Storage Resource Sharing with CASTOR

The Cern Advanced STORage (CASTOR) system is a hierarchical storage management system developed at CERN to meet the requirements for high energy physics applications. The existing disk cache management subsystem in CASTOR, the stager, was developed more than a decade ago and was intended for relatively moderate (large at the time) sized disk caches and request load. Due to internal limitations a single CASTOR stager instance will not be able to efficiently manage distributed disk cashes of several PetaBytes foreseen for the experiments at the Large Hadron Collider (LHC) which will be commissioned in 2007. The Mass Storage challenge comes not only from the sheer data volume and rates but also from the expected request load in terms of number of file opens per second. This paper presents the architecture design for a new CASTOR stager now being developed to address the LHC requirements and overcome the limitations with the current CASTOR stager. Efficient management of PetaByte disk caches made up of clusters of 100s of commodity file servers (e.g. linux PCs) resembles in many aspects the CPU cluster management, for which sophisticated batch scheduling systems have been available since more than a decade. Rather than reinventing scheduling and resource sharing algorithms and apply them to disk storage resources, the new CASTOR stager design aims to leverage some of the resource management concepts from existing CPU batch scheduling systems. This has led to a pluggable framework design, where the scheduling task itself has been externalized allowing the reuse of commercial or open source schedulers. The development of the new CASTOR stager also incorporates new strategies for data migration and recall between disk and tape media where the resource allocation takes place just-in-time for the data transfer. This allows for choosing the best disk and network resources based on current load.