Deconstructing Amazon EC2 Spot Instance Pricing
read more
Citations
Portfolio-driven Resource Management for Transient Cloud Servers
Price forecasting for spot instances in Cloud computing
A Fairness-Aware Pricing Methodology for Revenue Enhancement in Service Cloud Infrastructure
An index policy for dynamic pricing in cloud computing under price commitments
Managing custom runtimes in an on-demand code execution system
References
Incentives in Teams
Service-Oriented Computing: State of the Art and Research Challenges
Performance Analysis of Cloud Computing Services for Many-Tasks Scientific Computing
Related Papers (5)
Frequently Asked Questions (14)
Q2. What is the effect of a random reserve price?
By creating an impression of false activity (demand and supply changes), the random reserve price can mask times of low demand and price inactivity, thus possibly driving up the provider’s stock.
Q3. Why do the authors ignore the transition period between the second and third epochs?
Due to (1) the gradual move to the new minimal values and to (2) a bug in the pricing mechanism that was fixed in mid-January 2010 [9], the authors choose to disregard data from the transition period between the second and third epochs.
Q4. What are the main reasons for the use of spot price histories?
Optimizing Client Goals Using Spot Price Traces: Andrzejak, Kondo and Yi used spot price histories to advise the client how to minimize monetary costs while meeting an SLA [15], and to schedule checkpoints [16] and migrations [17].
Q5. How many minutes did the client submit a bid?
It is interesting to note that such quiet times can be monetized by clients to gain free computation power with a probability of about 25%, by submitting an instance with a bid of the current spot price 31 minutes after a price change.
Q6. What is the effect of the simulation results?
The simulation results also suggest that Amazon set prices via a market-driven auction with a constant reserve price during the second epoch (December, 2009 until January, 2010), and that prices above the band are market-driven.
Q7. What are the benefits of understanding how Amazon prices its spare capacity?
Understanding how Amazon prices its spare capacity is useful for clients, who can decide how much to bid for instances; for providers, who can learn how to build more profitable systems; and for researchers, who can differentiate between prices set by an artificial process and prices likely to have been set by real client bids.
Q8. What is the common explanation for the dependence on the declared price?
1–3 show that availability strongly depends on declared price for all regions and all instance types, and that this dependency has a typical recurring shape, which can be explained by assuming that Amazon uses the same mechanism to set the price in different regions.
Q9. What is the main epoch of the pricing mechanism?
6. Since the pricing mechanism changes notably and qualitatively between epochs, data regarding these epochs should be separated if an associated statistical analysis is to be sound.
Q10. What is the common way that Amazon sets the price of a product?
The authors conjecture that usually, contrary to impressions conveyed by Amazon [1] and assumptions made by researchers [7], [8], the spot price is set according to a constantly changing reserve price, disregarding client bids.
Q11. What is the availability of different spot instance types?
Fig. 1 shows the availability of different spot instance types as a function of declared price (price-availability graphs), for all examined Windows spot instance types in all regions.
Q12. How many times do the authors show that indiscriminately using Amazon’s current price?
The authors have shown that indiscriminately using Amazon’s current traces to model client behavior is unfounded on average 98% of the time.
Q13. What are the different pricing models for Amazon’s EC2 clients?
Amazon’s EC2 clients rent virtual machines called instances, such that each instance has a type describing its computational resources as follows: m1.small, m1.large and m1.xlarge, respectively denote small, large, and extra large “standard” instances; m2.xlarge, m2.2xlarge, and m2.4xlarge respectively denote extra large, double extra large, and quadruple extra large “high memory” instances;and c1.medium and c1.xlarge respectively denote medium and extra-large “high CPU” instances.
Q14. What is the purpose of the dynamic reserve price?
It also serves to occasionally clear queues of low bids within the band, a purpose that is not served by a constant reserve price that is equal to the ceiling price.