Modeling and Prediction of I/O Performance in Virtualized Environments

Modern, future-oriented data centers increasingly rely on virtualization technology to host their services and applications efficiently and flexibly by sharing the resources and allocating them on-demand. The dramatically increasing amount of data generated and stored by today’s applications, however, poses significant challenges for the data center operators to respect Service-Level Agreements (SLAs) and guarantee adequate performance for their users. To cope with such challenges, storage resources in today’s data centers have evolved from simple disk-based arrays to sophisticated tiered systems with complex caching and optimization strategies. Still, the storage resources are usually highly underutilized and overprovisioned to avoid bottlenecks along the infrastructure’s data I/O path, thus leaving highly expensive resources lie waste. Performance modeling techniques in the area of performance engineering can usually support to anticipate I/O performance bottlenecks and employ the storage resources more reasonably. For a practical applicability, however, these techniques need to be refined and tailored to capture the I/O performance in modern virtualized environments. The increasing complexity of the I/O infrastructure is a major challenge for practical performance engineering approaches. Existing approaches usually address this issue at a low modeling abstraction level and with specific focus on certain application scenarios such that important capacity planning questions frequently remain unanswered. In this thesis, we present a novel performance modeling approach tailored to I/O performance prediction in virtualized environments. The main idea is to identify important performance-influencing factors and model them at a practical abstraction level. Capturing these factors, we develop tailored storage-level I/O performance models using complementary modeling formalisms based on statistical regression analysis and queueing theory. To increase the practical applicability of these models, we combine the low-level