MAPLE - Microbenchmark-based Architecture Performance Learning and Estimation

Description: 

As cloud computing continues to dominate modern software deployment, organizations face increasing challenges in selecting and evaluating appropriate hardware platforms for their data-intensive workloads. Low-level microbenchmarks are commonly used in this context because they provide quick, standardized measurements of individual hardware components and microarchitectural characteristics without requiring full application deployment. These benchmarks help assess fundamental system capabilities such as memory throughput, cache performance, and instruction execution efficiency.

However, there is a significant gap between these low-level microbenchmark results and actual application-level performance, making it difficult to predict how cloud data applications will perform on unseen hardware configurations. Traditional benchmarking approaches often require extensive testing on each target platform, which is time-consuming, costly, and impractical given the rapidly evolving landscape of cloud infrastructure offerings.

To address these challenges, we propose MAPLE, a research initiative that explores the relationship between microbenchmark performance and application-level performance in cloud data environments. This project brings together Snowflake's industry-leading expertise in cloud data warehousing and real-world workload characteristics with our chair's academic and practical expertise in performance modeling and evaluation. Together, we aim to develop methodologies that can better predict application behavior across different hardware platforms using microbenchmark data.

This approach enables organizations to make more informed infrastructure decisions without requiring exhaustive testing on every potential platform. Through this collaboration, MAPLE aims to establish new methodologies for cloud performance evaluation that bridge the gap between micro-level measurements and macro-level application behavior, ultimately benefiting both the research community and industry practitioners in the cloud computing ecosystem.

People involved: Samuel Kounev, Yannik Lubas

Project start: 2026

Project end:  2027