Abstract
At present, all kinds of computer-based systems are growing, and this growth momentum has not diminished. For a large storage cluster, at present, a large storage cluster to store data is likely to reach the ZB magnitude, therefore, there must be a large number of computers or storage devices to cooperate, so as to produce a large storage cluster management and organization problems, then there is a distributed storage system. Distributed storage has the characteristics of strong scalability, good availability and fast running speed, especially Ceph, as an open source distributed data storage platform, can easily expand the memory to PB level. And has a very high access speed, so it has been valued and loved by the industry, in the current release, has replaced the previous Filestore; In addition, Bcache has optimized the local performance, so that the underlying memory engine has undergone qualitative changes, so that the memory performance has been greatly improved. Ceph and other distributed storage systems based on block storage have a big flaw, that is, the quality of service is not high enough to provide users with high quality services, or to provide users with priority. Although a storage system is not required, with quality of service quality of service, it can become a complete, advanced distributed storage system. At present, although CephQoS based on Dmclock is still in the experimental stage, it is not yet fully mature. Based on Bluestore's Ceph, this paper will focus on a new data coding fault-tolerant technology -- local repair code, analyze its performance in detail, and test it in the actual system. Compared with RS erasure code, local repair code can reduce node repair cost at the expense of storage efficiency. Improve node fault tolerance and repair efficiency. '''Key words:''' Bluestore; Distributed storage; Ceph; Erasure code performance optimization
At present, all kinds of computer-based systems are growing, and this growth momentum has not diminished. For a large storage cluster, at present, a large storage cluster to store data is likely to reach the ZB magnitude, therefore, there must be a large number of computers or [...]