Research On On-line Scaling Scheme For Erasure-coded Storage Clusters

As the amount of data in the storage system continues to grow,the storage system will face the challenge of insufficient storage capacity,the usual response is to add new storage devices to storage system,namely storage scaling.For erasure-coded storage cluster,storage scaling involves two operations: data migration and parity update.In data center and other scenarios,the storage system needs to provide all-weather data storage services,that is,the storage system can not stop storage service when the storage scaling is in progress.In this study,we research on on-line scaling scheme for erasure-coded storage clusters.In this paper,we propose a piggybacked on-line scaling scheme,called PiggySS(Piggybacked Scaling Scheme),which uses the existing user request data to reduce the scaling I/Os PiggySS has three salient features.First PiggySS alleviates the I/O interference of scaling operations to user requests by exploiting the existence of requested user data.Second,PiggySS accomplishes balanced I/Os during the on-line scaling procedure by distributing data-block reads/writes and parity-block reads/writes among all involved nodes with taking node workload as well as scaling I/Os into account.Third,PiggySS achieves high read parallelism by migrating a portion of requested data blocks to newly-added nodes,and provides high write performance by applying full-stripe data filling to scaled chunk groups,thus improving the I/O performance of scaled storage clusters.Apart from four algorithms used to handle both scaling and user I/Os,we also present three optimization approaches for the PiggySS scheme.We implement PiggySS along with three alternative scaling schemes in a Reed-Solomon-coded storage cluster,on which PiggySS and other three scaling schemes(i.e.,Scale-RS,McPod,RR)quantitatively evaluated by replaying real-world I/O traces.Experimental results demonstrate that PiggySS outperforms the other scaling schemes in terms of user response time,the average user response time of PiggySS is 33.5%,29.92%,19.34% of Scale-RS,McPod,RR during the scaling process,and 87.09%,94.53%,85.87% after scaling,respectively.