Programmable Storage Workshop: Managing Bufferbloat in Storage Systems, Esmaeil Mirvakili

176 Views
Published
Managing Bufferbloat in Storage Systems
Esmaeil Mirvakili (UCSC)
Abstract: In most systems, the Quality of Service is usually achieved by queueing and scheduling the incoming requests based on their priority and costs. However, storage systems have two levels of queuing, a frontend and a backend. The requests are queued and scheduled in the frontend and submitted to the backend queue. The backend can be different kinds of devices and storage systems, and scheduling the backend requests is not always possible. This design poses a new challenge. The new incoming requests with high priority cannot be scheduled with the requests which are already submitted to the backend. As a result, if a large number of requests are submitted to the backend, the scheduling algorithm in the frontend cannot guarantee the QoS as intended. On the other hand, if a small number of requests are sent to the backend, we cannot utilize the backend at the full capacity (in terms of throughput and performance). The solution to this problem is to batch a reasonable number of requests, so we can maximize the scheduling ability with a minimum impact on the throughput. In this presentation, we propose a design based on a mitigation of the bufferbloat problem in networks. The use of bufferbloat mitigation algorithms like Controlled Delay (CoDel) in the backend as a backpressure mechanism can help us adapt the request batch size based on workload and QoS assumptions. We have implemented a CoDel based prototype in Ceph with BlueStore as the backend system to validate our solution. Based on the preliminary results, it seems that this solution can effectively minimize the batch size with little impact on throughput.

Bio: Esmaeil Mirvakili is a 2nd-year computer science Ph.D. student at UCSC working with Carlos Maltzahn at the Systems Research Lab. His focus is on Storage and distributed systems. He received his master degree in Software Engineering at the Sharif University of Technology.
Category
Network Storage
Be the first to comment