As of the 372 release and above (Tornado Back End 1.0.3, Twister
2.1.0, Typhoon 2.1.0), spool objects support raw devices as their
underlying storage mechanism.
Using raw devices has several behaviors that should be noted
-
Filesystems are not used, and therefore the benefits/consequences of
using them are lost. Most notably, most OS'es use free RAM to cache
filesystem pages of frequently accessed files. Raw devices have the
effect of direct I/O and no such caching is done. Since direct I/O is
the preferred method of operation for reader server spool objects (NOT
Cyclone spool objects), this behavior is a win.
-
Spool creation is near instantaneous. Blocks of null bytes are not
written to the spool files as they are when filesystem based spools
are used sine sequential access is guaranteed by definition.
Since every block on the filesystem is not excercised, underlying hardware
errors may lurk after a raw spool has been formatted.
-
Spools have a maximum size of 2TB (7/2005) and only one spool is
allowed per device. With a 20 TB RAID - 10 2 TB LUNs should be
created.
-
Do not make an entire device a raw spool, this can have
unintended side-effects when an OS tries to read the partition
table. We strongly recommend partitions be created and
they be made into raw devices, in the simplest case
a single partition taking up the entire disk is more than adequate.
-
Devices must be labeled with the
rawtool device labelling tool,
included in the /tools directory.
-
Device names are important! rawtool allows for Tornado
Backend/Twister/Typhoon to identify a renumbered device, but be sure
to create names that are unique per Backend in a shared spool
environment.