Costing cloud

Introduction

Let’s do some ballpark numbers without further ado… these are for purposes of “order of magnitude” only. Getting more precise takes a bit more work.

• A fairly powerful cloud computer (instance) costs $1 per hour: 8 cores, 128GB RAM
  • Other instance type costs scale with horsepower from ‘raspberry pi’ to ‘8-GPU-monster’
• It costs $0.023 per Gigabyte per month to store data in a cloud bucket, also called object storage
  • Archival cost is one sixth at $0.004 / GB / month or $48 per Terabyte per year
• Data ingress (upload to the cloud) is free
• Data egress (download from the cloud) is $0.10 per GByte
  • At UW a certain amount of your egress charges are waived
• Attaching a drive to a cloud machine costs $0.10 per GByte per month
• Cloud machines can be stopped when you are not using them; and re-started in the same state
  • While they are stopped they cost very very little
  • When you re-start them you can do so on more powerful computers
    • This allows you to develop on cheap / small machines and run big processing jobs on big machines
• The prices for cloud computers given here are on-demand prices: You get the machine for as long as you like.
  • You can also get preemptible machines at typically 20% to 40% of the on-demand rate
• A cloud machine with an attached current-generation NVIDIA Tesla V100 GPU costs $3 per hour
  • A preemptible version of the same costs $0.63 per hour

This page provides some additional ideas and guidelines for costing cloud usage.

Links

• AWS Grants
• AWS monthly cost calculator
• Google Cloud Platform (GCP) research credit program
• Azure Research Credits
  • This program may currently be discontinued
• UW research overhead waiver information

Warnings

• Research credits are stopgap or transitional support, not ‘ongoing funding’
  • By stopgap we mean the researcher uses cloud research credits to…
    • …evaluate the feasibility of the public cloud
    • …tide over until standard research budgets can cover cloud costs
  • Research credit awards typically last 12 months; amounts are based on cost estimation for the proposed work

Building cloud computing into a research proposal

• Review your status and options with regard to cloud research credit grants
• Verify with the funding agency that cloud computing will be a welcome component of your proposal
• We hold office hours and are happy to consult with you
• Create a cost plan per year
  • Estimate the compute power per machine and number of VMs you will want to secure
  • Estimate the cost of associated memory
  • Estimate the cost of separate storage
  • Estimate your egress cost (cost of retrieving data from the cloud)
  • Estimate the cost of other services you intend to use
• As necessary: Provide justification of your choice of cloud
  • You can compare to alternative cloud providers
  • You can refer to the advantages of using the cloud over purchasing and maintaining equipment
  • Cite case studies presented here at cloudmaven.org
  • Remember that the case for cloud includes strengths:
    • Computing scale
    • Data security
    • Advanced tools and services
    • Similarity of compute environment to local machines
    • Rapidly growing user base and commensurate open resources from which to build your solutions
• If you are budgeting for cloud resources and you use a UW IT cloud service
  • You do not pay indirect cost overhead (F&A)
  • See this web page for more information

Cloud Research Credits

Generous research credit grants (circa 2017) are commonly provided by cloud vendors, specifically Microsoft Azure and GCP. These grants can have a value as high as $20,000 and a typical duration of 12 months. They are considered non-renewable after that time.

Having a research credit grant of this type is independent of writing a grant proposal to a funding agency such as NIH or NSF. If a cloud research credit grant is secured by you en route to writing a proposal to an agency like NSF: You may wish to reference that grant in your proposal as a means of defraying computing costs in year one of the proposed effort.

FAQ

Q: How does the cloud work?

A: This answer is the tip of the iceberg: ‘How you log in to a computer on the cloud and run programs.’ See ensuing remarks for a quick terminology overview. First you would get a cloud account from some provider; we work with Microsoft Azure, Amazon Web Services and Google Cloud Platform. Log in through a browser to their Console and start an instance. This is equivalent to renting a computer by the hour. This computer will have an ip address; and you can log on to that machine for example using ssh or PuTTY (a terminal window). Once there you can start building your computing enviornment, importing data, compiling and running programs, and so on; as you would on a desktop machine. In other words: Get back to your research.
This snapshot of ‘moving your work to the cloud’ does not yet get to the why, which we address below. There is much more to the story so to get a sense of this we suggest looking at the case studies provided here.

Q: Why would I consider moving to the cloud?

A: Reduced risk, reduced hassle, lower cost, and massive on-demand compute power should you need it. (If you need additional reasons keep reading.) Let’s cover these points in slightly more detail. Traditionally you must purchase and maintain computers and then worry about disks crashing. In the cloud someone else has set up the computers already; and the cloud provider is constantly creating backups of your disk partitions. Cloud operating systems are maintained and updated and patched by the cloud provider (although you can also do security patches as needed). In short you never buy and/or maintain hardware and you won’t worry about losing your work / data. Migrating to more powerful machines is trivially easy so your hardware does not gradually become obsolete. You do have to learn and implement cost management and security practices; so there is a learning curve. When you are not working on cloud machines you turn them off. Your work environment as preserved at negligible cost (10 cents per GB per month). You do not pay for computing horsepower when you are not using it.
And should you need massive computing power the cloud has thousands of machines available. You can scale your compute task out and complete your processing quickly.

Q: How fast/powerful is a given cloud computer?

A: Cloud computers – also called instances or Virtual Machines – exist in many configurations and power levels. A small low-power machine will cost perhaps $0.01 USD to operate per hour and a pretty powerful machine might cost $1.00 per hour. The most powerful machines you can find on the cloud can cost $15 to $25 per hour. You can read about cost and power at the cloud vendor’s website. At UW we focus on three vendors: AWS, Microsoft Azure and Google Cloud Platform. From an instance command line you can issue commands to learn more; such as…

% lscpu
% cat /proc/cpuinf

Q: How much does a cloud computer cost?

A: One penny per hour for a basic low-power instance. $0.40 for a powerful machine, $1.00 for very powerful.

Q: How much does cloud disk space cost?

A: RAM memory on a cloud instance is part of the cost of that instance. Disk space costs $0.10 per Gigabyte per month on AWS; other vendors such as Microsoft Azure and Google are comparable. These disk resources are called Volumes and are directly associated with cloud compute instances, in contrast to cloud storage.

Q: How much does cloud storage cost?

A: Storage is unlimited and separate from the concept of disk space (see above). Rapid-access storage costs $0.023 (2.3 cents) per Gigabyte-month on the AWS cloud. Other vendor costs are comparable. Low-access storage is cheaper, as low as $0.004 USD per GByte-month. This translates to $48 per Terabyte-year. This cost has dropped steadily over the past few years. Cloud storage is extremely secure and reliable. This links to a comprehensive page on AWS storage costs.

Q: How is cloud object storage different from disk space on the cloud?

A: Object storage is distinct from and cheaper than disk space by a factor of four. Content that you do not need to access repeatedly frequently is cheaper to place in storage.

Q: Is there an overview of costs?

A: For AWS this diagram

might help but it may also get out-of-date.