HP-UX 11i v3 Memory Management Subsystem

19
Appendix 1 -- Variations with OS release
Memory management enhancements since the initial HP-UX 11i v3
release
• Floating cells were first supported in the September 2007 Update.
• The Kernel Access Infrastructure feature to allow the mapping of user virtual addresses into the
kernel was introduced in the March 2008 Update.
• The numa_policy kernel tunable parameter to control the default placement of application
memory objects was introduced in the March 2008 Update.
• The system base page size was made tunable by the system administrator in the September 2008
Update. For earlier releases, the system base page size was hardwired to 4 KB.
• The swapoff command was provided in the September 2008 Update to allow swap devices to
be removed immediately without requiring a system reboot.
• The September 2009 Update reintroduced kernel tunable parameters to control the behavior of the
pager, which parameters had been eliminated from the initial release. Specifically, the parameter
lotsfree_pct(5) governs the point at which the pager begins to move unused pages from
memory to the swap devices. When the fraction of free memory in the system is less than the
threshold established by
lotsfree_pct(5), the pager will eject pages to swap, otherwise it will
leave as many pages as possible in memory. The parameter
desfree_pct(5) establishes the
point at which the pager becomes more aggressive in freeing memory by ejecting pages to swap.
That is, when the fraction of free memory falls below
desfree_pct(5), the pager will take more
aggressive measures to free memory, to include putting low priority processes to sleep.
• The March 2010 Update provides enhancements to the implementation of the Mostly Private
Address Space (MPAS) model to increase performance significantly, and also promotes the use of
larger superpages in vPars instances with ejectable memory.
Differences between HP-UX 11i v2 and HP-UX 11i v3
Unified Filecache
HP-UX versions prior to HP-UX 11i v3 maintained two separate memory caches to speed up repeated
accesses to data residing on disk files: the page cache was used for memory mapped files while the
buffer cache was used for files accessed through the read and write interfaces. This required extra
administrative overhead to size the two caches and used memory space inefficiently. Moreover, since
the operating system did not maintain coherency between the two caches, applications that used both
the
mmap(2) and read(2)/write(2) interfaces to access the same files had to perform explicit
synchronization in order to behave properly.
In HP-UX 11i v3, there is one single unified filecache to hold the memory images of file data,
regardless of the interface applications use for access. This is simpler to manage, more efficient in
terms of memory usage, and usually gives an increase in performance.
Unconditional use of pseudo-swap
Prior to HP-UX 11i v3, the kernel tunable parameter swapmem_on(5) controlled whether or not
memory could be used as pseudo-swap space. Since there is no downside and considerable benefit
to using pseudo-swap, the feature is enabled unconditionally in HP-UX 11i v3.
Appendix 2 -- Variations with platform architecture
• The system base page size was made tunable by the system administrator on Integrity platforms.
On HP9000 PA-RISC, the system base page size is hardwired to 4 KB.
• The Mostly Private Address Space (MPAS) model is provided on Integrity platforms only. PA-RISC
platforms implement only the Mostly Global Address Space (MGAS) model.