How can we control the window in RSS when mapping a large file? Now let me explain what i mean.
For example, we have a large file that exceeds RAM by several times, we do shared memory mmaping for several processes, if we access some object whose virtual address is located in this mapped memory and catch a page fault, then reading from disk, the sub-question is, will the opposite happen if we no longer use the given object? If this happens like an LRU, then what is the size of the LRU and how to control it? How is page cache involved in this case?
RSS graph
This is the RSS graph on testing instance(2 thread, 8 GB RAM) for 80 GB tar file. Where does this value of 3800 MB come from and stay stable when I run through the file after it has been mapped? How can I control it (or advise the kernel to control it)?
As long as you're not taking explicit action to lock the pages in memory, they should eventually be swapped back out automatically. The kernel basically uses a memory pressure heuristic to decide how much of physical memory to devote to swapped-in pages, and frequently rebalances as needed.
If you want to take a more active role in controlling this process, have a look at the madvise() system call.
This allows you to tweak the paging algorithm for your mmap, with actions like:
MADV_FREE (since Linux 4.5)
The application no longer requires the pages in the range specified by addr and len. The kernel can thus free these pages, but the freeing could be delayed until memory pressure occurs. ...
MADV_COLD (since Linux 5.4)
Deactivate a given range of pages. This will make the pages a more probable reclaim target should there be a memory pressure.
MADV_SEQUENTIAL
Expect page references in sequential order. (Hence, pages in the given range can be aggressively read ahead, and may be freed soon after they are accessed.)
MADV_WILLNEED
Expect access in the near future. (Hence, it might be a good idea to read some pages ahead.)
MADV_DONTNEED
Do not expect access in the near future. (For the time being, the application is finished with the given range, so the kernel can free resources associated with it.) ...
Issuing an madvise(MADV_SEQUENTIAL) after creating the mmap might be sufficient to get acceptable behavior. If not, you could also intersperse some MADV_WILLNEED/MADV_DONTNEED access hints (and/or MADV_FREE/MADV_COLD) during the traversal as you pass groups of pages.
Related
This question already has answers here:
Closed 10 years ago.
Possible Duplicate:
mmap() vs. reading blocks
I heard (read it on the internet somewhere) that mmap() is faster than sequential IO. Is this correct? If yes then why it is faster?
mmap() is not reading sequentially.
mmap() has to fetch from the disk itself same as read() does
The mapped area is not sequential - so no DMA (?).
So mmap() should actually be slower than read() from a file? Which of my assumptions above are wrong?
I heard (read it on the internet somewhere) that mmap() is faster than sequential IO. Is this correct? If yes then why it is faster?
It can be - there are pros and cons, listed below. When you really have reason to care, always benchmark both.
Quite apart from the actual IO efficiency, there are implications for the way the application code tracks when it needs to do the I/O, and does data processing/generation, that can sometimes impact performance quite dramatically.
mmap() is not reading sequentially.
2) mmap() has to fetch from the disk itself same as read() does
3) The mapped area is not sequential - so no DMA (?).
So mmap() should actually be slower than read() from a file? Which of my assumptions above are wrong?
is wrong... mmap() assigns a region of virtual address space corresponding to file content... whenever a page in that address space is accessed, physical RAM is found to back the virtual addresses and the corresponding disk content is faulted into that RAM. So, the order in which reads are done from the disk matches the order of access. It's a "lazy" I/O mechanism. If, for example, you needed to index into a huge hash table that was to be read from disk, then mmaping the file and starting to do access means the disk I/O is not done sequentially and may therefore result in longer elapsed time until the entire file is read into memory, but while that's happening lookups are succeeding and dependent work can be undertaken, and if parts of the file are never actually needed they're not read (allow for the granularity of disk and memory pages, and that even when using memory mapping many OSes allow you to specify some performance-enhancing / memory-efficiency tips about your planned access patterns so they can proactively read ahead or release memory more aggressively knowing you're unlikely to return to it).
absolutely true
"The mapped area is not sequential" is vague. Memory mapped regions are "contiguous" (sequential) in virtual address space. We've discussed disk I/O being sequential above. Or, are you thinking of something else? Anyway, while pages are being faulted in, they may indeed be transferred using DMA.
Further, there are other reasons why memory mapping may outperform usual I/O:
there's less copying:
often OS & library level routines pass data through one or more buffers before it reaches an application-specified buffer, the application then dynamically allocates storage, then copies from the I/O buffer to that storage so the data's usable after the file reading completes
memory mapping allows (but doesn't force) in-place usage (you can just record a pointer and possibly length)
continuing to access data in-place risks increased cache misses and/or swapping later: the file/memory-map could be more verbose than data structures into which it could be parsed, so access patterns on data therein could have more delays to fault in more memory pages
memory mapping can simplify the application's parsing job by letting the application treat the entire file content as accessible, rather than worrying about when to read another buffer full
the application defers more to the OS's wisdom re number of pages that are in physical RAM at any single point in time, effectively sharing a direct-access disk cache with the application
as well-wisher comments below, "using memory mapping you typically use less system calls"
if multiple processes are accessing the same file, they should be able to share the physical backing pages
The are also reasons why mmap may be slower - do read Linus Torvald's post here which says of mmap:
...page table games along with the fault (and even just TLB miss)
overhead is easily more than the cost of copying a page in a nice
streaming manner...
And from another of his posts:
quite noticeable setup and teardown costs. And I mean noticeable. It's things like following the page tables to unmap everything cleanly. It's the book-keeping for maintaining a list of all the mappings. It's The TLB flush needed after unmapping stuff.
page faulting is expensive. That's how the mapping gets populated, and it's quite slow.
Linux does have "hugepages" (so one TLB entry per 2MB, instead of per 4kb) and even Transparent Huge Pages, where the OS attempts to use them even if the application code wasn't written to explicitly utilise them.
FWIW, the last time this arose for me at work, memory mapped input was 80% faster than fread et al for reading binary database records into a proprietary database, on 64 bit Linux with ~170GB files.
mmap() can share between process.
DMA will be used whenever possible. DMA does not require contiguous memory -- many high end cards support scatter-gather DMA.
The memory area may be shared with kernel block cache if possible. So there is lessor copying.
Memory for mmap is allocated by kernel, it is always aligned.
"Faster" in absolute terms doesn't exist. You'd have to specify constraints and circumstances.
mmap() is not reading sequentially.
what makes you think that? If you really access the mapped memory sequentially, the system will usually fetch the pages in that order.
mmap() has to fetch from the disk itself same as read() does
sure, but the OS determines the time and buffer size
The mapped area is not sequential - so no DMA (?).
see above
What mmap helps with is that there is no extra user space buffer involved, the "read" takes place there where the OS kernel sees fit and in chunks that can be optimized. This may be an advantage in speed, but first of all this is just an interface that is easier to use.
If you want to know about speed for a particular setup (hardware, OS, use pattern) you'd have to measure.
I want to mmap a big file into memory and parse it sequentially. As I understand if bytes have been lazily read into memory once, they stay there. Is there a way to periodically tell the system to release the previously read contents?
This understanding is only a very superficial view.
To understand what really happens you have to take into account the difference of the virtual memory of your process and the actual real memory of the machine. Mapping a huge file means reserving space in your virtual address-space. It's probably platform-dependent if anything is already read at this point.
When you actually access the data the OS has to fill an actual page of memory. When you access other parts these parts have to be brought into memory. It's completely up to the OS when it will re-use the memory. Normally this happens when some data is accessed by you or an other process and no free memory is available. But could happen at any time. If you access it again later it might be still in memory or will brought back by the OS. No way for your process to tell the difference.
In short: You don't need to care about that. The OS manages all that in the background.
One point might be that if you map a really huge file this takes up space in your virtual address-space which is limited. So if you deal with many huge mappings and or huge allocations you might want to only map parts of the file at a given time.
ADDITION: after thinking a bit about it, I came up with a reason why it might be smarter to do it blockwise-sequential. Although I doubt you will be able to measure that.
Any reasonable OS will look for a block to unload when in need in something like the following order:
unmapped files ( not needed anymore)
LRU unmodified mapped file (can be retrieved from disc)
LRU modified mapped file (same as 2. but needs to be updated on disc before unload)
LRU allocated memory (needs to be written to swap)
So unmapping blocks known to be never used again as you go, you give the OS a hint that these should be freed earlier. This will give data that has been used less recently but might be accessed in the future a bigger chance to stay in memory.
My questions are as follows:
I mmap(memory mapping) a file into the virtual memory space.
When I access the first byte of the file using a pointer at the first time, the OS will try to access the data in memory, but it will fails and raises the page fault, because the data doesn't present in memory now. So the OS will swap the data from disk into memory. Finally my access will success.
(question is coming)
When I modify the data(in-memory) and write back into disk file, how could I just free the physical memory for other using, but remain virtual memory for fetching the data back into memory as needed?
It sounds like the page-out and page-in behaviors where the OS know the memory is exhaust, it will swap the LRU(or something like that) memory page into disk(swap files) and free the physical memory for other process, and fetch the evicted data back into memory as needed. But this mechanism is controlled by OS.
For some reasons, I need to control the page-out and page-in behaviors by myself. So how should I do? Hack the kernel?
You can use the madvise system call. Its behaviour is affected by the advice argument; there are many choices for advice and the optimal one should be picked based on the specifics of your application.
The flag MADV_DONTNEED means that the given range of physical backing frames should be unconditionally freed (i.e. paged out). Also:
After a successful MADV_DONTNEED operation, the semantics of
memory access in the specified region are changed: subsequent
accesses of pages in the range will succeed, but will result
in either repopulating the memory contents from the up-to-date
contents of the underlying mapped file (for shared file
mappings, shared anonymous mappings, and shmem-based
techniques such as System V shared memory segments) or zero-
fill-on-demand pages for anonymous private mappings.
This could be useful if you're absolutely certain that it will be very long until you access the same position again.
However it might not be necessary to force the kernel to actually page out; instead another possibility, if you're accessing the mapping sequentially is to use madvise with MADV_SEQUENTIAL to tell kernel that you'd access your memory mapping mostly sequentially:
Expect page references in sequential order. (Hence, pages in the given range can be aggressively read ahead, and may be freed soon after they are accessed.)
or MADV_RANDOM
Expect page references in random order. (Hence, read ahead may be less useful than normally.)
These are not as aggressive as explicitly calling MADV_DONTNEED to page out. (Of course you can combine these with MADV_DONTNEED as well)
In recent kernel versions there is also the MADV_FREE flag which will lazily free the page frames; they will stay mapped in if enough memory is available, but are reclaimed by the kernel if the memory pressure grows.
You can checout mlock+munlock to lock/unlock the pages. This will give you control over pages being swapped out.
You need to have CAP_IPC_LOCK capability to perform this operation though.
I am trying to understand how mmap works while looking at man mmap.
As I understand it, it adds a mapping to the page table that maps between the file and the virtual address (which is the address that is given void *addr)
So, what happens when 2 programs map the same file?
Are there 2 entries in the page table, one for each program?
So, what happens when 2 programs map the same file? Are there 2 entries in the page table, one for each program?
In modern operating systems, each process has its own page table for its memory, that may point to pages of physical memory shared with other user and kernel processes.
With MAP_SHARED, this mapping is shared: updates to the mapping are visible to other processes that map this file, and are carried through to the underlying file. The file may not actually be updated until msync(2) or munmap() is called.
This seems very interesting, but there are numerous caveats:
The actual pages mmapped by both processes for the same file may reside at the same address or at a different address in each process, storing pointers into this shared memory may not allow the other process to use them as they might point to inconsistent addresses.
The implementation may use the same physical memory pages for both mappings or not: for subtile reasons (cache strategies, out of sync reading...), even if it is the same physical memory, modifications done by one process to its memory may not be immediately reflected in the memory of the other process.
So the modification may or may not be visible to the other processes mmapping the file nor reading it via read or the FILE* stream API.
If one of the processes calls msync(), the modifications should be visible in all maps and for all yet unread portions of the file, bearing in mind that the FILE* streaming APIs may have buffered some data in internal unshared buffers: modifications in this area will not be reflected.
Conclusion: it is risky and unreliable to use these mechanisms to implement inter process communication. The behavior may depend on system specific characteristics such as the OS strategies, the CPU and cache architectures, the type of RAM in use, the clock speed, and who knows what else. It is safer to rely on proven APIs that may indeed be implemented using mmapped memory, but only if it is know to provide the correct semantics.
The actual system implementation is different. At the risk of over simplification (and omitting paging here):
A mmap will map physical page frames to a file.
So, what happens when 2 programs map the same file? Are there 2 entries in the page table, one for each program?
If two processes (P and Q) map to the same file, then P and Q will each have there own page table; each page table will have entry mapping to the same physical page frame (which could be mapped to different addresses within P and Q).
I want to keep virtual address space reserved in my process for memory that was previously used but is not presently needed. I'm interested in the situation where the host kernel is Linux and it's configured to prevent overcommit (which it does by detailed accounting for all committed memory).
If I just want to prevent the data that my application is no longer using from occupying physical memory or getting swapped to disk (wasting resources either way), I can madvise the kernel that it's unneeded, or mmap new zero pages over top of it. But neither of these approaches will necessarily reduce the amount of memory that counts as committed, which other processes are then prevented from using.
What if I replace the pages with fresh zero pages that are marked read-only? My intent is that they don't count towards committed memory, and further that I can later use mprotect to make them writable, and that it would fail if making them writable would go over the committed memory limit. Is my understanding correct? Will this work?
If you're not using the page (reading or writing to it), it won't be commited to your address space (only reserved).
But your address space is limited, so you can't play as you want/like with it.
See for example ElectricFence which may fail for large number of allocations, because of insertion of "nul page/guard page" (anonymous memory with no access).
Have a look at these thread : "mprotect() failed: Cannot allocate memory" :
http://thread.gmane.org/gmane.comp.lib.glibc.user/538/focus=976052
On Linux, assuming overcommit has not been disabled, you can use the MAP_NORESERVE flag to mmap, which will ensure that the page in question will not be accounted as allocated memory prior to being accessed. If overcommit has been completely disabled, see below about multiple-mapping pages.
Note that Linux's behavior for zero pages has changed at times in the past; with some kernel versions, simply reading the page would cause it to be allocated. With others, a write is necessary. Note that the protection flags do not cause allocation directly; however they can prevent you from accidentally triggering an allocation. Therefore, for most reliable results you should avoid accessing the page at all by mprotecting with PROT_NONE.
As another, more portable option, you can map the same page at multiple locations. That is, create and open an empty temp file, unlink it, ftruncate to some reasonable number of pages, then mmap repeatedly at offset 0 into the file. This will absolutely guarantee the memory only counts once against your program's memory usage. You can even use MAP_PRIVATE to auto-reallocate it when you write to the page.
This may have higher memory usage than the MAP_NORESERVE technique (both for kernel tracking data, and for the pages of the temp file itself), however, so I would recommend using MAP_NORESERVE instead when available. If you do use this technique, try to make the region being mapped reasonably large (and put it in /dev/shm if on Linux, to avoid actual disk IO). Each individual mmap call will consume a certain amount of (non-swappable) kernel memory to track it, so it's good to keep that count down.