Related
I am working on a sorting project and I've come to the point where a main bottleneck is reading in the data. It takes my program about 20 seconds to sort 100,000,000 integers read in from stdin using cin and std::ios::sync_with_stdio(false); but it turns out that 10 of those seconds is reading in the data to sort. We do know how many integers we will be reading in (the count is at the top of the file we need to sort).
How can I make this faster? I know it's possible because a student in a previous semester was able to do counting sort in a little over 3 seconds (and that's basically purely read time).
The program is just fed the contents of a file with integers separated by newlines like $ ./program < numstosort.txt
Thanks
Here is the relevant code:
std::ios::sync_with_stdio(false);
int max;
cin >> max;
short num;
short* a = new short[max];
int n = 0;
while(cin >> num) {
a[n] = num;
n++;
}
This will get your data into memory about as fast as possible, assuming Linux/POSIX running on commodity hardware. Note that since you apparently aren't allowed to use compiler optimizations, C++ IO is not going to be the fastest way to read data. As others have noted, without optimizations the C++ code will not run anywhere near as fast as it can.
Given that the redirected file is already open as stdin/STDIN_FILENO, use low-level system call/C-style IO. That won't need to be optimized, as it will run just about as fast as possible:
struct stat sb;
int rc = ::fstat( STDIN_FILENO, &sb );
// use C-style calloc() to get memory that's been
// set to zero as calloc() is often optimized to be
// faster than a new followed by a memset().
char *data = (char *)::calloc( 1, sb.st_size + 1 );
size_t totalRead = 0UL;
while ( totalRead < sb.st_size )
{
ssize_t bytesRead = ::read( STDIN_FILENO,
data + totalRead, sb.st_size - totalRead );
if ( bytesRead <= 0 )
{
break;
}
totalRead += bytesRead;
}
// data is now in memory - start processing it
That code will read your data into memory as one long C-style string. And the lack of compiler optimizations won't matter one bit as it's all almost bare-metal system calls.
Using fstat() to get the file size allows allocating all the needed memory at once - no realloc() or copying data around is necessary.
You'll need to add some error checking, and a more robust version of the code would check to be sure the data returned from fstat() actually is a regular file with an actual size, and not a "useless use of cat" such as cat filename | YourProgram, because in that case the fstat() call won't return a useful file size. You'll need to examine the sb.st_mode field of the struct stat after the call to see what the stdin stream really is:
::fstat( STDIN_FILENO, &sb );
...
if ( S_ISREG( sb.st_mode ) )
{
// regular file...
}
(And for really high-performance systems, it can be important to ensure that the memory pages you're reading data into are actually mapped in your process address space. Performance can really stall if data arrives faster than the kernel's memory management system can create virtual-to-physical mappings for the pages data is getting dumped into.)
To handle a large file as fast as possible, you'd want to go multithreaded, with one thread reading data and feeding one or more data processing threads so you can start processing data before you're done reading it.
Edit: parsing the data.
Again, preventing compiler optimizations probably makes the overhead of C++ operations slower than C-style processing. Based on that assumption, something simple will probably run faster.
This would probably work a lot faster in a non-optimized binary, assuming the data is in a C-style string read in as above:
char *next;
long count = ::strtol( data, &next, 0 );
long *values = new long[ count ];
for ( long ii = 0; ii < count; ii++ )
{
values[ ii ] = ::strtol( next, &next, 0 );
}
That is also very fragile. It relies on strtol() skipping over leading whitespace, meaning if there's anything other than whitespace between the numeric values it will fail. It also relies on the initial count of values being correct. Again - that code will fail if that's not true. And because it can replace the value of next before checking for errors, if it ever goes off the rails because of bad data it'll be hopelessly lost.
But it should be about as fast as possible without allowing compiler optimizations.
That's what crazy about not allowing compiler optimizations. You can write simple, robust C++ code to do all your processing, make use of a good optimizing compiler, and probably run almost as fast as the code I posted - which has no error checking and will fail spectacularly in unexpected and undefined ways if fed unexpected data.
You can make it faster if you use a SolidState hard drive. If you want to ask something about code performance, you need to post how are you doing things in the first place.
You may be able to speed up your program by reading the data into a buffer, then converting the text in the buffer to internal representation.
The thought behind this is that all stream devices like to keep streaming. Starting and stopping the stream wastes time. A block read transfers a lot of data with one transaction.
Although cin is buffered, by using cin.read and a buffer, you can make the buffer a lot bigger than cin uses.
If the data has fixed width fields, there are opportunities to speed up the input and conversion processes.
Edit 1: Example
const unsigned int BUFFER_SIZE = 65536;
char text_buffer[BUFFER_SIZE];
//...
cin.read(text_buffer, BUFFER_SIZE);
//...
int value1;
int arguments_scanned = snscanf(&text_buffer, REMAINING_BUFFER_SIZE,
"%d", &value1);
The tricky part is handling the cases where the text of a number is cut off at the end of the buffer.
Can you ran this little test in compare to your test with and without commented line?
#include <iostream>
#include <cstdlib>
int main()
{
std::ios::sync_with_stdio(false);
char buffer[20] {0,};
int t = 0;
while( std::cin.get(buffer, 20) )
{
// t = std::atoi(buffer);
std::cin.ignore(1);
}
return 0;
}
Pure read test:
#include <iostream>
#include <cstdlib>
int main()
{
std::ios::sync_with_stdio(false);
char buffer[1024*1024];
while( std::cin.read(buffer, 1024*1024) )
{
}
return 0;
}
I have a complex interpreter reading in commands from (sometimes) multiples files (the exact details are out of scope) but it requires iterating over these multiple files (some could be GB is size, preventing nice buffering) multiple times.
I am looking to increase the speed of reading in each command from a file.
I have used the RDTSC (program counter) register to micro benchmark the code enough to know about >80% of the time is spent reading in from the files.
Here is the thing: the program that generates the input file is literally faster than to read in the file in my small interpreter. i.e. instead of outputting the file i could (in theory) just link the generator of the data to the interpreter and skip the file but that shouldn't be faster, right?
What am I doing wrong? Or is writing suppose to be 2x to 3x (at least) faster than reading from a file?
I have considered mmap but some of the results on http://lemire.me/blog/archives/2012/06/26/which-is-fastest-read-fread-ifstream-or-mmap/ appear to indicate it is no faster than ifstream. or would mmap help in this case?
details:
I have (so far) tried adding a buffer, tweaking parameters, removing the ifstream buffer (that slowed it down by 6x in my test case), i am currently at a loss for ideas after searching around.
The important section of the code is below. It does the following:
if data is left in buffer, copy form buffer to memblock (where it is then used)
if data is not left in the buffer, check to see how much data is left in the file, if more than the size of the buffer, copy a buffer sized chunk
if less than the file
//if data in buffer
if(leftInBuffer[activefile] > 0)
{
//cout <<bufferloc[activefile] <<"\n";
memcpy(memblock,(buffer[activefile])+bufferloc[activefile],16);
bufferloc[activefile]+=16;
leftInBuffer[activefile]-=16;
}
else //buffers blank
{
//read in block
long blockleft = (cfilemax -cfileplace) / 16 ;
int read=0;
/* slow block starts here */
if(blockleft >= MAXBUFELEMENTS)
{
currentFile->read((char *)(&(buffer[activefile][0])),16*MAXBUFELEMENTS);
leftInBuffer[activefile] = 16*MAXBUFELEMENTS;
bufferloc[activefile]=0;
read =16*MAXBUFELEMENTS;
}
else //read in part of the block
{
currentFile->read((char *)(&(buffer[activefile][0])),16*(blockleft));
leftInBuffer[activefile] = 16*blockleft;
bufferloc[activefile]=0;
read =16*blockleft;
}
/* slow block ends here */
memcpy(memblock,(buffer[activefile])+bufferloc[activefile],16);
bufferloc[activefile]+=16;
leftInBuffer[activefile]-=16;
}
edit: this is on a mac, osx 10.9.5, with an i7 with a SSD
Solution:
as was suggested below, mmap was able to increase the speed by about 10x.
(for anyone else who searches for this)
specifically open with:
uint8_t * openMMap(string name, long & size)
{
int m_fd;
struct stat statbuf;
uint8_t * m_ptr_begin;
if ((m_fd = open(name.c_str(), O_RDONLY)) < 0)
{
perror("can't open file for reading");
}
if (fstat(m_fd, &statbuf) < 0)
{
perror("fstat in openMMap failed");
}
if ((m_ptr_begin = (uint8_t *)mmap(0, statbuf.st_size, PROT_READ, MAP_SHARED, m_fd, 0)) == MAP_FAILED)
{
perror("mmap in openMMap failed");
}
uint8_t * m_ptr = m_ptr_begin;
size = statbuf.st_size;
return m_ptr;
}
read by:
uint8_t * mmfile = openMMap("my_file", length);
uint32_t * memblockmm;
memblockmm = (uint32_t *)mmfile; //cast file to uint32 array
uint32_t data = memblockmm[0]; //take int
mmfile +=4; //increment by 4 as I read a 32 bit entry and each entry in mmfile is 8 bits.
This should be a comment, but I don't have 50 reputation to make a comment.
What is the value of MAXBUFELEMENTS? From my experience, many smaller reads is far slower than one read of larger size. I suggest to read the entire file in if possible, some files could be GBs, but even reading in 100MB at once would perform better than reading 1 MB 100 times.
If that's still not good enough, next thing you can try is to compress(zlib) input files(may have to break them into chunks due to size), and decompress them in memory. This method is usually faster than reading in uncompressed files.
As #Tony Jiang said, try experimenting with the buffer size to see if that helps.
Try mmap to see if that helps.
I assume that currentFile is a std::ifstream? There's going to be some overhead for using iostreams (for example, an istream will do its own buffering, adding an extra layer to what you're doing); although I wouldn't expect the overhead to be huge, you can test by using open(2) and read(2) directly.
You should be able to run your code through dtruss -e to verify how long the read system calls take. If those take the bulk of your time, then you're hitting OS and hardware limits, so you can address that by piping, mmap'ing, or adjusting your buffer size. If those take less time than you expect, then look for problems in your application logic (unnecessary work on each iteration, etc.).
Does anyone know how to read in a file with raw encoding? So stumped.... I am trying to read in floats or doubles (I think). I have been stuck on this for a few weeks. Thank you!
File that I am trying to read from:
http://www.sci.utah.edu/~gk/DTI-data/gk2/gk2-rcc-mask.raw
Description of raw encoding:
hello://teem.sourceforge.net/nrrd/format.html#encoding (change hello to http to go to page)
- "raw" - The data appears on disk exactly the same as in memory, in terms of byte values and byte ordering. Produced by write() and fwrite(), suitable for read() or fread().
Info of file:
http://www.sci.utah.edu/~gk/DTI-data/gk2/gk2-rcc-mask.nhdr - I think the only things that matter here are the big endian (still trying to understand what that means from google) and raw encoding.
My current approach, uncertain if it's correct:
//Function ripped off from example of c++ ifstream::read reference page
void scantensor(string filename){
ifstream tdata(filename, ifstream::binary); // not sure if I should put ifstream::binary here
// other things I tried
// ifstream tdata(filename) ifstream tdata(filename, ios::in)
if(tdata){
tdata.seekg(0, tdata.end);
int length = tdata.tellg();
tdata.seekg(0, tdata.beg);
char* buffer = new char[length];
tdata.read(buffer, length);
tdata.close();
double* d;
d = (double*) buffer;
} else cerr << "failed" << endl;
}
/* P.S. I attempted to print the first 100 elements of the array.
Then I print 100 other elements at some arbitrary array indices (i.e. 9,900 - 10,000). I actually kept increasing the number of 0's until I ran out of bound at 100,000,000 (I don't think that's how it works lol but I was just playing around to see what happens)
Here's the part that makes me suspicious: so the ifstream different has different constructors like the ones I tried above.
the first 100 values are always the same.
if I use ifstream::binary, then I get some values for the 100 arbitrary printing
if I use the other two options, then I get -6.27744e+066 for all 100 of them
So for now I am going to assume that ifstream::binary is the correct one. The thing is, I am not sure if the file I provided is how binary files actually look like. I am also unsure if these are the actual numbers that I am supposed to read in or just casting gone wrong. I do realize that my casting from char* to double* can be unsafe, and I got that from one of the threads.
*/
I really appreciate it!
Edit 1: Right now the data being read in using the above method is apparently "incorrect" since in paraview the values are:
Dxx,Dxy,Dxz,Dyy,Dyz,Dzz
[0, 1], [-15.4006, 13.2248], [-5.32436, 5.39517], [-5.32915, 5.96026], [-17.87, 19.0954], [-6.02961, 5.24771], [-13.9861, 14.0524]
It's a 3 x 3 symmetric matrix, so 7 distinct values, 7 ranges of values.
The floats that I am currently parsing from the file right now are very large (i.e. -4.68855e-229, -1.32351e+120).
Perhaps somebody knows how to extract the floats from Paraview?
Since you want to work with doubles, I recommend to read the data from file as buffer of doubles:
const long machineMemory = 0x40000000; // 1 GB
FILE* file = fopen("c:\\data.bin", "rb");
if (file)
{
int size = machineMemory / sizeof(double);
if (size > 0)
{
double* data = new double[size];
int read(0);
while (read = fread(data, sizeof(double), size, file))
{
// Process data here (read = number of doubles)
}
delete [] data;
}
fclose(file);
}
I'm trying to copy a file, but whatever I try, the copy seems to be a few bytes short.
_file is an ifstream set to binary mode.
void FileProcessor::send()
{
//If no file is opened return
if(!_file.is_open()) return;
//Reset position to beginning
_file.seekg(0, ios::beg);
//Result buffer
char * buffer;
char * partBytes = new char[_bufferSize];
//Packet *p;
//Read the file and send it over the network
while(_file.read(partBytes,_bufferSize))
{
//buffer = Packet::create(Packet::FILE,std::string(partBytes));
//p = Packet::create(buffer);
//cout<< p->getLength() << "\n";
//writeToFile(p->getData().c_str(),p->getLength());
writeToFile(partBytes,_bufferSize);
//delete[] buffer;
}
//cout<< *p << "\n";
delete [] partBytes;
}
_writeFile is the file to be written to.
void FileProcessor::writeToFile(const char *buffer,unsigned int size)
{
if(_writeFile.is_open())
{
_writeFile.write(buffer,size);
_writeFile.flush();
}
}
In this case I'm trying to copy a zip file.
But opening both the original and copy in notepad I noticed that while they look identical , It's different at the end where the copy is missing a few bytes.
Any suggestions?
You are assuming that the file's size is a multiple of _bufferSize. You have to check what's left on the buffer after the while:
while(_file.read(partBytes,_bufferSize)) {
writeToFile(partBytes,_bufferSize);
}
if(_file.gcount())
writeToFile(partBytes, _file.gcount());
Your while loop will terminate when it fails to read _bufferSize bytes because it hits an EOF.
The final call to read() might have read some data (just not a full buffer) but your code ignores it.
After your loop you need to check _file.gcount() and if it is not zero, write those remaining bytes out.
Are you copying from one type of media to another? Perhaps different sector sizes are causing the apparent weirdness.
What if _bufferSize doesn't divide evenly into the size of the file...that might cause extra bytes to be written.
You don't want to always do writeToFile(partBytes,_bufferSize); since it's possible (at the end) that less than _bufferSize bytes were read. Also, as pointed out in the comments on this answer, the ifstream is no longer "true" once the EOF is reached, so the last chunk isn't copied (this is your posted problem). Instead, use gcount() to get the number of bytes read:
do
{
_file.read(partBytes, _bufferSize);
writeToFile(partBytes, (unsigned int)_file.gcount());
} while (_file);
For comparisons of zip files, you might want to consider using a non-text editor to do the comparison; HxD is a great (free) hex editor with a file compare option.
Profiling my program and the function print is taking a lot of time to perform. How can I send "raw" byte output directly to stdout instead of using fwrite, and making it faster (need to send all 9bytes in the print() at the same time to the stdout) ?
void print(){
unsigned char temp[9];
temp[0] = matrix[0][0];
temp[1] = matrix[0][1];
temp[2] = matrix[0][2];
temp[3] = matrix[1][0];
temp[4] = matrix[1][1];
temp[5] = matrix[1][2];
temp[6] = matrix[2][0];
temp[7] = matrix[2][1];
temp[8] = matrix[2][2];
fwrite(temp,1,9,stdout);
}
Matrix is defined globally to be a unsigned char matrix[3][3];
IO is not an inexpensive operation. It is, in fact, a blocking operation, meaning that the OS can preempt your process when you call write to allow more CPU-bound processes to run, before the IO device you're writing to completes the operation.
The only lower level function you can use (if you're developing on a *nix machine), is to use the raw write function, but even then your performance will not be that much faster than it is now. Simply put: IO is expensive.
The top rated answer claims that IO is slow.
Here's a quick benchmark with a sufficiently large buffer to take the OS out of the critical performance path, but only if you're willing to receive your output in giant blurps. If latency to first byte is your problem, you need to run in "dribs" mode.
Write 10 million records from a nine byte array
Mint 12 AMD64 on 3GHz CoreDuo under gcc 4.6.1
340ms to /dev/null
710ms to 90MB output file
15254ms to 90MB output file in "dribs" mode
FreeBSD 9 AMD64 on 2.4GHz CoreDuo under clang 3.0
450ms to /dev/null
550ms to 90MB output file on ZFS triple mirror
1150ms to 90MB output file on FFS system drive
22154ms to 90MB output file in "dribs" mode
There's nothing slow about IO if you can afford to buffer properly.
#include <stdio.h>
#include <assert.h>
#include <stdlib.h>
#include <string.h>
int main (int argc, char* argv[])
{
int dribs = argc > 1 && 0==strcmp (argv[1], "dribs");
int err;
int i;
enum { BigBuf = 4*1024*1024 };
char* outbuf = malloc (BigBuf);
assert (outbuf != NULL);
err = setvbuf (stdout, outbuf, _IOFBF, BigBuf); // full line buffering
assert (err == 0);
enum { ArraySize = 9 };
char temp[ArraySize];
enum { Count = 10*1000*1000 };
for (i = 0; i < Count; ++i) {
fwrite (temp, 1, ArraySize, stdout);
if (dribs) fflush (stdout);
}
fflush (stdout); // seems to be needed after setting own buffer
fclose (stdout);
if (outbuf) { free (outbuf); outbuf = NULL; }
}
The rawest form of output you can do is the probable the write system call, like this
write (1, matrix, 9);
1 is the file descriptor for standard out (0 is standard in, and 2 is standard error). Your standard out will only write as fast as the one reading it at the other end (i.e. the terminal, or the program you're pipeing into) which might be rather slow.
I'm not 100% sure, but you could try setting non-blocking IO on fd 1 (using fcntl) and hope the OS will buffer it for you until it can be consumed by the other end. It's been a while, but I think it works like this
fcntl (1, F_SETFL, O_NONBLOCK);
YMMV though. Please correct me if I'm wrong on the syntax, as I said, it's been a while.
Perhaps your problem is not that fwrite() is slow, but that it is buffered.
Try calling fflush(stdout) after the fwrite().
This all really depends on your definition of slow in this context.
All printing is fairly slow, although iostreams are really slow for printing.
Your best bet would be to use printf, something along the lines of:
printf("%c%c%c%c%c%c%c%c%c\n", matrix[0][0], matrix[0][1], matrix[0][2], matrix[1][0],
matrix[1][1], matrix[1][2], matrix[2][0], matrix[2][1], matrix[2][2]);
As everyone has pointed out IO in tight inner loop is expensive. I have normally ended up doing conditional cout of Matrix based on some criteria when required to debug it.
If your app is console app then try redirecting it to a file, it will be lot faster than doing console refreshes. e.g app.exe > matrixDump.txt
What's wrong with:
fwrite(matrix,1,9,stdout);
both the one and the two dimensional arrays take up the same memory.
Try running the program twice. Once with output and once without. You will notice that overall, the one without the io is the fastest. Also, you could fork the process (or create a thread), one writing to a file(stdout), and one doing the operations.
So first, don't print on every entry. Basically what i am saying is do not do like that.
for(int i = 0; i<100; i++){
printf("Your stuff");
}
instead allocate a buffer either on stack or on heap, and store you infomration there and then just throw this bufffer into stdout, just liek that
char *buffer = malloc(sizeof(100));
for(int i = 100; i<100; i++){
char[i] = 1; //your 8 byte value goes here
}
//once you are done print it to a ocnsole with
write(1, buffer, 100);
but in your case, just use write(1, temp, 9);
I am pretty sure you can increase the output performance by increasing the buffer size. So you have less fwrite calls. write might be faster but I am not sure. Just try this:
❯ yes | dd of=/dev/null count=1000000
1000000+0 records in
1000000+0 records out
512000000 bytes (512 MB, 488 MiB) copied, 2.18338 s, 234 MB/s
vs
> yes | dd of=/dev/null count=100000 bs=50KB iflag=fullblock
100000+0 records in
100000+0 records out
5000000000 bytes (5.0 GB, 4.7 GiB) copied, 2.63986 s, 1.9 GB/s
The same applies to your code. Some tests during the last days show that probably good buffer sizes are around 1 << 12 (=4096) and 1<<16 (=65535) bytes.
You can simply:
std::cout << temp;
printf is more C-Style.
Yet, IO operations are costly, so use them wisely.