Faculdade de Ciências e Tecnologia (FCT) >
FCT Departamentos >
FCT: Departamento de Informática >
FCT: DI - MA Dissertations >
Please use this identifier to cite or link to this item:
|Title: ||A shared-disk parallel cluster file system|
|Authors: ||Lopes, Paulo Orlando Reis Afonso|
|Advisor: ||Medeiros, José|
|Issue Date: ||2009|
|Publisher: ||FCT - UNL|
|Abstract: ||Today, clusters are the de facto cost effective platform both for high performance
computing (HPC) as well as IT environments. HPC and IT are quite different environments
and differences include, among others, their choices on file systems and storage: HPC favours parallel file systems geared towards maximum I/O bandwidth, but which are not fully POSIX-compliant and were devised to run on top of (fault prone) partitioned storage; conversely, IT data centres favour both external disk arrays (to provide highly available storage) and POSIX compliant file systems, (either general purpose or shared-disk cluster file systems, CFSs).
These specialised file systems do perform very well in their target environments provided that applications do not require some lateral features, e.g., no file locking on parallel file systems, and no high performance writes over cluster-wide shared files on CFSs. In brief, we can say
that none of the above approaches solves the problem of providing high levels of reliability and performance to both worlds.
Our pCFS proposal makes a contribution to change this situation: the rationale is to take advantage on the best of both – the reliability of cluster file systems and the high performance of parallel file systems. We don’t claim to provide the absolute best of each, but we aim at full POSIX compliance, a rich feature set, and levels of reliability and performance good enough
for broad usage – e.g., traditional as well as HPC applications, support of clustered DBMS engines that may run over regular files, and video streaming. pCFS’ main ideas include:
· Cooperative caching, a technique that has been used in file systems for distributed disks but, as far as we know, was never used either in SAN based cluster file systems or in parallel file systems. As a result, pCFS may use all infrastructures (LAN and SAN) to move data.
· Fine-grain locking, whereby processes running across distinct nodes may define nonoverlapping byte-range regions in a file (instead of the whole file) and access them in parallel, reading and writing over those regions at the infrastructure’s full speed (provided that no major metadata changes are required).
A prototype was built on top of GFS (a Red Hat shared disk CFS): GFS’ kernel code was
slightly modified, and two kernel modules and a user-level daemon were added. In the
prototype, fine grain locking is fully implemented and a cluster-wide coherent cache is maintained through data (page fragments) movement over the LAN.
Our benchmarks for non-overlapping writers over a single file shared among processes
running on different nodes show that pCFS’ bandwidth is 2 times greater than NFS’ while
being comparable to that of the Parallel Virtual File System (PVFS), both requiring about 10 times more CPU. And pCFS’ bandwidth also surpasses GFS’ (600 times for small record sizes, e.g., 4 KB, decreasing down to 2 times for large record sizes, e.g., 4 MB), at about the same CPU usage.|
|Description: ||Dissertação apresentada para obtenção do Grau de Doutor em Informática Pela Universidade Nova de Lisboa, Faculdade de Ciências e Tecnologia|
|Appears in Collections:||FCT: DI - MA Dissertations|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.