Chapter 7 HPSS User Interface Configuration
HPSS Installation Guide September 2002 439
Release 4.5, Revision 2
The bulk of DFS SMT is implemented in the DFS file server, but there is also a user space shared
library that implements all APIs in the XDSM specification. The kernel component maintains
XDSM sessions, XDSM tokens, event queues, and the metadata which describes the events for
whichvarious file systems have registered. The kernel component is also responsible for receiving
events and dispatching them to the DMAP.
The DFS File Server fetches and stores DM attributes, provides persistent managed regions and
events, perform invisible I/O, purges data from files, and verifies file residency. This component
determines if a DMAP has registered to receive a notification for an event related to a particular
operation,then generates the event. If the event is synchronous, it causes the file system operation
towait for a response to the event before proceeding. It also provides for interlocking between DFS
SMT requests and file system calls.
To support persistent DM related metadata, the Episode File System has an extended attribute
facility. DM attributes, event masks, managed regions, and attribute change times (dtime values)
are stored as extended attributes. These extended attributes are treated as file metadata. Changes
to extended attributes are logged.
Episodewas also modified to support files that become sparse by punching holes that release disk
resources.With a conventional sparse file, readingfrom a hole returns zeroes. To assume these same
semantics for a hole that exists because the DMAP migrated the data to tertiary storage would be
incorrect. In this case, the DMAP must retrieve the data from tertiary storage. Hence, Episode
marks these file blocks as being off-line (in tertiary store) instead of as a hole. This allows the file
server to handle partially resident files.
To prevent blocking file server (kernel) threads while waiting for a response to an event, Episode
has a mechanism to notify the client to retry after a specified interval of time.
The DFS fileset dump and restore capability includes extended attributes and migrated regions.
Migrated data is not recalled when a dump is taken, producing an abbreviated dump.
The current XDSM release for Episode does not support cloning. Consequently, some fileset
commands,such as, fts clone,fts move, etc., will not work for XDSM enabled DFS. As a result
fileset backups that rely on cloning must be done withfts dump.
7.6.2.5 XDSM Implementation for XFS
The XDSM implementation supported by SGI for Linux XFS is compliant with the corresponding
standard XDSM specification. In addition, it provides optional features: persistent opaque Data
Management (DM) attributes, persistent event masks, and persistent managed regions.
There is a user space shared library that implements all APIs in the XDSM specification. A kernel
component maintains XDSM sessions, XDSM tokens, event queues, and the metadata which
describes the events for which various file systems have registered. The kernel component is also
responsible for receiving events and dispatching them to the DMAP.
XFS fetches and stores DM attributes, provides persistent managed regions and events, performs
invisible I/O, purges data from files, and verifies file residency. It determines if a DMAP has
registeredto receive a notification for an event related to a particular operation, then generates the
event. If the event is synchronous, it causes the file system operation to wait for a response to the
event before proceeding.