(79, 16) exit
Solution size: 437
[(0, 18), (1, 18), (0, 18), ..., (79, 14), (79, 15), (79, 16)]

In the map, the pink asterisks (*) observe the trail utilized by the agent between the beginning (S), key (Okay), and exit (E) places.

ASCII-based map of the solved maze.

Increasing observability with a sidecar container
When working complicated jobs utilizing a number of parts, it helps to have extra visibility into what these parts are doing. For instance, if there may be an error or a efficiency drawback, this data might help you discover the place and what the difficulty is.

To instrument my utility, I exploit AWS Distro for OpenTelemetry:

Using telemetry knowledge collected on this manner, I can arrange dashboards (for instance, utilizing CloudWatch or Amazon Managed Grafana) and alarms (with CloudWatch or Prometheus) that assist me higher perceive what is going on and scale back the time to unravel a difficulty. More typically, a sidecar container might help combine telemetry knowledge from AWS Batch jobs together with your monitoring and observability platforms.

Things to know
AWS Batch help for multi-container jobs is out there in the present day within the AWS Management Console, AWS Command Line Interface (AWS CLI), and AWS SDKs in all AWS Regions the place Batch is obtainable. For extra data, see the AWS Services by Region record.

There isn’t any further value for utilizing multi-container jobs with AWS Batch. In reality, there isn’t any further cost for utilizing AWS Batch. You solely pay for the AWS assets you create to retailer and run your utility, corresponding to EC2 cases and Fargate containers. To optimize your prices, you should utilize Reserved Instances, Savings Plan, EC2 Spot Instances, and Fargate in your compute environments.

Using multi-container jobs accelerates growth occasions by decreasing job preparation efforts and eliminates the necessity for customized tooling to merge the work of a number of groups right into a single container. It additionally simplifies DevOps by defining clear part duties in order that groups can rapidly determine and repair points in their very own areas of experience with out distraction.

To study extra, see the best way to arrange multi-container jobs within the AWS Batch User Guide.



Please enter your comment!
Please enter your name here