The virtual memory (VM) subsystem was completely rewritten by Digital in V2.0 in order to improve upon the Mach design adopted by the OSF. Specifically, Digital added the following functionality to improve performance and maintainability (V4.0 added several enhancements to these areas):
mmap
MAP_PRIVATE semantics to make them compatible with
Sun Microsystems
and
System V Release 4.0
The following sections discuss these improvements to VM.
In lazy allocation, swap space is reserved dynamically as the system needs to reclaim physical memory instead of having to allocate it in advance for every page of anonymous memory (that is, memory devoted to the stack and heap of a process, and to data that is not file-backed).
However, when the list of active pages falls below the preconfigured limit and the OSF memory manager attempts to reclaim pages for the free list by paging out virtual pages, if the available swap space has already been exhausted, the OSF memory manager does not back off of the page-out and instead simply discards the page. Thus, when the process whose page has been discarded takes a page fault and attempts to reactivate the missing page, unpredictable behavior results, including system hangs, panics, and at times data corruption.
To correct this problem, Digital reworked the page-out algorithm to ensure that pages are not lost if the memory manager is unable to allocate swap space for a virtual page.
As swap space decreases, the Digital UNIX Version 4.0 memory manager logs warning messages at the console until finally, if the memory manager is unable to allocate swap space for a page-out, it selects the oldest idle process and kills it, thereby freeing up swap space and returning virtual pages to the free list.
Unlike lazy allocation, the eager reservation policy reserves a page of swap space for every page of anonymous memory that is allocated. Although this policy is expensive in terms of reserved disk space, it eliminates the chance that the memory manager will have to kill a process to reclaim virtual pages and free up swap space.
The eager reservation policy is set by default, although either the lazy or eager policy can be configured. For more information, see the System Tuning and Performance Management guide
To increase file system performance,
Digital
implemented a Unified Buffer Cache
(UBC)
fully integrated with the file system
that caches file system data and can grow or shrink
upon demand.
Unlike the conventional Buffer Cache
which is configured and allocated at boot-time
and which relies on
bcopy
routines
to move data in and out of memory,
the UBC
references the same physical pages as virtual memory and
can use map operations rather than
bcopy
routines
to access data,
thereby increasing system performance.
In addition,
since the
UBC
contains only file system data,
the Buffer Cache only needs to cache metadata,
requiring only 3% of physical memory rather
than the 25% required by previous versions of the operating system.
By default,
the UBC
can grow to consume
all of physical memory
so that the system
can determine
dynamically
the percent of
memory that should be allocated to the UBC.
However,
the maximum percent of memory that the
UBC can grow to is configurable and can be set in the
system configuration file by defining the
ubc-maxpercent
variable.
In an effort to improve performance, Digital changed the OSF paging algorithm to support simultaneous paging to multiple swap partitions. By contrast, OSF paging is to one swap partition at a time, waiting until the first swap partition is filled before moving to the next. As a result, since disk transfer rates are several thousand times slower than the speed of memory and the Alpha CPU, system administrators can greatly reduce this disparity in speed by spreading swap partitions among different disks and different controllers. In fact, by supporting simultaneous paging to multiple swap partitions, Digital UNIX Version 4.0 allows multiple tasks to take simultaneous page faults, thereby further increasing performance.
Whereas OSF/1 supports a paging algorithm that moves pages in and out one at a time, to improve performance, Digital UNIX adds support for page in and page out clustering. Although in most cases, it is more efficient to do multiple multipage DMA operations rather than multiple single page DMA operations, this is a configurable option.
Rather than mapping the I/O space into memory, the memory-mapped device interface points to a data structure that defines the I/O space. As a result, large quantities of kernel virtual address space are saved as well as physical memory in general.
Since the OSF departed from the
Sun Microsystems
and
System V Release 4.0
mmap
semantics,
Digital
rewrote
mmap
so that
Sun and
System V
applications could compile and run on
Digital UNIX Version 4.0.
Using the same kind of permission bits that the file system employs, shared memory segments can be read and write protected to prevent unwanted access.
Shared text segments
allow
multiple processes to share the same page tables
that map shared text.
All processes that share the same text segment
benefit from one process taking a page fault, because less memory
is needed and the performance of
fork
is improved.
The Alpha EV4 CPU contains a direct mapped physical OFF chip secondary cache, which is organized so that if the secondary cache size is N pages, then every Nth page of the physical pages of memory hashes into the same page. Digital UNIX VM manages the physical pages of memory in such a way that, if an entire resident working set of a process can fit into the secondary cache, VM places it there. As a result, because VM strives to ensure that a process's entire working set is always in the secondary cache, the number of physical memory accesses is greatly reduced as a process executes.
The Alpha EV5 CPU optionally contains 3 levels of caches: internal, secondary, and tertiary caches. Digital UNIX manages physical memory in such a way that the most active subset of a process working set will remain in the fastest cache.
A new kernel memory allocator (kernel
malloc)
was added to
Digital UNIX
to use kernel-wired memory
more efficiently.
All calls to
the
mbuf
allocator are now
mapped to the new kernel memory allocator.
In addition, several components of the I/O subsystems can use
the kernel memory allocator directly, rather than
having to manage memory on their
own.
As a result,
we save the amount of memory these allocators were
reserving.
In addition,
the new allocator handles allocation under
interrupt context better than the
kalloc
allocator
and
has a
garbage collection thread to free memory.
Although the OSF provides guidelines for writing an external pager, these guidelines are (at best) provisional. Digital believes that the complexity and efficiency of the Digital UNIX Version 4.0 memory management system makes it impractical at this time to provide a sensible interface for an external pager.
The memory reclamation policy was enhanced in Digital UNIX Version 4.0 to improve system performance. In previous releases of the operating system, once the system began running out of physical memory, global paging would begin to reclaim memory, attempting to select pages fairly between the Unified Buffer Cache (UBC) and VM. However, as the system runs and files are opened and closed, a large percentage of memory in the UBC is always referencing old, closed files--owing to its file caching algorithms. As a result, attempting to select some pages from VM and some pages from the UBC is actually unnecessary initially, since theoretically only 10% of the pages in the UBC are dirty, whereas almost all the pages in VM are dirty (in actual practice, however, the amount of dirty pages in the UBC is much smaller than 10% since most of the pages in the UBC are not in use). Because the UBC, unlike VM, contains so few dirty pages, it is much more efficient to reclaim pages from the UBC first, down to some configurable level, than it is to begin global paging immediately.
In
Digital UNIX Version 4.0,
the page reclamation policy was further enhanced
to take advantage of the many
unreferenced pages that are in both VM and UBC.
When a system begins to exhaust its
physical memory,
memory is first reclaimed from the
UBC
down to a threshold
defined by the parameter
ubc-borrowpercent.
This
parameter is set
by default
to be 10% of the memory in the UBC
and is
configurable.
If,
after
all unused
memory is
reclaimed from the UBC and
the system still requires more physical memory,
global paging is then invoked.
In effect, this policy can double the load that can be placed on a system before demands for memory begin to noticeably degrade performance.
In Digital UNIX Version 4.0, the swap allocation mechanism was restructured to further reduce the fragmentation of swap space which could occur over time. Swap space is now allocated and deallocated in contiguous units, so that seek time is greatly reduced, thereby greatly improving performance.
For information on how to tune and configure the virtual memory subsystem, see the System Tuning and Performance Management guide and the System Administration guide .