Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider (2003.03423v3)

Abstract: Function as a Service (FaaS) has been gaining popularity as a way to deploy computations to serverless backends in the cloud. This paradigm shifts the complexity of allocating and provisioning resources to the cloud provider, which has to provide the illusion of always-available resources (i.e., fast function invocations without cold starts) at the lowest possible resource cost. Doing so requires the provider to deeply understand the characteristics of the FaaS workload. Unfortunately, there has been little to no public information on these characteristics. Thus, in this paper, we first characterize the entire production FaaS workload of Azure Functions. We show for example that most functions are invoked very infrequently, but there is an 8-order-of-magnitude range of invocation frequencies. Using observations from our characterization, we then propose a practical resource management policy that significantly reduces the number of function coldstarts,while spending fewerresources than state-of-the-practice policies.

• The paper empirically characterizes the complete Azure Functions workload, revealing function invocation frequencies spanning 8 orders of magnitude.
• It introduces a histogram-based, adaptive resource management policy that minimizes cold starts and improves serverless performance.
• Simulation and implementation on Apache OpenWhisk validate its effectiveness, offering actionable insights for cloud providers and researchers.

Characterizing and Optimizing Serverless Workloads

The paper "Serverless in the Wild: Characterizing and Optimizing the Serverless Workload at a Large Cloud Provider" provides an empirical paper of Function as a Service (FaaS) workloads, specifically focusing on Azure Functions. FaaS, a paradigm of serverless computing, enables developers to execute code without explicitly managing infrastructure. The paper addresses the lack of publicly available information on serverless workload characteristics and proposes a resource management policy aimed at optimizing serverless environments.

Key Contributions

1. Workload Characterization: The paper conducts an exhaustive analysis of Azure Functions' entire production workload, presenting insights into function invocation patterns, triggers, and resource consumption. One notable finding is the wide range of function invocation frequencies, spanning 8 orders of magnitude, and the observation that most functions are infrequently invoked. This variability highlights the importance of tailored resource management strategies.
2. Resource Management Policy: Based on the workload characterization, the authors propose a resource management policy designed to minimize cold starts—instances where functions experience additional latency due to slow initialization. The policy uses histograms to adaptively track inter-invocation intervals, allowing for pre-warming of functions based on predicted invocation times. A combination of this histogram-based approach and time-series analysis aims to optimize resource allocation without additional overhead.
3. Implementation and Evaluation: Simulation results demonstrate that the policy significantly reduces cold starts and improves resource utilization compared to static keep-alive strategies used by major providers. The policy was evaluated by implementing it on the Apache OpenWhisk platform, showing a feasible integration and achieving improved performance without notable execution overhead.
4. Data Release: The authors provide sanitized traces of production FaaS workloads, enabling other researchers to develop and evaluate their systems using realistic data.

Implications and Future Directions

The findings have both practical and theoretical implications for the serverless computing model. By offering a detailed characterization of real-world serverless workloads, the paper supports the development of more efficient function scheduling and resource management policies. The proposed adaptive approach could influence how cloud providers design their serverless platforms, balancing performance and cost efficiency.

Possible future directions suggested by this research include exploring more sophisticated machine learning models to predict function invocation patterns and further optimizing the orchestration of serverless applications. Additionally, understanding the impact of function characteristics on execution center workloads could lead to more refined strategies for managing serverless resources.

In conclusion, this paper represents a significant step in analyzing and optimizing serverless workload management, providing actionable insights and methodological advancements for improving the FaaS paradigm in large-scale cloud environments.