 |
Configuring LSM for maximum performance depends on
the volume type and the underlying hardware.
The following are some general guidelines to keep
in mind when creating an LSM volume.
- Choose the type of volume based on application
requirements and available hardware. Note the tradeoffs
between performance, price and availability.
- For LSM mirrored volumes, mirror across disk controllers
for high availability. If each mirror of a mirrored
volume is striped, then stripe down controllers.
- If you are configuring Dirty-Region-Logging (DRL)
for mirrored volumes to improve mirrored volume recovery
time after a system crash, avoid creating DRLs on
the same disks that store volume data.
- If the application typically issues multiple simultaneous
I/Os to an LSM striped volume (mirrored or not),
avoid splitting individual I/Os. Choose a stripe
size that equals the application I/O transfer size,
thus balancing I/O across the disks in the stripe
set. Otherwise, split individual large I/Os across
the physical disks in the stripe set using a multiple
of the number of disks in the stripe set. Note that
LSMs default stripe width, 64k, works best for most
environments. It is not uncommon for a workload to
change or not be easily understood, so do not over-optimize
for such applications.
- If the application typically issues a large number
of small I/Os to an LSM striped, non-mirrored volume
on a system with multiple CPUs (GS1280, for example),
use the LSM sysconfig variable, Max_LSM_IO_PERFORMANCE,
to reduce the need for certain spinlocks, thus further
improving performance. Note that this sysconfig variable
is available in V5.1B of Tru64 UNIX.
- For LSM RAID5 volumes, choose a stripe size that
results in a full-stripe write, avoiding a read-modify-write
cycle. The default stripe size, 16k, works best for
most environments.
- If you are using LSM along with hardware RAID,
avoid striping at both the LSM and hardware RAID
level unless you configured a large stripe width
for LSM over multiple hardware RAID units. So, for
a performance benefit, the LSM volume should be striped
over multiple RAID sets on different controllers
with the LSM stripe size set to a multiple of the
full hardware RAID stripe size. The following diagram
illustrates this configuration.
Otherwise, use LSM to mirror across hardware RAID
units for high-availability.
- If in a single-system environment, use LSMs volstat
command to help benchmark performance. If in a TruCluster
environment, run volstat from the cluster member
driving I/O.
LSM maintains 4-8 copies of a configuration database which describes the LSM
configuration. These copies should be spread across controllers for high
availability. LSM cannot run without loading its configuration database.
The following are some general guidelines to keep
in mind when initializing LSM.
- Ensure configuration database copies are spread
across controllers. In the V5.X stream of Tru64 UNIX,
LSM will do this for you for storage that is not
fibre-connected. For fibre-connected storage, manually
spread LSM configuration databases across controllers
within the fabric for high availability.
- Disks under LSM control have a private region
which in part stores the LSM configuration database.
In the V5.X stream of Tru64 UNIX, the default private
region size was increased from 1024 sectors to 4096
sectors. If creating a small LSM configuration (generally
fewer than 10 physical disks under LSM control),
you may want to consider using a smaller private
region size, as it will affect the time it takes
to startup LSM and the time it takes to make configuration
changes. While we're talking a matter of seconds,
if every second counts and the configuration is tight,
a smaller private region size is the right choice.
- When physical disks are added to LSM, they are
organized into diskgroups. In a TruCluster environment,
keep LSM diskgroups on shared storage as much as
possible, avoiding hybrid diskgroups where more than
one cluster member must be available to access those
volumes. Hybrid disk groups provide a host of problems
in light of cluster members leaving and joining the
cluster as configuration changes are occurring. The
following diagrams illustrate these configurations.
LSM Highly Available Disk Group
LSM Hybrid Diskgroup (avoid!)
|
