16 Implementing the IBM General Parallel File System (GPFS) in a Cross-Platform Environment
Figure 2-7 shows the SVC architecture and the physical and logical views.
Figure 2-7 IBM SVC physical and logical view
Be sure to use zoning to limit the number of paths to four. The hosts must run a multipathing
device driver to show a single device. The multipathing driver supported and delivered by
SVC is the IBM Subsystem Device Driver (SDD). Native MPIO drivers on selected hosts are
2.2.3 The characteristic of storage hardware
This section provides information about the hardware storage functions.
RAID level
Select a RAID protection level for your array. For most situations, you need a degree of
availability protection, which eliminates RAID 0. You also have to select the best RAID level
according to your applications and business needs. To help you with the selection, use
Table 2-1 on page 17 as a guideline for selecting RAID protection for GPFS.
Mdisk Mdisk
Mdisk group
LUNs from
IBM SVC Logical View
IBM SVC Physical View
IBM storage
EMC storage
IBM SVC Cluster
SAN Fabric
Heterogeneous Storages
Notes: For OS-specific information about MPIO support, go to the following address:
To download Subsystem Device Driver Device Specific Module (SDDDSM) information,
see the following address:
Chapter 2. Infrastructure planning 17
Table 2-1 Guidelines for selecting RAID protection for GPFS
Cache policy
This cache policy is an area of tuning where there are significant differences based on the
style of storage server and the I/O workload. Generally, consider the following guidelines:
Sequential I/O workloads: Enable read cache, disable any read prefetch
Random I/O workloads: Enable read and write cache
Metadata Disks: Enable read and write cache
It may be more apparent why you would enable cache for random workloads but not so clear
why you disable read cache on sequential workloads.
With GPFS, disable any read prefetch mechanisms for the caching of data in the storage
server. Although it works well for high concurrency environments and is high performance,
read-prefetch at the storage must be disabled because the GPFS mechanism for allocating
space is difficult for storage algorithms to “predict.” GPFS can compensate by prefetching
data. To a storage controller, this read access does not look sequential, so the read-ahead
that is provided in most storage servers makes incorrect assumptions, therefore actually
degrading read-performance. In fact, up to a 20% penalty can occur on certain models of
The following DS4000/DS5000 cache settings are supported:
Read cache enabled or disabled
Write cache enabled or disabled
The cache mirroring is enabled or disabled (depending upon the write cache mirroring
setting). The performance benefits of the read or the write caching depend on the application.
Because the cache settings can be easily adjusted from the Storage Manager (SM),
preferably carry out your performance tests during the implementation. Use the following
sample configuration as a starting point for the NSD servers or other GPFS nodes that are
directly attached to a SAN over a Fibre Channel network; use the following cache settings:
Read cache = enabled
Read ahead multiplier = 0
Write cache = disabled
Write cache mirroring = disabled
Write block size = 16 KB
Description RAID 5 RAID 6 RAID 10 Advantage
Random write performance 2 3 1 RAID 10
Sequential write performance 1 2 3 RAID 5
Availability 3 1 2 RAID 6
Space efficiency 1 2 3 RAID 5
RAID level for GPFS Yes Yes No RAID 5 and 6
Configuration for GFPS 4+P 8+2P N/A N/A
Minimum number of disks per LUNs 5 10 8 N/A
Note: In the event of a power failure or hard controller failure, a risk of file system
corruption exists when enabling write cache.

Get Implementing the IBM General Parallel File System (GPFS) in a Cross Platform Environment now with O’Reilly online learning.

O’Reilly members experience live online training, plus books, videos, and digital content from 200+ publishers.