IBM General Parallel File System for AIX: Administration and Programming Reference

Changes for GPFS 1.3

GPFS 1.3 provides several performance, scalability and usability enhancements:

• The term Stripe Group Manager has been replaced by the term File System Manager.
• The term GPFS configuration has been replaced by the term GPFS nodeset.
• The term stripe group panic has been replaced by the term file system forced unmount.
• Possible maximum values for file size and file system size have been relaxed. The introduction of multiple levels of indirection allows file sizes up to the largest supported GPFS file system size. The supported value for maximum file system size has increased to 9TB.

  With the use of multiple levels of indirection, GPFS now internally controls the values for indirect block size and i-node size. System control of these values allows for more effective caching of i-nodes and may improve the performance of some applications. The -I IndirectSize and -i InodeSize options have been removed from the mmcrfs command.

• GPFS exploits the Low-Level Application Programming Interface (LAPI) function of the Parallel System Support Program (PSSP). Using LAPI instead of Transmission Control Protocol/Internet Protocol (TCP/IP) should improve the performance of communication between the GPFS daemons.

  However, it is important you understand what is required when choosing LAPI as the communications protocol. A switch adapter window must be reserved on each node in the GPFS nodeset. To successfully reserve an adapter window for use by GPFS, all LoadLeveler^(R) jobs must be stopped prior to issuing either the mmconfig command or the mmchconfig command. If LoadLeveler is not stopped, it will reserve all adapter windows for its use. If an adapter window cannot be reserved on each node for use by GPFS, the GPFS commands will fail.

• GPFS exploits the SP Security Services Library routines.
• The GPFS File System Manager may be migrated to a specific node.
• GPFS file systems may now be exported via the Distributed File Service (DFS) protocol.
• GPFS file systems can now be managed and monitored by using the Data Management Application Programming Interface (DMAPI) for GPFS. The interface is compliant with The Open Group's XDSM standard (System Management: Data Storage Management (XDSM) API).

  Note:

  DFS^(TM) export of a DMAPI enabled file system is not supported in GPFS 1.3.

• A series of interfaces which will allow the Message Passing Interface (MPI) layer of an application to specify characteristics of its data access and allow GPFS to optimize its access to the data.
• The mmcrvsd command is now available to create virtual shared disks for use by GPFS. It is suggested that this command be used to create virtual shared disks for the mmcrfs, mmadddisk, mmchdisk, and the mmrpldisk commands.
• The mmdefragfs command is now available to reduce disk fragmentation thereby potentially increasing the number of full free blocks available to the file system.
• The mmshutdown command is now available to cleanly unmount all file system and shutdown GPFS either on a single node or across a set of nodes.
• The mmchdisk command has been enhanced to allow all disks belonging to a file system to be resumed or started at the same time.
• The mmadddisk, mmchdisk, mmdeldisk, mmrestripefs, and mmrpldisk commands have been enhanced to allow the user to specify which nodes of a GPFS nodeset will participate in the restripe of a file system.
• Additional block sizes of 512K and 1024K are supported. These larger block size values may allow you to improve performance by matching RAID block sizes.
• GPFS programming interfaces now support both 32-bit and 64-bit applications.
• An external programming call, which provides exact status for a file, has been added to GPFS. Exact status for a file using the file descriptor (gpfs_fstat( )) or the file system name (gpfs_stat( ) ), can now be obtained.
• The mmaddnode, mmchconfig, mmconfig, mmcrfs, mmdelnode, and mmlsnode commands are now executable from any node running GPFS or the control workstation. Each of these GPFS command now has a -C NodesetId option which allows you to designate which GPFS nodeset the command is to act upon. If this option is not specified when issued from a node, the command will act upon the nodeset to which the issuing node belongs. This option must be specified when the command is issued from the control workstation, or it will fail.
• Because file system names must be unique across nodesets. the mmadddisk, mmchdisk, mmchfs, mmchmgr, mmdefragfs, mmdeldisk, mmdelfs, mmdf, mmfsck, mmlsdisk, mmlsfs, mmlsmgr, mmrestripefs, and mmrpldisk commands are executable from any node running GPFS or the control workstation (CWS).
• The mmcheckquota command has been expanded to allow replacement of the user and group quota files for a file system.
• Support for the following memory mapped file capabilities as described in the X/OPEN 4.2 standard:
  1. mmap
  2. munmap
  3. msync
• Support for the shmat memory mapped file capability as described in AIX.
• Support for the fsize user process resource limit, as defined by ulimit.