Source string Read only

(itstool) path: sect1/para
875/8750
Context English State
The entire radix tree bitmap is also preallocated in order to avoid having to allocate kernel memory during critical low memory swapping operations. After all, the system tends to swap when it is low on memory so we should avoid allocating kernel memory at such times in order to avoid potential deadlocks.
To reduce fragmentation the radix tree is capable of allocating large contiguous chunks at once, skipping over smaller fragmented chunks.
I did not take the final step of having an <quote>allocating hint pointer</quote> that would trundle through a portion of swap as allocations were made in order to further guarantee contiguous allocations or at least locality of reference, but I ensured that such an addition could be made.
When to free a page
Since the VM system uses all available memory for disk caching, there are usually very few truly-free pages. The VM system depends on being able to properly choose pages which are not in use to reuse for new allocations. Selecting the optimal pages to free is possibly the single-most important function any VM system can perform because if it makes a poor selection, the VM system may be forced to unnecessarily retrieve pages from disk, seriously degrading system performance.
How much overhead are we willing to suffer in the critical path to avoid freeing the wrong page? Each wrong choice we make will cost us hundreds of thousands of CPU cycles and a noticeable stall of the affected processes, so we are willing to endure a significant amount of overhead in order to be sure that the right page is chosen. This is why FreeBSD tends to outperform other systems when memory resources become stressed.
The free page determination algorithm is built upon a history of the use of memory pages. To acquire this history, the system takes advantage of a page-used bit feature that most hardware page tables have.
In any case, the page-used bit is cleared and at some later point the VM system comes across the page again and sees that the page-used bit has been set. This indicates that the page is still being actively used. If the bit is still clear it is an indication that the page is not being actively used. By testing this bit periodically, a use history (in the form of a counter) for the physical page is developed. When the VM system later needs to free up some pages, checking this history becomes the cornerstone of determining the best candidate page to reuse.
What if the hardware has no page-used bit?
For those platforms that do not have this feature, the system actually emulates a page-used bit. It unmaps or protects a page, forcing a page fault if the page is accessed again. When the page fault is taken, the system simply marks the page as having been used and unprotects the page so that it may be used. While taking such page faults just to determine if a page is being used appears to be an expensive proposition, it is much less expensive than reusing the page for some other purpose only to find that a process needs it back and then have to go to disk.
FreeBSD makes use of several page queues to further refine the selection of pages to reuse as well as to determine when dirty pages must be flushed to their backing store. Since page tables are dynamic entities under FreeBSD, it costs virtually nothing to unmap a page from the address space of any processes using it. When a page candidate has been chosen based on the page-use counter, this is precisely what is done. The system must make a distinction between clean pages which can theoretically be freed up at any time, and dirty pages which must first be written to their backing store before being reusable. When a page candidate has been found it is moved to the inactive queue if it is dirty, or the cache queue if it is clean. A separate algorithm based on the dirty-to-clean page ratio determines when dirty pages in the inactive queue must be flushed to disk. Once this is accomplished, the flushed pages are moved from the inactive queue to the cache queue. At this point, pages in the cache queue can still be reactivated by a VM fault at relatively low cost. However, pages in the cache queue are considered to be <quote>immediately freeable</quote> and will be reused in an LRU (least-recently used) fashion when the system needs to allocate new memory.
It is important to note that the FreeBSD VM system attempts to separate clean and dirty pages for the express reason of avoiding unnecessary flushes of dirty pages (which eats I/O bandwidth), nor does it move pages between the various page queues gratuitously when the memory subsystem is not being stressed. This is why you will see some systems with very low cache queue counts and high active queue counts when doing a <command>systat -vm</command> command. As the VM system becomes more stressed, it makes a greater effort to maintain the various page queues at the levels determined to be the most effective.
An urban myth has circulated for years that Linux did a better job avoiding swapouts than FreeBSD, but this in fact is not true. What was actually occurring was that FreeBSD was proactively paging out unused pages in order to make room for more disk cache while Linux was keeping unused pages in core and leaving less memory available for cache and process pages. I do not know whether this is still true today.
Pre-Faulting and Zeroing Optimizations
Taking a VM fault is not expensive if the underlying page is already in core and can simply be mapped into the process, but it can become expensive if you take a whole lot of them on a regular basis. A good example of this is running a program such as <citerefentry><refentrytitle>ls</refentrytitle><manvolnum>1</manvolnum></citerefentry> or <citerefentry><refentrytitle>ps</refentrytitle><manvolnum>1</manvolnum></citerefentry> over and over again. If the program binary is mapped into memory but not mapped into the page table, then all the pages that will be accessed by the program will have to be faulted in every time the program is run. This is unnecessary when the pages in question are already in the VM Cache, so FreeBSD will attempt to pre-populate a process's page tables with those pages that are already in the VM Cache. One thing that FreeBSD does not yet do is pre-copy-on-write certain pages on exec. For example, if you run the <citerefentry><refentrytitle>ls</refentrytitle><manvolnum>1</manvolnum></citerefentry> program while running <command>vmstat 1</command> you will notice that it always takes a certain number of page faults, even when you run it over and over again. These are zero-fill faults, not program code faults (which were pre-faulted in already). Pre-copying pages on exec or fork is an area that could use more study.
A large percentage of page faults that occur are zero-fill faults. You can usually see this by observing the <command>vmstat -s</command> output. These occur when a process accesses pages in its BSS area. The BSS area is expected to be initially zero but the VM system does not bother to allocate any memory at all until the process actually accesses it. When a fault occurs the VM system must not only allocate a new page, it must zero it as well. To optimize the zeroing operation the VM system has the ability to pre-zero pages and mark them as such, and to request pre-zeroed pages when zero-fill faults occur. The pre-zeroing occurs whenever the CPU is idle but the number of pages the system pre-zeros is limited in order to avoid blowing away the memory caches. This is an excellent example of adding complexity to the VM system in order to optimize the critical path.
Page Table Optimizations
The page table optimizations make up the most contentious part of the FreeBSD VM design and they have shown some strain with the advent of serious use of <function>mmap()</function>. I think this is actually a feature of most BSDs though I am not sure when it was first introduced. There are two major optimizations. The first is that hardware page tables do not contain persistent state but instead can be thrown away at any time with only a minor amount of management overhead. The second is that every active page table entry in the system has a governing <literal>pv_entry</literal> structure which is tied into the <literal>vm_page</literal> structure. FreeBSD can simply iterate through those mappings that are known to exist while Linux must check all page tables that <emphasis>might</emphasis> contain a specific mapping to see if it does, which can achieve O(n^2) overhead in certain situations. It is because of this that FreeBSD tends to make better choices on which pages to reuse or swap when memory is stressed, giving it better performance under load. However, FreeBSD requires kernel tuning to accommodate large-shared-address-space situations such as those that can occur in a news system because it may run out of <literal>pv_entry</literal> structures.
Both Linux and FreeBSD need work in this area. FreeBSD is trying to maximize the advantage of a potentially sparse active-mapping model (not all processes need to map all pages of a shared library, for example), whereas Linux is trying to simplify its algorithms. FreeBSD generally has the performance advantage here at the cost of wasting a little extra memory, but FreeBSD breaks down in the case where a large file is massively shared across hundreds of processes. Linux, on the other hand, breaks down in the case where many processes are sparsely-mapping the same shared library and also runs non-optimally when trying to determine whether a page can be reused or not.
Page Coloring
We will end with the page coloring optimizations. Page coloring is a performance optimization designed to ensure that accesses to contiguous pages in virtual memory make the best use of the processor cache. In ancient times (i.e. 10+ years ago) processor caches tended to map virtual memory rather than physical memory. This led to a huge number of problems including having to clear the cache on every context switch in some cases, and problems with data aliasing in the cache. Modern processor caches map physical memory precisely to solve those problems. This means that two side-by-side pages in a processes address space may not correspond to two side-by-side pages in the cache. In fact, if you are not careful side-by-side pages in virtual memory could wind up using the same page in the processor cache—leading to cacheable data being thrown away prematurely and reducing CPU performance. This is true even with multi-way set-associative caches (though the effect is mitigated somewhat).
FreeBSD's memory allocation code implements page coloring optimizations, which means that the memory allocation code will attempt to locate free pages that are contiguous from the point of view of the cache. For example, if page 16 of physical memory is assigned to page 0 of a process's virtual memory and the cache can hold 4 pages, the page coloring code will not assign page 20 of physical memory to page 1 of a process's virtual memory. It would, instead, assign page 21 of physical memory. The page coloring code attempts to avoid assigning page 20 because this maps over the same cache memory as page 16 and would result in non-optimal caching. This code adds a significant amount of complexity to the VM memory allocation subsystem as you can well imagine, but the result is well worth the effort. Page Coloring makes VM memory as deterministic as physical memory in regards to cache performance.
Conclusion
Virtual memory in modern operating systems must address a number of different issues efficiently and for many different usage patterns. The modular and algorithmic approach that BSD has historically taken allows us to study and understand the current implementation as well as relatively cleanly replace large sections of the code. There have been a number of improvements to the FreeBSD VM system in the last several years, and work is ongoing.
Bonus QA session by Allen Briggs <email>briggs@ninthwonder.com</email>
What is <quote>the interleaving algorithm</quote> that you refer to in your listing of the ills of the FreeBSD 3.X swap arrangements?
FreeBSD uses a fixed swap interleave which defaults to 4. This means that FreeBSD reserves space for four swap areas even if you only have one, two, or three. Since swap is interleaved the linear address space representing the <quote>four swap areas</quote> will be fragmented if you do not actually have four swap areas. For example, if you have two swap areas A and B FreeBSD's address space representation for that swap area will be interleaved in blocks of 16 pages:
A B C D A B C D A B C D A B C D
FreeBSD 3.X uses a <quote>sequential list of free regions</quote> approach to accounting for the free swap areas. The idea is that large blocks of free linear space can be represented with a single list node (<filename>kern/subr_rlist.c</filename>). But due to the fragmentation the sequential list winds up being insanely fragmented. In the above example, completely unused swap will have A and B shown as <quote>free</quote> and C and D shown as <quote>all allocated</quote>. Each A-B sequence requires a list node to account for because C and D are holes, so the list node cannot be combined with the next A-B sequence.
Why do we interleave our swap space instead of just tack swap areas onto the end and do something fancier? It is a whole lot easier to allocate linear swaths of an address space and have the result automatically be interleaved across multiple disks than it is to try to put that sophistication elsewhere.
The fragmentation causes other problems. Being a linear list under 3.X, and having such a huge amount of inherent fragmentation, allocating and freeing swap winds up being an O(N) algorithm instead of an O(1) algorithm. Combined with other factors (heavy swapping) and you start getting into O(N^2) and O(N^3) levels of overhead, which is bad. The 3.X system may also need to allocate KVM during a swap operation to create a new list node which can lead to a deadlock if the system is trying to pageout pages in a low-memory situation.

Loading…

No matching activity found.

Browse all component changes

Glossary

English English
No related strings found in the glossary.

Source information

Source string comment
(itstool) path: sect1/para
Flags
read-only
Source string location
article.translate.xml:529
String age
a year ago
Source string age
a year ago
Translation file
articles/vm-design.pot, string 67