No Reservations: A First Look at Amazons Reserved Instance - PowerPoint PPT Presentation

No Reservations: A First Look at Amazon’s Reserved Instance Marketplace Pradeep Ambati, David Irwin, and Prashant Shenoy University of Massachusetts Amherst

Cost vs. Risk Tradeoffs in IaaS Clouds Expensive Cost ( per hour ) On-Demand Reserved Cheap Spot Rejection Risk, Demand Risk, Rejection Risk, Non-Revocable Non-Revocable Revocable Risk 2

Reserved Instance (RI) Risks • RI’s only cheaper than on-demand if highly utilized • Accurately forecasting demand over long periods is challenging - Unforeseen events like COVID-19 can substantially change demand • Reserved VMs expose users to substantial demand risk - Due to the gap between forecasted and actual demand • To mitigate demand risk… - …Amazon operates the Reserved Instance Marketplace (RIM) 3

Reserved Instance Marketplace (RIM) • RIM allows users to sell VM reservations at a price they set - After listing reserved VMs, EC2 posts them on the RIM • Enables purchasing variable and shorter-term VM reservations - AWS only offers 1-year and 3-year reservations on the market 4

Comparing RIM with the Spot Market • RIM is a competitive market with multiple buyers and sellers - In contrast, EC2’s spot market has only a single seller (Amazon) - Spot price not set based on instantaneous supply/demand • RIM is akin to the housing market with many unique listings - Listing value defined by a large set of attributes - E.g., VM type, number of instances listed, term duration, supply/demand, etc. • Spot market is akin to the stock market - Uniform pricing of many identical assets (VMs) across regions and AZs 5

RIM Pricing • Clearly, RIM’s usefulness is a function of its prices - But, EC2 does not automatically archive RIM price data • We started monitoring and archiving RIM prices in 2018/9 - Query RIM using EC2’s python Boto3 API every 30 minutes - Collect data for all VM types in 69 AZs across 22 regions • Publicly released data at UMass Trace Repository - http://traces.cs.umass.edu/index.php/Main/Cloud 6

A First Look at RIM Data • Analyze RIM data from 2018/9 to 2020/5 - Reveal important market characteristics - Identify potential reserved VM optimization opportunities • Key Market Characteristics - Market volume - VM type - Term duration - Time-on-market - Comparison with on-demand and spot 7

Market Volume across Regions Instances ECUs Average Market Volume Average Market Volume 1600 24K (#instances listed) (#ECUs listed) 1200 18K 800 12K 400 6K 0 0K u u u e e u a s a a s s s u u s p p - - - - - - - - - e w e w c w n e s a a e o a o e e e s s n s u s s s r t t t t t t - t - t t h - - r - h 1 2 - 1 2 a 1 e 1 e l a - a 1 s s t - t - 1 1 Region us-east-1 and us-west-2 are largest regions 8

Market Volume across Regions Instances ECUs Average Market Volume Average Market Volume 1600 24K (#instances listed) (#ECUs listed) 1200 18K 800 12K 400 6K 0 0K u u u e e u a s a a s s s u u s p p - - - - - - - - - e w e w c w n e s a a e o a o e e e s s n s u s s s r t t t t t t - t - t t h - - r - h 1 2 - 1 2 a 1 e 1 e l a - a 1 s s t - t - 1 1 Region Lower market volume may increase risk of using RIM 9

Market Volume by VM Type us-east-1 Instances ECUs Average Market Volume Average Market Volume 1000 8K (#instances listed) (#ECUs listed) 750 6K 500 4K 250 2K 0 0K c4.large c5.large m4.large i3.xlarge c5.4xlarge r5.large r4.large c5d.large m5.2xlarge c5d.9xlarge Instance Type c4.large and c5.large are most popular VM types 10

Market Volume by VM Type us-east-1 Instances ECUs Average Market Volume Average Market Volume 1000 8K (#instances listed) (#ECUs listed) 750 6K 500 4K 250 2K 0 0K c4.large c5.large m4.large i3.xlarge c5.4xlarge r5.large r4.large c5d.large m5.2xlarge c5d.9xlarge Instance Type RIM lagging indicator of instance type popularity 11

Listing Volume vs. Term Duration us-east-1 Instances ECUs 9K 100K #Instances Listed 80K #ECUs Listed 6K 60K 40K 3K 20K 0K 0K 1 4 7 1 1 2 - - - 0 3 5 3 6 9 - - - 1 2 3 2 4 6 Duration (months) Short and long term durations more plentiful 12

Listing Volume vs. Term Duration us-east-1 Instances ECUs 9K 100K #Instances Listed 80K #ECUs Listed 6K 60K 40K 3K 20K 0K 0K 1 4 7 1 1 2 - - - 0 3 5 3 6 9 - - - 1 2 3 2 4 6 Duration (months) Discrepancy at 25-36 months may indicate buyer’s remorse 13

Listing Price (c4.large) vs. Term Duration us-east-1 Minimum Price Average Price 8 Price (cents/hr) 1Y Reservation Price 6 3Y Reservation Price 4 2 0 1-3 4-6 7-9 10-12 13-24 25-36 Duration (months) Average listing price decreases (nearly) linearly 14

Time-on-the-Market across Regions Average Time on the Market 90 60 (Days) 30 0 u u u e e u a s a s s s u u s p a p - - - - - - - - - e w e w e w c n s a a e o a o e e e s s n s u s s r s t t t t t t - t - t t h - - r - h 1 2 - 1 a e 2 1 1 e l a - a 1 s s t - t - 1 1 Region Largest regions have the longest average time on the market 15

Time-on-the-Market across Regions Average Time on the Market 90 60 (Days) 30 0 u u u e e u a s a s s s u u s p a p - - - - - - - - - e w e w e w c n s a a e o a o e e e s s n s u s s r s t t t t t t - t - t t h - - r - h 1 2 - 1 a e 2 1 1 e l a - a 1 s s t - t - 1 1 Region Indicates that demand is less relative to the supply 16

Time-on-the-Market by VM type us-east-1 Average Time on the Market 150 120 (Days) 90 60 30 0 c4.large c5.large m4.large i3.xlarge c5.4xlarge r5.large r4.large c5d.large m5.2xlarge c5d.9xlarge Instance Type Popular VM types have longest average time on the market 17

Comparing VM Purchasing Options us-east-1 0.1 Effective Price ($/hr) 0.075 0.05 0.025 0 Reserved (1yr) RIM (<1yr) Reserved (3yr) RIM (>=1yr) Spot Ondemand Purchasing Option Users neither lose nor make money on average in the RIM 18

Comparing VM Purchasing Options us-east-1 0.1 Effective Price ($/hr) 0.075 0.05 0.025 0 Reserved (1yr) RIM (<1yr) Reserved (3yr) RIM (>=1yr) Spot Ondemand Purchasing Option RIM listings (>=1yr) price close to spot 19

Conclusion • Provides a first look at Amazon RIM data - Analyzed RIM data from 2018/9 to 2020/5 to reveal key market characteristics - Publicly released data at UMass Trace Repository ‣ http://traces.cs.umass.edu/index.php/Main/Cloud • Future work - More in-depth analysis of RIM data - Use RIM data to mitigate demand risk and optimize long-term cloud costs 20

Email: lambati@umass.edu 21