1. 6/25/2018 1:59 AM
BRK3219
Meet the most demanding HPC customer needs on Azure with Cycle Computing and Batch
Mark Scurrell
Principal Program Manager
Jason Stowe
Principal Group Program Manager
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

2. Today we’ll cover
Moving big compute to the cloud - customer lessons learnt
Azure infrastructure for big compute and HPC workloads
Azure Batch for SaaS and new cloud-native apps
CycleCloud for clustered, enterprise workloads

3. Big Compute and HPC workloads
Industries
Aerospace
Automotive
Science
Energy
Architecture
Financial Services
Manufacturing
Media & Entertainment
Workloads
Test execution
Car crash simulations
Audio & video transcoding
Rendering
Data ETL
Video pre- & post-processing
Genomics
Financial risk simulations
Deep learning & AI training
OCR
Compiled MATLAB
Engineering stress analysis
R at scale
Oil reservoir simulation
Motorsports simulations
Image analysis
Computational fluid dynamics

4. #1 – Zero waiting in line for compute

5. #2 – Ask questions of any scale
Ask the right question,
regardless of scale
Customers use 100s to 1,000s
Of cores to answer business-critical
Questions they couldn’t have done before.

6. #3 – Users with unique requirements are OK
Trivial to support different use cases
Different RAM ratios, GPU, FPGA, Application/OS needs
Move workloads that don’t fit internally to Cloud

7. #4 – Cloud gets faster & cheaper over time…

8. #5 – Time and Cost are the sole metrics that matter

9. Everything you don’t think about!

10. #6 – Accelerating answers, accelerates people
720 (hours)
720
Computing
Analysis
2880 hours / 120 Days to Decision
SCALABLE COMPUTING (in hours)
1456 hours / 60.6 Days to Decision 8
ANTICIPATED BENEFIT (in hours)

11. #6 – Accelerating answers, accelerates people
720 (hours)
720
Computing
Analysis
2880 hours / 120 Days to Decision
SCALABLE COMPUTING (in hours)
Higher Quality Output,
Iterative Analysis, Less Context Switching
Computing & Analysis
POST ADOPTION: AGILE DESIGN PROCESS 8

12. #7 – Every smart person gets their own sandbox
User
User
User
User
User
User
User
User
User
User
User
User
Old: Shared internal cluster
New: Cluster Per Researcher
Competition for resources
Waiting in line for compute
Shared downtime
Remove bottlenecks
Cost controls to manage $
No waiting = 2x faster users 12

13. Infrastructure for big compute and HPC

14. Application types Embarrassingly parallel: Tightly coupled:
Applications do not communicate
May share common store & data
May have dependencies
E.g. Monte Carlo simulations, transcoding, rendering
Tightly coupled:
Applications communicate; mainly use MPI
Requires low latency, high bandwidth networking for scale
E.g. car crash simulation, fluid dynamics, AI training

15. Azure compute platform
Parallel R
CycleCloud
SaaS / Client
Solution
Batch AI Training
VFX Plug-Ins
Batch Shipyard
Azure Batch
VM Management & Job Scheduling
Platform
VMs & VMSS
Infrastructure
Hardware

16. No-compromise HPC and AI VMs
Up to 16 cores, 3.2 GHz E V3 Haswell processor
Up to 224 GiB DDR4 memory
FDR InfiniBand (56 Gbps, 2.6 microsecond latency)
2 TB of local SSD ND
Up to 4 NVIDIA Pascal P40 GPUs
Up to 24 cores
Up to 448 GiB memory
Up to 3 TB of local SSD
FDR InfiniBand NC
Up to 4 NVIDIA Tesla K80 GPUs
Up to 24 cores
Up to 224 GiB memory
Up to 1440 GiB of local SSD
FDR InfiniBand
NCv2
Up to 4 NVIDIA Pascal P100 GPUs
Up to 24 cores
Up to 448 GiB memory
Up to 3 TB of local SSD
FDR InfiniBand NV
Up to 4 NVIDIA Tesla M60 GPUs
Up to 24 cores
Up to 224 GiB memory
Up to 1440 GiB of local SSD

17. Why InfiniBand RDMA matters?

18. Why InfiniBand RDMA matters?

19. A Visual Cloud Collaboration Platform for Engineering
6/25/2018 1:59 AM
A Visual Cloud Collaboration Platform for Engineering
Instant on
Collaboration
Data management
Familiar desktop interface
On-demand licensing
Finite element analysis
Analyze and Optimize Assemblies
Best-in-class optimization
Motion Analysis and Optimization
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

20. Demo

21. Azure Batch

22. Microsoft Build 2017
6/25/2018 1:59 AM
Azure Batch
Rendering
Media transcoding & pre-/post- processing
Test execution
Monte Carlo simulations
Genomics
Deep Learning
OCR
Data ingestion, processing, ETL
R at scale
Compiled MATLAB
Engineering simulations
Image analysis & processing
Enable applications and algorithms to easily and efficiently run in parallel at scale
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

23. Applications / Algorithms
6/25/2018 1:59 AM
Azure Batch Concepts
Queue
Jobs & Tasks
Applications / Algorithms
Pool of VMs
INPUT
OUTPUT
© 2013 Microsoft Corporation. All rights reserved. Microsoft, Windows, and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

24. Azure Batch focus areas
Microsoft Build 2017
6/25/2018 1:59 AM
Azure Batch focus areas
Capacity on demand
Jobs on demand
Scale according to load
Pay by the minute
Efficient
Elastic
1 to 10,000’s VMs
1 to millions of tasks
No charge for Batch;
pay for used resources
No head node
Use low-priority VMs
Cost
effective
Scale
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

25. Elasticity

26. Scale
Number of cores

27. Azure Batch capabilities
Access via API’s, CLI’s, and UI’s:
.NET, Java, Node.js, Python, REST
PowerShell, x-plat Azure CLI
Azure Portal, Batch Labs x-plat client UI
Choice of VMs:
Windows or Linux
Standard or custom images
Windows pool can use AHUB (new)
Use low-priority VMs (now GA)
Rich app management:
Get apps from blobs, Batch app packages, package managers, custom VM images
Docker container images (coming soon)
Pool scaling:
Manual or automatic
VM networking:
Pool VMs can be in a VNET
Job scheduling:
Supports both embarrassingly parallel and tightly coupled MPI jobs
Run > 1 task in parallel per node
Detect and retry failed tasks
Can set max execution time for jobs and tasks
Task dependencies
Job prep and cleanup tasks
Monitoring:
VM monitoring and auto-recover
Metrics and logs available via Portal and API

28. Demo: Azure Batch

29. Batch Low-priority VMs – now GA
Microsoft Build 2017
6/25/2018 1:59 AM
Batch Low-priority VMs – now GA
Cheaper compute - up to 80% discount; fixed price
Uses surplus capacity - availability can vary; VMs can be preempted
All Batch VM sizes and regions
Get work done for lower cost, faster, or do more for same price
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

30. Batch low-priority VMs
Microsoft Build 2017
6/25/2018 1:59 AM
Batch low-priority VMs
Batch pools:
Can contain both low-priority and/or dedicated VMs
Scale both VM types independently using resize operation or auto-scale
If preemption, pool automatically seeks to target to replace preempted VMs
Batch tasks:
If VMs preempted, interrupted tasks are automatically re-scheduled & re-executed
// Create a Pool with mixture of dedicated and low-priority
CloudPool pool = PoolOperations.CreatePool (“Pool1”, “Standard_D14_v2”, vmConfig);
pool.TargetDedicatedComputeNodes = 20;
pool.TargetLowPriorityComputeNodes = 80;
pool.Commit();
// Get Pool info
CurrentDedicated = pool.CurrentDedicatedComputeNodes;
CurrentLowPriority = pool.CurrentLowPriorityComputeNodes;
// Resize Pool – no dedicated, all low-pri
pool.Resize(0, 100);
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

31. Batch low-priority VMs
Microsoft Build 2017
6/25/2018 1:59 AM
Batch low-priority VMs
Suitable workload characteristics:
Tolerant of interruption
Tolerant of reduced capacity
Suitable jobs:
Job completion time flexibility
Shorter tasks
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

32. Batch low-priority flexibility
Microsoft Build 2017
6/25/2018 1:59 AM
Lowest cost
Batch low-priority flexibility
Lower cost, with guaranteed baseline capacity
Lowest cost, while maintaining capacity
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

33. https://github.com/Azure/batch-shipyard
Azure Batch Shipyard
Drive Batch using Python command line tool and JSON recipes (no development or API usage required)
Supports Linux Docker container images
Data movement:
Azure Files, Azure Blobs, NFS, GlusterFS
Create and manage NFS and GlusterFS file systems

34. https://github.com/Azure/doAzureParallel
R - doAzureParallel
Scale up R execution using Batch
Parallel backend package for popular foreach package
Each iteration of foreach loop runs as a Batch task

35. R - doAzureParallel Hiscox is a leading specialist insurer
Flood risk is key growth area in the US
Need to be able to run very large-scale catastrophe research
Without Azure Batch, the analysis would take 8 months and would limit their research options
With Azure Batch, the research can be run in a 48 hour window; allows Hiscox to extend the research performed and remain competitive in the Flood market

37. What does Cycle Do? Hybrid Big Compute Orchestration
Measuring Carbon Footprint in Sub-Saharan Africa
Fast on-boarding, no changes to the application
Leverage internal & Cloud resources
Use case could not be done without CycleCloud 1 3 2
Local
Cloud 4

38. The Challenges: Cloud & HPC Big Compute
Existing Workflows
Applications
Budget Controls
User
Inputs
Security Stack
Authorization
Data Dependencies
Scheduler
Scalability
Job scripts & data
Cloud accounts
Storage / Data sources
OS variations
AD / LDAP Authorization
IT LOB
Inputs

39. The Solution: CycleCloud
Existing Workflows
Applications
Budget Controls
User
Inputs
Security Stack
Authorization
Data Dependencies
Scheduler
Scalability
Job scripts & data
Cloud accounts
Storage / Data sources
OS variations
AD / LDAP Authorization
Internal
Audit/Compliance data
Usage data (User, Group, App)
Job run-time by instance data
AppServer platform
IT LOB
Inputs

40. Who is Cycle Computing?
Leader in Big Compute Cloud Management Software
Pioneering Cloud Big Compute/HPC for 10 years
Compute hour growth: 7x every 2 years
Managing 380M hours in 2016
CycleCloud Value Proposition
Simple Managed Access to Big Compute
Accelerating Innovation for the Enterprise
Faster time to result, with cost control
Our customers
Fortune 500, startups, and public sector
Big 5 Life sciences & pharma, Big 5 financial services, Fortune 100 Insurance, Fortune 100 manufacturing, aerospace & electronics

41. Cycle Unleashes Productivity…
Existing workflows, existing schedulers
Instant access to resources
Access scale up/down, with control
Reliable tools for IT
Workflows linking internal and external clouds
Tools to manage & control costs
Secure, consistent cloud access
Ability to link usage to costs
Support for Active Directory authorization

42. Cloud BigCompute Platform: Visibility & Reporting
Compute Instances: Performance & cost
Data utilization: Multiple Endpoints

43. Transform Business Process…
Fortune 100 Financial
Internal
2,500 core
cluster
CycleCloud Secure 8,000 core
Cluster
100,000 hr. jobs, now done in a day on cloud
Regulatory workload done 5x faster at 20% the cost

44. Transform Product Development…
Applications
QA team #1
Design team #2
Design team #1
Validation team #1
Manage many groups with many users running many applications
Increased compute resources 10X
200% faster time to market

45. “Transform Thinking”…
The Broad Institute
Internal Cluster
Cloud
On 8,000 cores in 1.5 hours
2 weeks, from concept to completion
Accessing 50,000 cores
30 years of compute in 6 hours

46. A Predictable Process to Production…
Plan
Pilot
Production
Promote
Identify workflow & path to Cloud resources
Test workflow, Clear success criteria Define controls
Finalize security plan, Review results,
Move to production
Leverage tools to expand user base, new applications

47. Demo

48. CycleCloud Summary Based on Azure infrastructure
Can be used on-premises and on Azure VMs
Focus areas:
Hybrid on-premises and cloud
Run existing HPC environments in the cloud
End-to-end solution for IT
Governance and management; e.g. core and $$$ quotas
If interested, send to

49. Reference Other MS Ignite sessions: More information:
BRK2275: Building a Hollywood blockbuster using Azure Batch rendering
BRK4034: Working with models for machine learning and Azure Batch AI
BRK2151: Past, present, and future: GPU and AI infrastructure on Microsoft Azure
More information:
Azure Batch documentation:
Azure Hybrid Use Benefit:

50. Please evaluate this session
Tech Ready 15
6/25/2018
Please evaluate this session
From your
Please expand notes window at bottom of slide and read. Then Delete this text box.
PC or tablet: visit MyIgnite
Phone: download and use the Microsoft Ignite mobile app
Your input is important!
© 2012 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.
The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

51. 6/25/2018 1:59 AM
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.