Hardware Reference
In-Depth Information
AFM has several data management features, such as pre-population of
cache datasets prior to application execution. Further, consistency points
across clusters can be taken through a feature called peer snapshots, which
makes it possible to use AFM as an asynchronous replication engine for dis-
aster recovery purposes.
9.3 Deployment and Usage
GPFS is deployed in thousands of data centers on hundreds of thousands
of nodes. The largest installations contain thousands of nodes in up to sixty
GPFS clusters and store petabytes of data, with many connected to an online
tape system with an even greater capacity. Some of these systems transfer data
at more than 400 GB/s and contain close to a billion files. Further, GPFS is
the key storage technology in several IBM products, including IBM Scale Out
Network Attached Storage (SONAS), DB2 Purescale, and IBM Linear Tape
File System Enterprise Edition (LTFS-EE).
A typical HPC GPFS installation uses the NSD server model, which allows
GPFS to scale beyond shared SCSI storage limitations and offers greater flex-
ibility in the storage architecture. Existing clusters typically use 1 or 10 GigE
networks between GPFS clients and NSD servers, but most newer clusters
are deploying InfiniBand for reduced latency and higher throughput. With
InfiniBand, GPFS clients utilize remote direct memory access (RDMA) to
communicate with the NSD servers and use the native InfiniBand protocol for
inter-client communication. In large deployments, which typically have many
more GPFS clients than NSD servers, it is common to create a separate stor-
age cluster of the NSD servers and one or more GPFS client clusters that use
cross-cluster mounts to access data in the storage cluster. This configuration
simplifies cluster management by allowing for the separate administration of
storage and compute clusters and fine-grained access control across datasets.
GPFS configurations based on the SNC model are currently being deployed
at several sites, either as a general cluster file system or as a scale-out data
analytics solution.
While GPFS AFM was originally designed to facilitate wide-area caching,
users have found new and unexpected use cases. These include scale-out appli-
cation access to legacy NAS devices, asynchronous replication of data to allow
for business continuity in case of a site failure, and construction of a global
namespace among GPFS clusters in one or more data centers, providing global
access to data at local network speeds.
GNR was first deployed on the IBM Power 775 server. Increasingly, more
GNR deployments are appearing with IBM System x GPFS Storage Server
(GSS), which is the first storage appliance that is designed specifically for
GPFS [3].
