TL;DR
This paper introduces an economic model to quantify the value of serverless computing for both providers and customers. It reveals that many customers could already benefit from serverless today, and identifies key areas for research and development to accelerate broader adoption.
What Problem Does This Paper Address?
The rise of cloud computing has shifted infrastructure management from on-premises to large-scale shared resources. Now, a new paradigm, "serverless," is emerging, promising to abstract away resource provisioning and management, with a pay-as-you-go model. However, it's unclear whether serverless will truly replace traditional "serverful" computing or remain a niche offering. This paper aims to address this by quantitatively analyzing the economic incentives that would drive adoption for both cloud providers and customers, moving beyond qualitative hypotheses.
Methods
The core of the paper is an economic model that defines conditions for mutual benefit between providers and customers regarding serverless adoption.
- Provider Condition: A rational provider will offer serverless products if the revenue from serverless functions (price per function * utilization ratio) exceeds the opportunity cost of renting out the resource as a virtual machine (VM rental price * time).
- Customer Condition: A rational customer will prefer serverless if the net value derived from serverless execution (considering performance, predictability, and transition costs) is worth the cost, relative to serverful execution. This is expressed as a ratio involving the price difference between serverless and serverful, and the customer's relative efficiency and net value gain from serverless.
The paper then categorizes customers into three groups (A, B, C) based on their "serverless preference" ($\alpha$), which is derived from the customer condition. It also analyzes the value of autoscaling, a key serverless feature, by simulating different workload types against various provisioning strategies.
Datasets
The paper uses workload statistics from Microsoft Azure Functions [16] to generate synthetic workloads for their experiments. These statistics include distributions of function execution durations and interarrival times (mean and coefficient of variation).
Key Insights and Ideas
- Serverless is Already Viable for Many: The economic model suggests that a significant portion of customers could already benefit from serverless today, even with current pricing structures. This is particularly true for customers with high utilization inefficiencies in their serverful setups.
- Customer Segmentation is Key: Customers can be broadly categorized into three groups based on their inherent preference for serverless. Group A is already incentivized, Group B would prefer serverless but providers aren't yet able to offer it profitably, and Group C would need improvements in serverless value (vs) to switch.
- Autoscaling's Broad Appeal: Surprisingly, serverless autoscaling can provide value even for workloads that are not extremely "bursty." This is due to factors like sparse utilization from periodic arrivals and the inherent inefficiencies of static peak provisioning.
- Research Directions are Clear: The paper highlights specific areas where research and development would have the most market impact, such as improving specialized hardware support for serverless (benefiting Group C), reducing data movement costs, and addressing cold starts and idle times (benefiting providers and enabling Group B).
- QoS Guarantees are Crucial: For customers with strict Quality of Service (QoS) requirements, simple serverless offerings might not suffice. Introducing features like guaranteed warm starts could unlock significant value and adoption.
Key Results
- Provider Condition: A provider should offer serverless if $c > \frac{p_r}{p_s}$, where $c$ is the utilization ratio of a VM executing serverless functions, $p_r$ is the VM rental price, and $p_s$ is the serverless function price.
- Customer Condition: A customer prefers serverless if $\frac{p_s}{p_r} < \alpha$, where $\alpha = \frac{t_r}{t_s} \cdot \frac{v_s}{v_r}$. Here, $t_r$ is serverful execution time, $t_s$ is serverless execution time, $v_s$ is net value from serverless, and $v_r$ is net value from serverful.
- Market Segmentation: The paper conceptually partitions the market into three groups based on $\alpha$ relative to $\frac{p_r}{p_s}$ and 1.
- Autoscaling Value: Simulations show that for bursty workloads, serverless is preferred after provisioning ~25% of peak resources. For random workloads, this breakeven point is around 50%. Even predictable workloads show some preference due to sparse utilization.
- QoS Impact: Achieving p95 queuing latency of 10ms might require $\alpha > 100$, and p99 latency might require $\alpha \approx 1000$. This indicates that serverless offerings with QoS guarantees could be highly valuable.
Limitations and Future Work
- Provider Infrastructure Black Box: The model assumes variable operational costs for serverless are not significantly different from serverful, which might not hold true in practice (e.g., electrical power for higher core utilization).
- Resource Heterogeneity: The model treats different VM configurations and QoS parameters as distinct products but doesn't deeply explore the economic implications of heterogeneity.
- Future Work: The authors suggest exploring time-varying pricing for serverless to flatten load patterns, similar to public utilities. They also emphasize the importance of developing serverless offerings for specific high-value use cases (like GPUs or data science) to unlock adoption for "locked-out" customers. Further research into QoS guarantees for serverless is also highlighted.
Summary updated