Bert's blog

Memory mapping files

Input and output, however tricky, are usually considered of secondary importance during a computationally intensive numerical simulation, since the time spent reading and writing files is usually a lot shorter than the time spent doing calculations. Having good input and output routines can however become important if these are called very frequently, since file manipulation involves (slow) hard drives and takes a long time compared to more basic CPU tasks like calculations.

In this post, I will discuss an interesting method to read and write files called memory mapping. In essence, this method replaces expensive file manipulations with much cheaper memory access, transferring the responsibility for how this memory is written to and read from an associated data file to the operating system, rather than your own software. This has a few important advantages, in terms of speed, number of reads/writes and even the possibility to share files between different programs without the need to write everything to disk.

What is memory mapping?

When a program reads or writes data from a file, then a location (usually encoded as a number of bytes since the start of the file) in that file is accessed and transferred from the hard drive to a buffer in memory. Depending on how the file was encoded, this memory location is then used to access the data, either as simple binary contents (for binary files) or to perform some more complicated type conversions (for text files). Reading a single line of data (whichever way you want to define line in this context) hence requires a complex succession of instructions that copy from hard drive to memory and between different locations in memory. The details of this are pretty much hidden from you as a user or code developer, but you can be sure that any file reading or writing routine makes use of some intermediary memory buffers where data is stored after it is read or before it is written.

The idea of memory mapping is to formalise this idea of intermediate buffers and make these buffers directly accessible to the program, rather than hiding them and only providing some auxiliary functions that expose them. A reasonably large part of a file is mapped directly from the hard drive to the memory, and the corresponding buffer in memory is made directly accessible to the program and can be manipulated as any other memory array.

This has a few advantages. First of all, the direct access to the memory buffer means you can read and write data more efficiently and even in parallel, since you are just accessing memory. But there is more. When a file is memory-mapped, then the operating system (or more specifically, the kernel) becomes responsible for making sure the data in the file on the hard drive and the data in the memory buffer is the same. This means that the operating system can choose when to access the hard drive, which does not necessarily need to happen right away, but can happen at a time when the system has nothing better to do. This can alleviate common input/output bottlenecks in programs, when the program is just waiting for the hard drive to finish reading or writing data. And it reduces the load on the hard drive by spreading hard drive access operations better in time. Finally, the operating system will keep track of files that are memory-mapped; if two separate programs memory-map the same file, then they will effectively share the same memory buffer. This means that programs can exchange data before it is even written to disk!

That sounds great, how do I do this?

The reason not everyone uses memory mapping for all input and output operations is that it is quite technical. And as with any input/output functionality, the exact details of how to memory map a file depend on the programming language you are using. Unsurprisingly, Python provides the most straightforward way to do this.

Python

NumPy has a straightforward memory mapping function called memmap. This function takes the name of the file as input and makes the memory buffer accessible as a regular NumPy array. In order for it to know which type the elements of the array should have, there is an additional argument called dtype (the default is a 64-bit integer, which is very likely not what you want). Other arguments are the type of access you want (read, write or both), the offset of the data you want to read in the file, and the desired shape of the array. An example:

import numpy as np

d = np.memmap("test.dat", dtype = 'd', mode = 'w+', shape=(5))
d[0] = 1.
d[1] = 4.
d[2] = 5.
d[3] = 1.
d[4] = 40.

This will create a binary file containing 5 double precision numbers. Overall, this is all very straightforward.

There is one additional feature of np.memmap that is worth noting. Since memory mapping effectively creates a memory buffer from the contents of the file, it is a quick way to create an array based on a file. In normal use however, changing something in this memory buffer will also change the file. So what if you want to read the data and manipulate it inside your script, but still want to preserve the old data? To support exactly this behaviour, np.memmap supports a third access mode, apart from r(+) and w+: c. This mode will memory map the file so that its contents ends up in memory, but as soon as you change something to the memory buffer, it will copy that part of the buffer to another memory location without affecting the original memory buffer, a so called copy-on-write. This means that changes you make to the memory mapped buffer will only affect the memory buffer and will not affect the underlying file.

C(++)

Since C and C++ are much more low-level languages than Python, memory mapping a file in these languages is a bit trickier. Luckily (for us) they both use the same syntax. Which essentially means that this feature only really exists in C, and that the C++ version uses the same C API.

Before I can explain how to do the memory mapping in C(++), it is important to introduce a concept called page size. The page size is the size (in bytes) of a single memory block used by the kernel. When the kernel allocates memory, the amount of memory that is allocated will always be a multiple of this page size, so the page size is basically a memory size unit for the kernel. Unfortunately, the page size is system and kernel dependent, although the most common value is , or (4 kB). The kernel provides functionality to determine the page size.

When creating a memory map, the offset in the file and size of the memory mapped buffer have to use kernel units, which means they are required to be multiples of the page size. If you use memory mapping to write a file, then this file will also need to have a size that is a multiple of the page size, but it is always possible to reduce it to its smaller actual size when you are done with it.

Since memory mapping uses a C API, it requires files created using a C API too. For both reading and writing, we use the open() function from fcntl.h (part of the POSIX library, which means this only works on UNIX systems):

#include <fcntl.h>
...

// reading
int file_read = open("test.dat", O_RDONLY, 0);
// writing
int file_write = open("test.dat", O_CREAT | O_RDWR,
                      S_IRUSR | S_IWUSR | S_IRGRP | S_IWGRP);

The syntax is clearly very low-level. The third argument in the second function call denotes the file permissions created for the new file. In this case, read and write permission is given to both the current user and the group the user belongs to.

Once a file descriptor has been created like this, it can be used in the memory mapping functions. Before we can write anything however (so only if we want to write), we need to make sure the file has at least the size that we want for our memory mapped buffer. If we do not do this, then writing to the file could potentially fail because of a lack of disk space. To do this, we use the function posix_fallocate:

#include <fcntl.h>
...

// file was created above
posix_fallocate(file_write, 0, 4096);

In this case, we make sure all bytes from 0 (the offset) to 4096 (the length, here a single page size on a typical UNIX system) are available. The function will return zero if this allocation was successful, after which we are good to go. The same function can be used later to increase the size of the file to another 4096 bytes, by replacing 0 with the end of the current file.

Once the file is ready, we can finally start memory mapping. This is done using the function mmap which is part of the POSIX header sys/mman.h:

#include <sys/mman.h>
...

// file was opened or created/allocated here

char *buffer = reinterpret_cast<char*>(
  mmap(NULL, 4096, PROT_WRITE, MAP_SHARED, file_write, 0));

In the example here, we create a memory map for writing. The map is allocated at a memory location chosen by the kernel (NULL), has a size of 4096 bytes (1 page size), has write access (PROT_WRITE) and is shared between all processes (MAP_SHARED). The map is created on top of the file descriptor file_write that we created and allocated before and starts at offset 0 bytes within that file.

If the memory mapping was successful, this function will return a pointer to the start of the memory buffer that contains the file contents. If not, it will return MAP_FAILED, a constant that can be checked.

That’s it! The memory buffer buffer can now be accessed as any other buffer. It could be cast to a type pointer so that it can be used as an ordinary array, or it can be used as a file buffer by copying data into it using memcpy.

When you are done with the memory buffer, you can unmap it using munmap, which simply takes the pointer to the start of the buffer and the size of the mapping as arguments. Strictly speaking, it can also be used to unmap part of an existing mapping, by providing the start address of the region that needs to be unmapped and an arbitrary length as arguments. The region will also be unmapped if the program finishes, but not if you close the file descriptor using close().

There are two additional technicalities worth mentioning. The first one concerns the actual file writing. As explained before, the operating system kernel is responsible for making sure that the data in the memory buffer and the data in the file are the same. It is however not specified when the kernel needs to do this. The rationale behind this is that the contents of the file on disk does not really matter for the program that memory mapped the file, since it will only access the file contents using the memory buffer, and this is always up to date. The same goes for other processes that memory map the same file; they also read the same memory buffer.

It can however sometimes be useful to force the kernel to write to disk. For this purpose, the msync function is provided. Its arguments are similar to those for munmap: a pointer to the start of a memory buffer and its length (this is the part of the buffer that will be written to disk), and an additional flag that signals when to do the write. The latter can have the values MS_SYNC (write now and wait until the write is done) or MS_ASYNC (start writing in the background). Note that munmap on the same memory region will behave exactly as MS_SYNC, except that after munmap the corresponding region is no longer memory mapped.

The second technicality concerns the final file size when using memory mapping to write files. As mentioned before, all memory mapped files need to be created with a size that is a multiple of the page size, simply because that is the size unit used by the kernel for memory management. The hard disk does not have this restriction, so you can save hard disk space by shrinking the file to its actual size once the memory has been unmapped. This can be done using ftruncate(). This function takes the file descriptor and the desired file size as an argument. Be sure to provide the right size, since it will invalidate everything in the file beyond the given size!

Can I see some examples of this?

Yes, you can. I have used memory mapping for a number of projects:

Most of my knowledge about memory mapping comes from the almost continuous output (called the log file) that my collaborators have been developing for the simulation code SWIFT.


Professional astronomer.