In this paper, we studied the problem of providing coarse-grained fairness to multiple hosts sharing a single storage system in a distributed manner. We propose a novel software system, PARDA, which uses average latency as an indicator for array overload and adjusts per-host issue queue lengths in a decentralized manner using flow control.
Our evaluation of PARDA in a hypervisor shows that it is able to provide fair access to the array queue, control overall latency close to a threshold parameter and provide high throughput in most cases. Moreover, combined with a local scheduler, PARDA is able to provide end-to-end prioritization of VM IOs, even in presence of variable workloads.
As future work, we are trying to integrate soft limits and reservations to provide a complete IO management framework. We would also like to investigate applications of PARDA to other non-storage systems where resource management must be implemented in a distributed fashion.
