Cluster Storage Capacity Increased by One Petabyte
Hardware
The following systems have been installed into a single 19 inch standard rack with redundant power supplies and rear-door water cooling.
Disk storage servers:
- 3 x IBM System Storage DCS3700 each with:
- 8 GB Cache
- 8 x 8Gbps Fibre Channel Ports
- 2 x DCS3700 Expansion Units
- 180 x 2 TB SAS disk drives.
Two additional IBM GPFS network shared disk servers:
- 2 x IBM System x3650 each with:
- 2 x Intel Xeon E5640 CPU, 4 Cores, 2.66GHz
- 12 GB RAM
- 1x QDR Infiniband Adapter, 2 Ports
- 3x 8Gbps Fibre Channel Adaper, 2 Ports
- 2x 146 GB 15k SAS drives
- 1x RAID Adapter
Two of the DCS3700 storage systems have been directly attached to the new IBM GPFS servers while one storage system has been attached to the SAN in order to provide cache capacity for TSM backups.
Deployment & Use
Setup of disk arrays was straightforward using the IBM storage manager. For each storage device 16 RAID-6 arrays with a net capacity of about 15 TByte each have been created. Arrays from two of the storage systems have been already mapped directly to SCSI disk devices on the new GPFS network shared disk servers and added to the existing parallel filesystems as a secondary storage pool.
Total capacity available for scientific users on the cluster:
- 250 TB in the scratch filesystem.
- 500 TB in the iplex filesystem.
New user quotas (soft/hard):
- 8/10 TByte per account in the scratch filesystem
- 20/25 TByte per project group in the iplex filesystem.
A policy has been tested In order to manage the placement of files on one of the two storage pools available. This policy uses a weight function to place files which are either very large or which have not been accessed for a long time onto the new secondary storage pool.
Outlook
After this significant upgrade of disk storage capacity it is required to completely redesign the backup and archive functions run against the parallel filesystems.
An upgrade of storage capacity for HOME filesystems is also already planned to be installed later in 2012.
