Tuning HP Tru64 UNIX V5.1A and V5.1B for Oracle

used by the database for its System Global Area (SGA) or any other application using shmget()

or nshmget() to allocate shared memory.

Using granularity hints memory offers the best performance and is ideally suited for large GS class

AlphaServer systems, especially those with applications that have many processes attached to a

large shared memory region (such as large Oracle databases). On smaller systems the overall

performance gain using granularity hints memory averages 7%; on those systems this may not be

an interesting option, due to its complexity in implementation and management and its relatively

small performance benefit. HP recommends this type of shared memory (statically allocated

granularity hints memory) for the most demanding systems using large shared memory segments.

It is important to ensure that all shared memory segments are allocated in the memory area reserved

for granularity hints memory. When the shared memory segments cannot be allocated in the

reserved memory, then messages will be displayed on the console and through syslog (for

example, /var/adm/messages). System performance will be negatively impacted, see the

ssm_threshold parameter for more information. You can prevent overallocation of this memory

area by setting the parameter gh_fail_if_no_mem = 1.

ipc:ssm_threshold must be disabled by setting ssm_threshold = 0 when using statically

allocated granularity hints memory (rad_gh_regions or gh_chunks) is used.

On those systems that are not NUMA aware, that is, all systems that only have a single RAD, you

can use the gh_chunks attribute to enable granularity hints memory. The gh_chunks parameter

is specified in multiples of the largest possible granularity hints page, that is, it specifies the number

of 4-MB pages to allocate for RAD 0. Both rad_gh_regions and gh_chunks are equally

effective. This document only describes rad_gh_regions; however, you can use gh_chunks on

single RAD systems as an alternative. To avoid confusion, HP recommends using the

rad_gh_regions parameter rather than the gh_chunks parameter.

To view the amount of granularity hints memory configured on a system, use the vmstat –P

command.

On the EV6x NUMA systems (AlphaServer GS80, GS160, and GS320) a RAD corresponds to a

Quad Building Block (QBB). On EV7 CPU based systems (AlphaServer ES47, ES80, and GS1280)

a RAD typically corresponds to a single CPU (unless it is changed through the

generic:cpus_in_rad parameter). The amount of memory per RAD to statically allocate for

granularity hints memory is specified in 1-MB increments through the corresponding

rad_gh_regions parameter.

The naming convention for this tunable parameter is to specify rad_gh_regions[<RAD

number>]. For example, to allocate 2 GB in RAD 0 you would specify 2048 for

rad_gh_regions[0], that is, rad_gh_regions[0] = 2048.

Take your planned/projected Oracle SGA size and divide by the number of RADs your system has

configured. For each rad_gh_regions[x] (where x represents the RAD identifier, from 0-63,

depending on the number of RADs in your system), you would specify the required value in

megabytes (MB). Note that the number of the RAD is determined by the physical location in the

system, that is, unpopulated RADs are still assigned a number.