Related
I have a job that requires having to write a single JSON file to s3 for each row in a Spark dataframe (which then gets picked up by another process).
df.repartition(col("id")).write.mode("overwrite").partitionBy(col("id")).json(
f"s3://bucket/path/to/file"
)
These datasets often consist of 100k rows (sometimes 1m+) and take a very long time to write. I understand that large numbers of small files is not great for read performance but is this also the case for writes? Or is there something that can be done with partitioning to speed things up?
Please don't do this, you will only suffer pain. S3 was designed to be cheap long-term storage optimized for large files. It was design so the 'prefix' (directory path) leads to a bucket that provides files. If you want to optimize reads and writes you want to develop several buckets to write to at the same time. This means you want to actually modify the directory path(prefix) to the bucket with the most amount of variation to increase the number of buckets that you write to.
Example of mulitple files being written to the same bucket:
S3:/mydrive/mystuff/2020-12-31
S3:/mydrive/mystuff/2020-12-30
S3:/mydrive/mystuff/2020-12-29
This is because they all share the same bucket prefix --> S3:/mydrive/mystuff/
What if instead you flipped the part that changes? Now you have different buckets being used as you are writing to different buckets.(prefix is different)
S3:2020-12-31/mydrive/mystuff/
S3:2020-12-30/mydrive/mystuff/
S3:2020-12-29/mydrive/mystuff/
This change will help with read/write speed as different buckets will be used. It will not solve the problem that S3 doesn't actually use directories to direct you to files. As I said a prefix is actually just a pointer to the bucket. It then searches against all files you have written, to find the file that exists in your bucket. This is why tons of small files makes things worse, the lookup time for files takes longer and longer the more files you write. Because this lookup is expensive it's much faster to write larger files and make the cost of lookup minimized.
Given a .tar archive, Matlab allows one to extract the contained files to disk via UNTAR command. One can then manipulate the extracted files in the ordinary way.
Issue: When several files are stored in a tarball, they are stored contiguously on disk and, in principle, they can be accessed serially. When such files are extracted, this contiguity doesn't hold any more and the file access can become random, hence slow & inefficient.
This is especially critical when the considered files are many (thousands) and small.
My question: is there any way to access to the archived files avoiding the preliminary extraction (in a sort of HDF5 fashion)?
In other words, would it be possible to cache the .tar so to access the contained files from the memory rather than from the disk?
(In general, direct .tar manipulation is possible, e.g. is C# tar-cs, in python).
No, as far as i know.
If you're using Matlab on Linux, try to extract to tmpnam. This will extract to tmpfs whitch should be faster accessible (bad idea if we are takling about several GB).
Otherway you can use system('untar xf file.tar only/needed/file') or python to get a more flexible untar behavior.
After some time I finally worked out a solution which gave me unbelievable speedups (like 10x or so).
In a word: ramdisk (tested on Linux (Ubuntu & CentOs)).
Recap:
Since the problem has some generality, let me state it again in a more complete fashion.
Say that I have many small files stored on disk (txt,pict, order of millions) which I want to manipulate (e.g. via matlab).
Working on such files (i.e. loading them/transmitting them on network) when they are stored on disk is tremendously slow since the disk access is mostly random.
Hence, tarballing the files in archives (e.g. of fixed size) looked to me like a good way to keep the disk access sequential.
Problem:
If case the manipulation of the .tar requires a preliminary extraction to disk (as it happens with matlab's UNTAR), the speed up given by sequential disk access is mostly loss.
Workaround:
The tarball (provided it is reasonably small) can be extracted to memory and then processed from there. In matlab, as I stated in the question, .tar manipulation in memory is not possible, though.
What can be done (equivalently) is untarring to ramdisk.
In linux, e.g. Ubuntu, a default ramdisk drive is mounted in /run/shm (tempfs). Files can be untarred via matlab there, having then extremely fast access.
In other words, a possible workcycle is:
untar to /run/shm/mytemp
manipulate in memory
possibly tar again the output to disk
This allowed me to change the scale-time of my processing from 8hrs to 40min and full CPUs load.
I was wondering what is better in this case?
I have to read in thousands of files. I was thinking of opening into each file and reading one and closing it. Or cat all the files into one file and read that.
Suggestions? This is all in Perl.
It shouldn't make that much of a difference. This sounds like premature optimization to me.
If the time for cating all files into one bigger file doesn't matter it will be faster (only when reading the file sequentially which is the default).
Of course if the process is taken into account it'll be much slower because you have to read, write and read again.
In general reading one file of 1000M should be faster than reading 100 files of 10M because for the 100 files you'll need to look for the metadata.
As tchrist says the performance difference might not be important. I think it depends on the type of file (e.g. for a huge number of files which are very small it would differ much more) and the overall performance of your system and its storage.
Note that cat * can fail if number of files is greater than your ulimit -n value. So sequential read can actually be safer.
Also, consider using opendir and readdir instead of glob if all your files are located in the same dir.
Just read the files sequentially. Perl's file i/o functions are pretty thin wrappers around native file i/o calls in the OS, so there isn't much point in fretting about performance from simple file i/o.
Primarily this seems to be a technique used by games, where they have all the sounds in one file, textures in another etc. With these files commonly reaching the GB size.
What is the reason behind doing this over maintaining it all in subdirectories as small files - one per texture which many small games use this, with the monolithic system being favoured by larger companies?
Is there some file system overhead with lots of small files?
Are they trying to protect their property - although most just seem to be a compressed file with a new extension?
The reasons we use an "archive" system like this where I work (a game development company):
lookup speed: We rarely need to iterate over files in a directory; we're far more often looking them up directly by name. By using a custom "file allocation table" that is essentially just a sequence of hash( normalized_filename ) -> [ offset, size ], we can look up files very quickly. We can also keep this index in RAM, potentially interleave it with other index tables, etc.
(When we do need to iterate, we can either easily iterate over all files in a .arc, or we can store a list of filenames, a list of hash-of-filenames, or just a list of [ offset, size ] pairs somewhere -- maybe even as a file in the archive. This is usually faster than directory-traversal on a FS.)
metadata: It's easy for us to tuck in any file metadata we want. For example, a single bit in the "size" field indicates whether the file is compressed or not (if it is, it has a header with more details about how to decompress it). We can even vary compression on pieces of a file if we know enough about the structure of the file ahead of time (we do this for sprite archives).
size: One of the devices we use has a "file size must be a multiple of X" requirement, where X is large compared to some of our files. For example, some of our lua scripts end up being just a few hundred bytes when compiled; taking extra overhead per .luc file adds up quickly.
alignment: on the other hand, sometimes we want to waste space. To take advantage of faster streaming (e.g. background DMA) from the filesystem, some of our files do want to obey certain alignment/size requirements. We can take care of that right in the tool, and the align/size we're shooting for doesn't necessarily have to line up with the underlying FS, allowing us to waste space only where we need it.
But those are the mundane reasons. The more fun stuff:
Each .arc registers in a list, and attempts to open a file know to look in the arcs. We search already-in-RAM archives first, then archives on the device FS, then the actual device FS. This gives us a ton of flexibility:
dynamic additions to the filesystem: at any time we can stream a new file or archive to the machine in question (over the network or the like) and have it appear as part of the "logical" filesystem; this is great when the actual FS resides in ROM or on a CD, and allows us to iterate much more quickly than we could otherwise.
(Doom's .wad system is a sort of example of the above, which allows modders to more easily override assets and scripts built into the game.)
possibility of no underlying fs: It's possible to use bin2obj to embed an entire arc directly in the executable (.rodata) at link time, at which point you don't ever need to look at the device FS -- we do this for certain small demo builds and the like. We can also send levels across the network or savegame-sneakernet this way. =)
organization and load/unload: since we can load and unload and override virtual "pieces" of our filesystem at any time, we can do some performance tricks with having the number of files in the FS be very small at any given time. We can additionally specify that an entire archive be loaded into memory, index table and data; our file load code is smart enough to know that if the file is already in memory, it doesn't need to do anything to read it other than move a pointer around. Some of the higher level code can actually detect that the file is in ram and just ask for the probably-already-looks-like-a-struct pointer directly.
portability: we only need to figure out how to get a few files on each new device we use, and then the remainder of the FS code is more or less the same. =) We do change the tool output a bit occasionally (for alignment reasons), but most of the processing remains the same.
de-duplication: with smarter archives, such as our sprite archives, we can (and do) de-duplicate data. If "jump" animation's fifth frame and "kick"'s third frame are the same, we can pull apart the file and only store one copy of that frame. We can do the same for whole files.
We ported a PC game to a system with much slower FS access recently. We didn't change the data format, and it turns out iterating through a dir on the raw device FS to load a hundred small XML files was absolutely killing our load times. The solution we used was to take each dir, make it into its own subdir.arc, and stick it in the master game.arc compressed. When the dir was needed (something like opendir was called) we decompressed the entire subdir.arc into RAM, added it to the filesystem, then iterated through it super-quickly.
It's the ability to throw something like this together in a few hours, and to ease the pain of porting across systems, that makes stuff like this worthwhile.
File systems do have an overhead. Usually, a file takes disk space rounded up to some power of 2 (e.g. up to 4 KB), so many small files would waste space. Some modern file systems try to mitigate that, but AFAIK it's not widespread yet. Additionally, file systems are often quite slow when accessing multiple files. E.g. it is usually considerably faster to copy one 400 MB file than 4000 100 KB files.
File systems come in handy when you have to modify files, because they handle changing file sizes much better than any simple home-grown solution. However, that's certainly not the case for constant game data.
On Apple systems, the most common way is to use, as you suggest, directories. They are called Bundles, and are in the Finder represented as just one file, but if you explore them more, they're actually directories. This makes writing code and conserving memory when loading individual items out of this bundle very easy. :-) Also, this makes taking incremental backups of gigantic databases easy, as for instance your iPhoto database is just a bundle, so you just backup changed and new files
On Windows, however, I believe this is much harder to do, it will look like a directory "no matter what" (I'm sure smart people have found a solution that will make Explorer see certain directories as a single file, but it's not common).
From a games developer point of view, you're not dealing with so small files that disk space overhead is something you're very much concerned with, so I doubt #doublep's suggestion, since it makes for such a hassle, but it makes it much easier with a single file if users are to copy an entire game over somewhere, then it's easy to check if the entire set is correct.
And, of course, it's harder to read for people that shouldn't have access to it. But it's also harder to modify, which means harder to patch, and harder to write extensions. Someone that uses extensions a lot, prefers the directory structure: The Sims.
Were I the games developer, I'd love to go for individual files. Then again, I'd be using bundles as I'd be writing for the Mac ;-)
Cheers
Nik
I can think of multiple reasons.
As doublep suggested, files occupy more space on the disc than they require. So an archive saves space. 10k files (of any size) should save you 20MB when packed into an archive. Not exactly a large amount of space nowadays, but still.
The other reason I can think of is disc fragmentation. I suspect a heavily fragmented disc will perform worse when accessing thousands of separate files on a fragmented space. But I'm no expert in this field, so I'd appreciate if someone more experienced verified this.
Finally, I think this may also have something to do with restricting access to separate game files. You can have a bunch of Lua scripts exposed, mess with them and break something. Or you could have the outro cinematic/sound/text/whatever exposed and get spoiled by accessing it. I do that myself as well: I encrypt images with a multipass XOR key, pack text files and config variables into a monolithic file (zipped for extra security) and only leave music freely accessible. This way, the game's secrets will remain undiscovered for a bit longer :).
Or there may be another reason I never thought about :D.
As you know games, especially with larger companies try to squeeze as much performance as they can. One technique is to have all the data in one large file and just DMA it to memory (think of it as a memcpy from CD to RAM). Since all the files are in one large one there will be no disk seeks and you can have a large number of files (which may cause large amount of seeks) all loaded quicky because of the technique.
A product that I am working on collects several thousand readings a day and stores them as 64k binary files on a NTFS partition (Windows XP). After a year in production there is over 300000 files in a single directory and the number keeps growing. This has made accessing the parent/ancestor directories from windows explorer very time consuming.
I have tried turning off the indexing service but that made no difference. I have also contemplated moving the file content into a database/zip files/tarballs but it is beneficial for us to access the files individually; basically, the files are still needed for research purposes and the researchers are not willing to deal with anything else.
Is there a way to optimize NTFS or Windows so that it can work with all these small files?
NTFS actually will perform fine with many more than 10,000 files in a directory as long as you tell it to stop creating alternative file names compatible with 16 bit Windows platforms. By default NTFS automatically creates an '8 dot 3' file name for every file that is created. This becomes a problem when there are many files in a directory because Windows looks at the files in the directory to make sure the name they are creating isn't already in use. You can disable '8 dot 3' naming by setting the NtfsDisable8dot3NameCreation registry value to 1. The value is found in the HKEY_LOCAL_MACHINE\System\CurrentControlSet\Control\FileSystem registry path. It is safe to make this change as '8 dot 3' name files are only required by programs written for very old versions of Windows.
A reboot is required before this setting will take effect.
NTFS performance severely degrades after 10,000 files in a directory. What you do is create an additional level in the directory hierarchy, with each subdirectory having 10,000 files.
For what it's worth, this is the approach that the SVN folks took in version 1.5. They used 1,000 files as the default threshold.
The performance issue is being caused by the huge amount of files in a single directory: once you eliminate that, you should be fine. This isn't a NTFS-specific problem: in fact, it's commonly encountered with user home/mail files on large UNIX systems.
One obvious way to resolve this issue, is moving the files to folders with a name based on the file name. Assuming all your files have file names of similar length, e.g. ABCDEFGHI.db, ABCEFGHIJ.db, etc, create a directory structure like this:
ABC\
DEF\
ABCDEFGHI.db
EFG\
ABCEFGHIJ.db
Using this structure, you can quickly locate a file based on its name. If the file names have variable lengths, pick a maximum length, and prepend zeroes (or any other character) in order to determine the directory the file belongs in.
I have seen vast improvements in the past from splitting the files up into a nested hierarchy of directories by, e.g., first then second letter of filename; then each directory does not contain an excessive number of files. Manipulating the whole database is still slow, however.
I have run into this problem lots of times in the past. We tried storing by date, zipping files below the date so you don't have lots of small files, etc. All of them were bandaids to the real problem of storing the data as lots of small files on NTFS.
You can go to ZFS or some other file system that handles small files better, but still stop and ask if you NEED to store the small files.
In our case we eventually went to a system were all of the small files for a certain date were appended in a TAR type of fashion with simple delimiters to parse them. The disk files went from 1.2 million to under a few thousand. They actually loaded faster because NTFS can't handle the small files very well, and the drive was better able to cache a 1MB file anyway. In our case the access and parse time to find the right part of the file was minimal compared to the actual storage and maintenance of stored files.
You could try using something like Solid File System.
This gives you a virtual file system that applications can mount as if it were a physical disk. Your application sees lots of small files, but just one file sits on your hard drive.
http://www.eldos.com/solfsdrv/
If you can calculate names of files, you might be able to sort them into folders by date, so that each folder only have files for a particular date. You might also want to create month and year hierarchies.
Also, could you move files older than say, a year, to a different (but still accessible) location?
Finally, and again, this requires you to be able to calculate names, you'll find that directly accessing a file is much faster than trying to open it via explorer. For example, saying
notepad.exe "P:\ath\to\your\filen.ame"
from the command line should actually be pretty quick, assuming you know the path of the file you need without having to get a directory listing.
One common trick is to simply create a handful of subdirectories and divvy up the files.
For instance, Doxygen, an automated code documentation program which can produce tons of html pages, has an option for creating a two-level deep directory hierarchy. The files are then evenly distributed across the bottom directories.
Aside from placing the files in sub-directories..
Personally, I would develop an application that keeps the interface to that folder the same, ie all files are displayed as being individual files. Then in the application background actually takes these files and combine them into a larger files(and since the sizes are always 64k getting the data you need should be relatively easy) To get rid of the mess you have.
So you can still make it easy for them to access the files they want, but also lets you have more control how everything is structured.
Having hundreds of thousands of files in a single directory will indeed cripple NTFS, and there is not really much you can do about that. You should reconsider storing the data in a more practical format, like one big tarball or in a database.
If you really need a separate file for each reading, you should sort them into several sub directories instead of having all of them in the same directory. You can do this by creating a hierarchy of directories and put the files in different ones depending on the file name. This way you can still store and load your files knowing just the file name.
The method we use is to take the last few letters of the file name, reversing them, and creating one letter directories from that. Consider the following files for example:
1.xml
24.xml
12331.xml
2304252.xml
you can sort them into directories like so:
data/1.xml
data/24.xml
data/1/3/3/12331.xml
data/2/5/2/4/0/2304252.xml
This scheme will ensure that you will never have more than 100 files in each directory.
Consider pushing them to another server that uses a filesystem friendlier to massive quantities of small files (Solaris w/ZFS for example)?
If there are any meaningful, categorical, aspects of the data you could nest them in a directory tree. I believe the slowdown is due to the number of files in one directory, not the sheer number of files itself.
The most obvious, general grouping is by date, and gives you a three-tiered nesting structure (year, month, day) with a relatively safe bound on the number of files in each leaf directory (1-3k).
Even if you are able to improve the filesystem/file browser performance, it sounds like this is a problem you will run into in another 2 years, or 3 years... just looking at a list of 0.3-1mil files is going to incur a cost, so it may be better in the long-term to find ways to only look at smaller subsets of the files.
Using tools like 'find' (under cygwin, or mingw) can make the presence of the subdirectory tree a non-issue when browsing files.
Rename the folder each day with a time stamp.
If the application is saving the files into c:\Readings, then set up a scheduled task to rename Reading at midnight and create a new empty folder.
Then you will get one folder for each day, each containing several thousand files.
You can extend the method further to group by month. For example, C:\Reading become c:\Archive\September\22.
You have to be careful with your timing to ensure you are not trying to rename the folder while the product is saving to it.
To create a folder structure that will scale to a large unknown number of files, I like the following system:
Split the filename into fixed length pieces, and then create nested folders for each piece except the last.
The advantage of this system is that the depth of the folder structure only grows as deep as the length of the filename. So if your files are automatically generated in a numeric sequence, the structure is only is deep is it needs to be.
12.jpg -> 12.jpg
123.jpg -> 12\123.jpg
123456.jpg -> 12\34\123456.jpg
This approach does mean that folders contain files and sub-folders, but I think it's a reasonable trade off.
And here's a beautiful PowerShell one-liner to get you going!
$s = '123456'
-join (( $s -replace '(..)(?!$)', '$1\' -replace '[^\\]*$','' ), $s )