High Performance Computing with do doAzur ureP ePar arallel - - PowerPoint PPT Presentation

High Performance Computing with

do doAzur ureP ePar arallel allel

Using Azure as your Parallel-Backend for Embarassingly Parallel work Microsoft JS Tan

Azure Big Compute

Commodity, most value for cost Fast processors, higher memory-to- core ratio, SSDs Most memory, Intel Xeon processors HPC/Low Latency VMs for compute intensive workloads GPU enabled VMs for Visualization/ Compute Fast processors, lower-memory to core ratio, SSDs

Azure Infrastructure

What is Batch?

APP Tasks are assigned to computers/VMs Many individual tasks Many computers/VMs

Scenarios

• A quant back-testing portfolio strategies
• A data scientist optimizing their model & parameter tuning
• A life-science researcher doing genome sequencing

What do they have in common?

• Scale – computationally expensive work - need to scale up in order to

get results back quickly

• Minimal IT Management – the user is the domain specialist, not an IT

specialist

• Elastic compute – temporary need for a lot of capacity
• Cost effective – low cost strategies are important!

+ They are all probably using R…

doAzureParallel is...

A R package that uses Azure as a parallel-backend for popular open source tools to use – foreach, caret, dplyr, etc.

Foreach using doAzureParallel

foreach (i = 1:100) %dopar% { myParallelAlgorithm(...) }

Microsoft Azure

doAzureParallel on Azure Batch

Azure Batch is a platform service that provides easy job scheduling and cluster management, allowing applications or algorithms to run in parallel at scale.

• Capacity on demand; jobs on demand
• Autoscale (more on this later)
• Minimal cluster management (node failure, install, etc)
• Hardware choice – use any VM size
• Pay by the minute
• Cost effective – no charge for using it, you only pay for the VMs
• More cost effective – low priority VMs (more on this later)

If you want to run jobs using elastic compute, Batch is a great fit!

Scale

• From 1 to 10,000 VMs for a cluster
• From 1 to millions of tasks
• Your selection of hardware:
• General compute VMs (A-Series / D-Series)
• Memory / storage optimized (G-Series)
• Compute Optimized (F-Series)
• GPU enabled (N-Series)
• Results from computing the mandelbrot set when

scaling up:

Local machine 5 parallel workers 10 parallel workers 20 parallel workers

Minimal Cluster Management

• Abstract away complex Azure/cloud concepts
• Zero IT-level management
• Work entirely in R Studio
• Monitor / Debug your jobs directly in R

studio

• Manage your cluster and multiple jobs

directly in R studio

• The results of your distributed, large scale

work can be returned directly to your R session

Minimal code change

• Minimal code change to use doAzureParallel
• Easy to use and you can get started in just a few lines of code

Elastic Compute

• Compute on-demand
• Create/delete your cluster as you

need

• Autoscaling pool = maximizing cloud

elasticity

• Long running batch jobs / overnight
• Daily scheduled work – pre-provision

cluster so its ready for you at the beginning of the day

• Bursty work

Cost Effective

• Low-Priority = (extremely) Low Costs
• Provisioning VMs from Azure’s surplus capacity at 80% discount
• Your Azure cluster can contain both regular (dedicated) VMs and low-priority VMs

My Local R Session

Azure Batch

Low Priority VMs

at up to 80% discount

Dedicated VMs

Cost Effective: More about Low Priority

When should I use it?

• Long running work that can be broken into smaller pieces and work that doesn't have a strict time

limit to complete

• Experimentation, testing, evaluating models

What you need to know when using it:

• Possibility that Azure
• will not allocate your VMs OR
• that it will take some or all of the capacity back
• If a node is pre-empted
• Azure Batch will replace your node for you
• Azure Batch will reschedule your work so that you job can successfully complete

Low Priority Scenarios

Dedicated Low-priority

Lowest Cost Lower cost + guaranteed baseline capacity Lower cost + maintaining capacity w/ autoscale

Azure Batch Pool Azure Batch Pool Azure Batch Pool Preempted Capacity Time Capacity Time Capacity Time

Questions? www.github.com/azure/doazureparallel

https://aka.ms/earl2017

What’s new with doAzureParallel?

• Low priority support a
• Richer Job Management experience a
• Resource Files to preload data a
• Parameter Tuning integration with Caret a
• Simple connector to Azure Blob Storage a

R + Azure Batch

So what R workloads work great on Azure Batch?

• Simulation based work (VaR calculation, back-testing, monte-carlo simulations, financial

modelling)

• Parameter Tuning / Model Evaluation (grid search, random search, cross validation, etc)
• Computing against data / ETL jobs / Data-prep jobs

What industries / verticals might be interested in using this?

• Financial Services
• Education & Research
• Sports analytics

doAzureParallel (since initial release)

• Initial release in March
• Grass roots strategy
• End-user focused
• Financial Services targeted / key messaging has been around simulation based work
• Interest from the field
• Feedback

Azure Batch

Low Priority VMs

at up to 80% discount

Dedicated VMs

My Local R Session