#Paper

CloudArray paper was published at IGARSS 2016: CloudArray: Easing huge image processing.

#Table of Contents

1. Overview
2. Architecture
3. Installation
4. Usage

CloudArray (try it here!) is a programming abstraction that eases big data programming in the cloud. CloudArray loads data from files then books and configures the right amount of resources (VMs, containers) able to process it. Data loading and resource management are entirely automatic and performed on-demand.

CloudArray builds on top of Julia native DistributedArrays abstraction, a multi-dimensional array whose data is transparently stored at distributed computers. Indeed, a CloudArray is a DistributedArray whose data and resource managements are automatic as the figure bellow illustrates: data load, VM booking, Julia workers configuration on top of Docker containers.

Therefore, existent codes that use DistributedArrays don't need to be adapted in order to use CloudArray. You just need to include CloudArray and use your cloud account, no need to manually interact with your cloud provider.

You are very welcome to try CloudArray from CloudArrayBox, a Web front end hosted at Azure.

CloudArray design is composed by two layers (c.f. Figure Architecture):

• CloudArray.jl is an extension of DistributedArrays.jl which automatically loads data from files (or other I/O stream) and store it at distributed Workers as DArray.
• Infra.jl books virtual machines (VMs) and creates, configures, and instantiates Docker containers on top of VMs. Then Julia Workers are configured and deployed on containers.

Download Julia 0.4

Debian-based Linux distros as Ubuntu or through

sudo apt-get install sshpass 

OS X through macports:

sudo port install sshpass

First load CloudArray package:

using CloudArray

Then tell CloudAarray the machine address and the password to passwordless SSH login:

CloudArray.set_host(host_address,ssh_password)

To use CloudArrayBox VMs to test CloudArray use the following parameters:

CloudArray.set_host("cloudarray.ddns.net","cloudarray@")

You may try CloudArray from your browser (CloudArrayBox), without installing any software at all. Just login with your Google account and run both Julia interface (Master) and cloud processing backend (Workers).

We just kindly ask you to not overload our few Azure VMs which are available to Julia community to test CloudArray for free. In other words, please do not run large parallel or high-processing codes.

You can also use CloudArray your computer with your local Julia 0.4 installation ([see installation instructions](Julia 0.4)) and use CloudArrayBox to deploy Workers.

Otherwise, you can define a customized runtime environment with your own cloud provider having Master and Workers configured as you prefer.

CloudArray main constructors are very simple and can be created by using an Array or a file.

You just need to tell DArray constructor which Array should be used to construct your CloudArray:

DArray(Array(...))

In this example, we first create the array arr with 100 random numbers then we create a CloudArray with the arr data:

arr = rand(100)
cloudarray_from_array = CloudArray.DArray(arr) # will take less than one minute

We can now access any cloudarray_from_array value as it would be a local array:

cloudarray_from_array[57]

If you are dealing with big data, i.e., your RAM memory is not enough to store your data, you can create a CloudArray from a file.

CloudArray.DArray(file_path)

file_path is the path to a text file in your local or distributed file system. All lines will be used to fill DArray elements sequentially. This constructor ignores empty lines.

Let's first create a simple text file with 100 random numbers.

f = open("data.txt","w+")
for i=1:100
    if i==100
        write(f,"$(rand())")
    else
        write(f,"$(rand())\n")
    end    
end
close(f)

Then we create a CloudArray with data.txt file.

CloudArray.cloudarray_from_file = DArray("data.txt")

Let's perform a sum operation at cloudarray_from_file:

sum(cloudarray_from_file)

This sum was performed locally at the Master, you can exploit DArray fully parallelism with further functions such as parallel Maps (pmap) and Reductions. See here more information on Parallel programming in Julia.

If you want to tune your CloudArray, you can directly use the CloudArray core constructor:

carray_from_task(generator::Task=task_from_text("test.txt"), is_numeric::Bool=true, chunk_max_size::Int=1024*1024,debug::Bool=false)

Arguments are:

• task_from_text same as file_path.
• is_numeric set to false if you need to load String instead of Float.
• chunk_max_size sets the maximum size that is allowed for each DArray chunk.
• debug enables debug mode.

As follows, we create a CloudArray by using the data.txt file which holds numeric values, then second argument is set to true. We'll set the third argument (chunk_max_size) to 500 so DArray chunks will not have more than 500 bytes each.

custom_cloudarray_from_file = DArray("data.txt", true, 500)

Now let's define and perform a parallel reduction at the just-created CloudArray:

parallel_reduce(f,darray) = reduce(f, map(fetch, { @spawnat p reduce(f, localpart(darray)) for p in workers()} ))
parallel_reduce(+,custom_cloudarray_from_file)

The result is the sum of all values of custom_cloudarray_from_file. Each DArray chunk performed in parallel the sum of the part of the DArrau it holds. The result is sent to the Master which performs the final sum. The function map is used to get the values with the fetch function.

You don't really need to know it, but if you are curious on how your data is stored, you can get further information such as:

@show custom_cloudarray_from_file.chunks
@show custom_cloudarray_from_file.cuts
@show custom_cloudarray_from_file.dims
@show custom_cloudarray_from_file.indexes
@show custom_cloudarray_from_file.pids

Please read DistributedArrays documentation to better understand these low-level details if you want.

Infra.jl provides an interface to manage Docker containers on top of Ubuntu machines. Infra.jl allows to configure, create, delete, list, and monitors containers as described next.

Configures passwordless SSH connections at host h whose password is p.

This function calls the cloud_setup.sh script which requires sshpass.

CloudArray.set_host("cloudarray.cloudapp.net","password")

Launches Docker containers and adds them as Julia workers configured with passwordless SSH. This function requires sshpass to be installed:

• Debian-based Linux distros as Ubuntu:

sudo apt-get install sshpass

• OS X through macports:

sudo port install sshpass

create_containers(2,3,1024) # 2 containers with 3 CPU Cores and 1gb RAM
create_containers(1,2,512)  # 1 container with 2 CPU Cores and 512mb RAM

Removes the specified container(s)/worker(s).

delete_containers(3)    # delete container 3
create_containers(1:5)  # delete from 1st to 5th container
create_containers(all)  # delete all containers

Returns the list of all containers' processes identifiers (IDs).

containers()

Gets the number of available container processes.

ncontainers()

List container(s) as a sorted list.

list_containers()

Returns the container memory usage.

mem_usage(number_of_container)

Returns the container CPU usage (%).

cpu_usage(number_of_container)

Returns the number of kilobytes read and written by the cgroup.

io_usage(number_of_container)

Returns networking TX/RX usage.

tx = number of bytes transmitted

rx = number of bytes reiceved

net_usage(number_of_container)

• Support other cloud infrastructures
  • Amazon EC2
  • OpenStack
• Set a price threshold
• Provide different QoS
  • E.g., Pricy and fastest vs. Cheapest and not so fast
• Add the following containers monitoring functions:
  • io_usage(key::Integer)
• Let users define which CSV separator should be used
• RESTful API
• CloudDataFrame: extend CloudArray to support DataFrame
• Use Azure REST API
• Redundancy: if Julia fails, it removes containers (mask this fault)
• Create tests
• Use @acc to improve performance

• Explicitly release resources (containers and VMs) after usage
• Use Julia Module to be able to call cloud_setup.sh
• CloudArrayBox logo transparent
• Replace sshpass by another means to authenticate through SSH
  • maybe require users' public key?

CloudArray is developed by the LaCCAN lab at the Computing Institute of the Federal University of Alagoas (Brazil). CloudArray is funded by Microsoft Azure Research Award, Brazilian National Council for Scientific and Technological Development (CNPq), and Alagoas Research Foundation (FAPEAL).