November 08, 2015
Jayanta Basak, Madhumita Bharde
Due to lack of generic, accurate, dynamic and comprehensive models for performance estimation, customers typically tend to under- provision or over-provision the storage systems today. With multi-tenancy, virtualization, scale and unified storage becoming the norm in the industry, it is highly desirable to strike an optimum balance for utilization and performance. However, performance prediction for enterprise storage systems is a tricky problem, consider- ing that there are multiple hardware and soft- ware layers cascaded in a complex way that affect the behavior of the system. Configuration factors such as CPU, cache size, RAM size, capacity, storage backend (HDD/Flash) and net- work cards etc. are known to have a significant effect on number of IOPS that can be pushed to the system. However, apart from system characteristics as these, storage workloads vary reason- ably and therefore IOPs numbers depend heavily on the type of workloads provisioned on storage systems. In this work, we treat storage system as a black-box and propose a solution that will equip admin make provisioning decisions based on knowledge of workloads.
The definitive version of the paper can be found at: https://lisa15.usenix.hotcrp.com/doc/lisa15-final10.pdf.