Database Reference
In-Depth Information
Because GPFS is often used in enterprise settings that have limited concur-
rent write access to shared files, GPFS incorporates a special optimization to
reduce lock acquisition overheads for single writers. Specifically, when the to-
ken manager sees a single writer accessing a file, it grants a lock to that writer
that covers the entire file, allowing that writer to operate on the file without
further interactions with the token manager. However, if another writer ap-
pears for that file, the lock is revised so that it covers only the region before, or
the region after, the region requested by the new client (depending on where it
requested its lock originally). This process of breaking up locks is continued if
more writers appear, allowing greater concurrency at the expense of increased
lock trac.
Because GPFS has a robust locking system, GPFS is able to extensively
cache both data and metadata on clients. Caching is performed by the GPFS
page pool. Data can be prefetched and held in page pool buffers on the client,
and writes can be buffered and written after control is returned to the client
process. Likewise, directory contents may be cached to speed directory listings.
These techniques, like lock optimization, are particularly helpful in environ-
ments where sharing is limited.
2.2.2 Object-Based Storage Systems
More recently, parallel file systems using
object-based
storage have emerged.
Object-based storage devices differ from traditional block storage devices in
that they present their storage as a collection of “objects” rather than a col-
lection of blocks. These objects can store data of variable size, as well as a
set of attributes on that data. Object-based storage devices also provide func-
tionality for searching based on the attributes of objects and for grouping
objects into collections. Overall these features shift the burden of local stor-
age management away from the file system and into the device, allowing file
system implementers to focus on other design challenges. Unfortunately, while
the interfaces for object-based storage devices have been specified, few prod-
ucts currently provide these capabilities. For this reason, parallel file system
designs relying on object-based storage capabilities provide these capabilities
by layering software or firmware on top of traditional block storage devices
(disks).
2.2.2.1
PanFS
Panasas provides a distributed, parallel file system and highly available,
clustered metadata and storage server hardware to large-scale, Linux-based
compute clusters used in government, academic, and commercial high-
performance computing.
5
-
7
Panasas is the primary storage used by the world's
first petaFLOP computer, Los Alamos National Laboratory's Roadrunner.
Panasas systems are constructed of one or more shelves, shown in Figure 2.3.
Each shelf is a bladeserver with 11 slots that can contain storage servers,
