Optimization of Distributed Storage on Commodity SSD using NVDIMM

Abstract

As cloud computing has been dominant, people become interested in distributed storage system used in data center. One of the most attractive things is ceph distributed storage. Thats because it has various advantages like high scalability, fault tolerance and rebalancing data objects automatically. Client can communicate the node which has the target data object directly, without extra communication with metadata server. Ceph distributed storage is said to be petabyte scalable for these reasons. In addition, the replication functionality of ceph ensures high availability, fault tolerance and no single point of failure. Thats why ceph has been selected as our target system. Even if ceph has various advantages like scalability, it has not been designed for the usage of NVDIMM with commodity SSD. So, the research focuses on how the ceph distributed storage has to be optimized for NVDIMM with commodity SSD. First, how the internal implementation of ceph has to be optimized for using NVDIMM as write buffer will be explained. This optimization has been implemented by making worker threads issue I/O request to NVDIMM and reply to client. This optimization improves the performance of write request. Second, how the mixed workload, which contains read and write request, should be dealt with will be explained. The above-mentioned optimization for write request cant improve the performance of mixed workload. The reason is that read requests can race with flush command. To remove the race, separation of read request and flush command should be used.