Fast files
-
A common beginner's mistake I see when doing file I/O is writing or reading one byte at a time (or a very small number of bytes). You can greatly increase I/O time by writing/reading good sized chunks at a time. Perhaps this is your issue? How big to make the buffers is pretty much application dependant, but even going to say several K at a time can make a huge difference.
Actually a have been considering that fact for a reason: As an algorithm restriction, I must deal with data byte by byte in order to accomplish the compression. So I begun wondering which was faster: to read byte by byte or to have to apply shift rights, bit wise orings and andings to extract the data (into a single byte) and then apply the rest of the algorithm on that byte which also includes those bit wise operations. I would guess that the 2nd option is faster because it works only with registers and main memory but there quite a bit of those operations on the algorithm. So my question is: does it really payoff to read memory chunks instead of byte by byte in this particular case? Thx! hint_54
-
hint_54 wrote:
You are talking about DMA rigth?
No. DMA is a way to access memory independently of the CPU.
hint_54 wrote:
I don't know how to use it with files!
A memory-mapped file is a spot in memory that can be accessed (e.g., open, close, read, write, seek) as though it were an actual file. By eliminating file I/O, a speed increase is (normally) realized. See these for examples:* http://www.ecst.csuchico.edu/~beej/guide/ipc/mmap.html
-
http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dngenlib/html/msdn\_manamemo.asp
-
http://www.codeproject.com/file/Memory\_Mapped\_Class\_\_\_PBD.asp
-
http://www.codeproject.com/win32/cmemmap.asp
"Take only what you need and leave the land as you found it." - Native American Proverb
-
-
hint_54 wrote:
...isn’t that memory consuming for larger files?
I guess that depends on your definition of "larger files." I wouldn't think twice about using MMF for files several MB in size. In any case, you can restrict the mapping to only the portion of the file you are interested in.
"Take only what you need and leave the land as you found it." - Native American Proverb
-
hint_54 wrote:
...isn’t that memory consuming for larger files?
I guess that depends on your definition of "larger files." I wouldn't think twice about using MMF for files several MB in size. In any case, you can restrict the mapping to only the portion of the file you are interested in.
"Take only what you need and leave the land as you found it." - Native American Proverb
There is no definition for larger files because this app is meant to be used by someone else, meaning that I can never know the maximum file size to be used (can be anything! And limiting that size is not an option.) and I can’t load portions of the file because the algorithm reads the file sequentially, woks on data (byte by byte) and stores the output way. But thx anyway ;) hint_54
-
Actually a have been considering that fact for a reason: As an algorithm restriction, I must deal with data byte by byte in order to accomplish the compression. So I begun wondering which was faster: to read byte by byte or to have to apply shift rights, bit wise orings and andings to extract the data (into a single byte) and then apply the rest of the algorithm on that byte which also includes those bit wise operations. I would guess that the 2nd option is faster because it works only with registers and main memory but there quite a bit of those operations on the algorithm. So my question is: does it really payoff to read memory chunks instead of byte by byte in this particular case? Thx! hint_54
You can read the file in large chunks and process that chunk byte by byte. Even when cached, doing disk I/O one byte at a time is painfully slow.
"Take only what you need and leave the land as you found it." - Native American Proverb
-
Bottlenecks are spots in your code where most of the execution time is spent. I would suggest you profile you app and see where most of the time is spent. If you're using VC6 there is a profile that ships with it. It takes some reading to figure out how to use it (http://msdn.microsoft.com/library/default.asp?url=/library/en-us/vccore98/html/_core_using_profile.2c_.prep.2c_.and_plist.asp[^]) There's also a way to run from VS but it doesn't always work(http://msdn.microsoft.com/library/default.asp?url=/library/en-us/vccore98/html/_core_using_profile.2c_.prep.2c_.and_plist.asp[^]) If you're using VS 2002 or 2003 you'll have to get the profiler from compuware. They have a freebie version available (http://www.compuware.com/products/devpartner/profiler/default.asp?cid=3019X36&focus=DevPartner&source=Web+%2D+Evaluation+Request&desc=Download+%2D+%27DevPartner+Profiler+Community+Edition%27&offering=DevPartner&sf=1&p=0[^]) If you're using VS 2005 there is a profile included. Once you see where most of the time is being spent you can then start to think about how to improve the performance.
My app spends most of its time reading from/writing to files. This is the main reason why I need a fast way to read/write files as the rest of the algorithm isn’t much time consuming, I’d say 85% of the time spent in the compressing algorithm is with files. I need a fast way to read/write files in sequence. thx. hint_54
-
You can read the file in large chunks and process that chunk byte by byte. Even when cached, doing disk I/O one byte at a time is painfully slow.
"Take only what you need and leave the land as you found it." - Native American Proverb
-
hint_54 wrote:
You are talking about DMA rigth?
No. DMA is a way to access memory independently of the CPU.
hint_54 wrote:
I don't know how to use it with files!
A memory-mapped file is a spot in memory that can be accessed (e.g., open, close, read, write, seek) as though it were an actual file. By eliminating file I/O, a speed increase is (normally) realized. See these for examples:* http://www.ecst.csuchico.edu/~beej/guide/ipc/mmap.html
-
http://msdn.microsoft.com/library/default.asp?url=/library/en-us/dngenlib/html/msdn\_manamemo.asp
-
http://www.codeproject.com/file/Memory\_Mapped\_Class\_\_\_PBD.asp
-
http://www.codeproject.com/win32/cmemmap.asp
"Take only what you need and leave the land as you found it." - Native American Proverb
DavidCrow wrote:
A memory-mapped file is a spot in memory that can be accessed (e.g., open, close, read, write, seek) as though it were an actual file.
Sorry, but I think that is supposed to be the other way around. Memory-mapping something allows it to be accessed like normal memory (through a pointer). For example, if you had a memory mapped buttons/switches in your hardware (common on arcade games), you would be able to read the state of the switches by reading values from one or more specific memory addresses. The same goes for files. If you memory-map a file, you get an address that points to some location in the file, and you can then access the contents of that file through the pointer. Peace! -=- James
If you think it costs a lot to do it right, just wait until you find out how much it costs to do it wrong!
Tip for new SUV drivers: Professional Driver on Closed Course does not mean your Dumb Ass on a Public Road!
DeleteFXPFiles & CheckFavorites (Please rate this post!) -
-
There is no definition for larger files because this app is meant to be used by someone else, meaning that I can never know the maximum file size to be used (can be anything! And limiting that size is not an option.) and I can’t load portions of the file because the algorithm reads the file sequentially, woks on data (byte by byte) and stores the output way. But thx anyway ;) hint_54
Using MMF can still improve performance, because you do not have to do the manual loading of data into a buffer - the OS basically does it for you. For example, if you were reading the file in 4KB chunks, you would be allocating a 4K buffer, copying from the file into that buffer, and then likely processing the contents of the buffer using the buffer's address. Using a MMF does that work for you. If you want to impose a limit on the size of the MMF section you want to create, that is fine. Choose a limit, say 2MB (or 4Mb, or 64MB, whatever). If the file is 2MB or smaller, you can MM the entire file. Of not, you can MM 2MB sections of the file one at a time. Peace! -=- James
If you think it costs a lot to do it right, just wait until you find out how much it costs to do it wrong!
Tip for new SUV drivers: Professional Driver on Closed Course does not mean your Dumb Ass on a Public Road!
DeleteFXPFiles & CheckFavorites (Please rate this post!) -
A common beginner's mistake I see when doing file I/O is writing or reading one byte at a time (or a very small number of bytes). You can greatly increase I/O time by writing/reading good sized chunks at a time. Perhaps this is your issue? How big to make the buffers is pretty much application dependant, but even going to say several K at a time can make a huge difference.
Actually file I/O is never done on a byte level. The least bit of information that can be read from ,or written to a file is one sector. A sector is always an even multiple of 128 bytes, the most common sector size is 512 bytes. So the runtime does buffer (at least) one sector (or more likely, a cluster). I think the main reason for slowing down file reading a byte at the time, is the function call overhead and all the checks that has to be done, in the runtime, before the runtime can return the byte in question.
-
Hi there! I'm writting an app that deals a lot with files (both reading and writting) and that is slowing down my app. What I need is a fast way to do the following operations on files: - read - writte - append - open - close I'm using c++ but inline assembly is an option. Thx!
Hello. I would suggest you to: 1. Get the cluster size of the disk you are using. Then create a buffer of that size. Do all reads and writes (if possible) with the cluster size. 2. Turn of stack checking, at least for the functions you use most frequently. 3. Do not, repeat NOT, use time (and cpu) consuming functions in your code. Especially avoid using the (x)printf functions at all times. It's incredibly time and cpu consuming! Another question: You say there shouldn't be any limitations to the file size. You are aware that the f-funcs has a file size limit of about 4 GB? If you want to avoid that limitaion, the you got to use the real Win32-functions CreateFile, ReadFile etc. If you decide to use them instead, then you have the possibility to use overlapped I/O, which might speed up the file I/O. Else, if you stay with the f-funcs, then considder to use the open(), read(), write(), close() e.t.c. They are closer to the file system than the f-funcs (not much, but its worth trying). Kakan
-
DavidCrow wrote:
A memory-mapped file is a spot in memory that can be accessed (e.g., open, close, read, write, seek) as though it were an actual file.
Sorry, but I think that is supposed to be the other way around. Memory-mapping something allows it to be accessed like normal memory (through a pointer). For example, if you had a memory mapped buttons/switches in your hardware (common on arcade games), you would be able to read the state of the switches by reading values from one or more specific memory addresses. The same goes for files. If you memory-map a file, you get an address that points to some location in the file, and you can then access the contents of that file through the pointer. Peace! -=- James
If you think it costs a lot to do it right, just wait until you find out how much it costs to do it wrong!
Tip for new SUV drivers: Professional Driver on Closed Course does not mean your Dumb Ass on a Public Road!
DeleteFXPFiles & CheckFavorites (Please rate this post!)James R. Twine wrote:
Sorry, but I think that is supposed to be the other way around.
Fair enough. Admittedly I've never used MMF before. Although I did change an application once that was reading a file a few bytes at a time to use
CMemFileinstead. Talk about a major speed improvement! Processing went from hours (some of the files were hundred MB in size) to just a few minutes.
"Take only what you need and leave the land as you found it." - Native American Proverb
-
Using MMF can still improve performance, because you do not have to do the manual loading of data into a buffer - the OS basically does it for you. For example, if you were reading the file in 4KB chunks, you would be allocating a 4K buffer, copying from the file into that buffer, and then likely processing the contents of the buffer using the buffer's address. Using a MMF does that work for you. If you want to impose a limit on the size of the MMF section you want to create, that is fine. Choose a limit, say 2MB (or 4Mb, or 64MB, whatever). If the file is 2MB or smaller, you can MM the entire file. Of not, you can MM 2MB sections of the file one at a time. Peace! -=- James
If you think it costs a lot to do it right, just wait until you find out how much it costs to do it wrong!
Tip for new SUV drivers: Professional Driver on Closed Course does not mean your Dumb Ass on a Public Road!
DeleteFXPFiles & CheckFavorites (Please rate this post!) -
Actually file I/O is never done on a byte level. The least bit of information that can be read from ,or written to a file is one sector. A sector is always an even multiple of 128 bytes, the most common sector size is 512 bytes. So the runtime does buffer (at least) one sector (or more likely, a cluster). I think the main reason for slowing down file reading a byte at the time, is the function call overhead and all the checks that has to be done, in the runtime, before the runtime can return the byte in question.
kakan [[]], a few things on that
kakan wrote:
So the runtime does buffer (at least) one sector (or more likely, a cluster).
Can't you be more precise on which of those does the runtime buffers at a single shot? A sector or a cluster? and What if I read more than just a sector/cluster, will it buffer the necessary sectors/clusters with a single operation or will it take the same amount of time it would if I read the 2, 3, or more sectors/clusters on different operations? thx!
-
Hello. I would suggest you to: 1. Get the cluster size of the disk you are using. Then create a buffer of that size. Do all reads and writes (if possible) with the cluster size. 2. Turn of stack checking, at least for the functions you use most frequently. 3. Do not, repeat NOT, use time (and cpu) consuming functions in your code. Especially avoid using the (x)printf functions at all times. It's incredibly time and cpu consuming! Another question: You say there shouldn't be any limitations to the file size. You are aware that the f-funcs has a file size limit of about 4 GB? If you want to avoid that limitaion, the you got to use the real Win32-functions CreateFile, ReadFile etc. If you decide to use them instead, then you have the possibility to use overlapped I/O, which might speed up the file I/O. Else, if you stay with the f-funcs, then considder to use the open(), read(), write(), close() e.t.c. They are closer to the file system than the f-funcs (not much, but its worth trying). Kakan
Hi there!
kakan wrote:
I would suggest you to: 1. Get the cluster size of the disk you are using. Then create a buffer of that size. Do all reads and writes (if possible) with the cluster size. 2. Turn of stack checking, at least for the functions you use most frequently. 3. Do not, repeat NOT, use time (and cpu) consuming functions in your code. Especially avoid using the (x)printf functions at all times. It's incredibly time and cpu consuming!
By the same order ;) 1. Ok with that. 2. How do i disable the stack checking? 3. Also ok with that, i'm not using them. :) I have also noted that you advise the use of CreateFile, ReadFile, etc instead of f-functions because of theyr limitation. Does that limitation also apply for the open(), read(), write (and so on) functions? And which are faster: Win32-functions or the DOS open/read/open.. ones? Thx! hint_54
-
Hi there!
kakan wrote:
I would suggest you to: 1. Get the cluster size of the disk you are using. Then create a buffer of that size. Do all reads and writes (if possible) with the cluster size. 2. Turn of stack checking, at least for the functions you use most frequently. 3. Do not, repeat NOT, use time (and cpu) consuming functions in your code. Especially avoid using the (x)printf functions at all times. It's incredibly time and cpu consuming!
By the same order ;) 1. Ok with that. 2. How do i disable the stack checking? 3. Also ok with that, i'm not using them. :) I have also noted that you advise the use of CreateFile, ReadFile, etc instead of f-functions because of theyr limitation. Does that limitation also apply for the open(), read(), write (and so on) functions? And which are faster: Win32-functions or the DOS open/read/open.. ones? Thx! hint_54
Hello and good morning. About the stack check, here is a snippet: #pragma check_stack(off) /* Funcs that are called often... */ char * _fastcall CWrTapeTh::w32fgets(char *string, int n) { .... } #pragma check_stack(on) The 4 GB limitation applies to all the old file handling funcs, bot the f-funcs (fopen, fwrite, ...) and open, write. The reason for this limitation is a 32-bit value (unsigned long, I think), that holds the actual position in the file. And that counter wraps at (approx) 4 GB. The Win32-funcs doesn't have that limit of file size. Which one is fastest? To be honest, I don't know, really. But the Win32-funcs are the only way to go if you want to be able to handle files of any size. My guess is that the Win32-funcs can be really fast. Besides, (as I said in my earlier post), the Win32-funcs can use overlapped I/O, which means that you can have several read/writes going on at the same time. Just try to write a file to a diskette with the f-funcs. Get the time for it. Then write the same file to the hard drive. Now, copy that file to the diskette. Get the time for the copy. Compare the times. You will see a remarkable difference. Why? I'm not 100% sure, but my guess is that Windows copy uses overlapped I/O. I know there is samples of overlapped file I/O at MSDN. Maybe I should dig deeper in this and post an article at CP? :) Kakan
-
kakan [[]], a few things on that
kakan wrote:
So the runtime does buffer (at least) one sector (or more likely, a cluster).
Can't you be more precise on which of those does the runtime buffers at a single shot? A sector or a cluster? and What if I read more than just a sector/cluster, will it buffer the necessary sectors/clusters with a single operation or will it take the same amount of time it would if I read the 2, 3, or more sectors/clusters on different operations? thx!