Is there a workaround for Java's poor performance on walking huge directories?

Question

I am trying to process files one at a time that are stored over a network. Reading the files is fast due to buffering is not the issue. The problem I have is just listing the directories in a folder. I have at least 10k files per folder over many folders.

Performance is super slow since File.list() returns an array instead of an iterable. Java goes off and collects all the names in a folder and packs it into an array before returning.

The bug entry for this is https://bugs.java.com/bugdatabase/view_bug;jsessionid=db7fcf25bcce13541c4289edeb4?bug_id=4285834 and doesn't have a work around. They just say this has been fixed for JDK7.

A few questions:

1. Does anybody have a workaround to this performance bottleneck?
2. Am I trying to achieve the impossible? Is performance still going to be poor even if it just iterates over the directories?
3. Could I use the beta JDK7 builds that have this functionality without having to build my entire project on it?

Bassorilievo

What's your target architecture? Do you need to maintain cross-platform execution? – Isosteric

I realize that this is a very old question, but for anyone using the latest JDK7, this functionality is now available via the Files.newDirectoryStream() API in the java.nio.file package. – Ambitendency

On a side note, we were able to speed up performance dramatically by using subdirectories extensively to limit the number of files/directories in any one single directory. Made certain file operations like file checking and writing much much faster but certain things like du and rm -rf did take longer. It's a trade off, depending on what is more important for your use case. – Muncy

Answer 1

Although it's not pretty, I solved this kind of problem once by piping the output of dir/ls to a file before starting my app, and passing in the filename.

If you needed to do it within the app, you could just use system.exec(), but it would create some nastiness.

You asked. The first form is going to be blazingly fast, the second should be pretty fast as well.

Be sure to do the one item per line (bare, no decoration, no graphics), full path and recurse options of your selected command.

EDIT:

30 minutes just to get a directory listing, wow.

It just struck me that if you use exec(), you can get it's stdout redirected into a pipe instead of writing it to a file.

If you did that, you should start getting the files immediately and be able to begin processing before the command has completed.

The interaction may actually slow things down, but maybe not--you might give it a try.

Wow, I just went to find the syntax of the .exec command for you and came across this, possibly exactly what you want (it lists a directory using exec and "ls" and pipes the result into your program for processing): good link in wayback (Jörg provided in a comment to replace this one from sun that Oracle broke)

Anyway, the idea is straightforward but getting the code right is annoying. I'll go steal some codes from the internets and hack them up--brb

/**
 * Note: Only use this as a last resort!  It's specific to windows and even
 * at that it's not a good solution, but it should be fast.
 * 
 * to use it, extend FileProcessor and call processFiles("...") with a list
 * of options if you want them like /s... I highly recommend /b
 * 
 * override processFile and it will be called once for each line of output.
 */
import java.io.*;

public abstract class FileProcessor
{
   public void processFiles(String dirOptions)
   {
      Process theProcess = null;
      BufferedReader inStream = null;

      // call the Hello class
      try
      {
          theProcess = Runtime.getRuntime().exec("cmd /c dir " + dirOptions);
      }
      catch(IOException e)
      {
         System.err.println("Error on exec() method");
         e.printStackTrace();  
      }

      // read from the called program's standard output stream
      try
      {
         inStream = new BufferedReader(
                                new InputStreamReader( theProcess.getInputStream() ));  
         processFile(inStream.readLine());
      }
      catch(IOException e)
      {
         System.err.println("Error on inStream.readLine()");
         e.printStackTrace();  
      }

   } // end method
   /** Override this method--it will be called once for each file */
   public abstract void processFile(String filename);


} // end class

And thank you code donor at IBM

Orange

this is a good solution for static trees or trees that don't change often, but you need to transverse it often – Bassorilievo

Yeah, I used it for cases where it was a run-once batch type of program. If you did it often, well I'm not sure that using exec() to recreate that file wouldn't work--but regardless it's an inelegant, error-prone solution. – Orange

I just tried this and a "dir /B /S ..." took about 30 mins to collect over 200MB of paths, while it took about 3.5 hours to transverse the same tree using File.list() – Bassorilievo

I am going to accept this solution, as it will probably solve the issue most of the time. – Bassorilievo

Bill: Mention that you must use a flag which tells dir/ls not to sort the results (so start getting names immediately): "ls -U". There doesn't seem to be such an option for "dir" :/ – Myosotis

@BillK: the link in your answer is dead. Any way you can add some details in the answer? – Courlan

Okay, now I officially hate Oracle. – Orange

@BillK there is a backup of your broken link available using the wayback machine here: web.archive.org/web/20080912133319/http://java.sun.com/… – Deglutinate

@AaronDigulla dir /b does the trick (under Windows 7, in 2014 though) – Sigh

Answer 2

How about using File.list(FilenameFilter filter) method and implementing FilenameFilter.accept(File dir, String name) to process each file and return false.

I ran this on Linux vm for directory with 10K+ files and it took <10 seconds.

import java.io.File;  
import java.io.FilenameFilter;

public class Temp {
    private static void processFile(File dir, String name) {
        File file = new File(dir, name);
        System.out.println("processing file " + file.getName());
    }

    private static void forEachFile(File dir) {
        String [] ignore = dir.list(new FilenameFilter() {
            public boolean accept(File dir, String name) {
                processFile(dir, name);
                return false;
            }
        });
    }

    public static void main(String[] args) {
        long before, after;
        File dot = new File(".");
        before = System.currentTimeMillis();
        forEachFile(dot);
        after = System.currentTimeMillis();
        System.out.println("after call, delta is " + (after - before));
    }  
}

Advantage

in windows, listing a remote folder form >20 seconds went to 500ms... +1 – Uranous

Answer 3

An alternative is to have the files served over a different protocol. As I understand you're using SMB for that and java is just trying to list them as a regular file.

The problem here might not be java alone ( how does it behaves when you open that directory with Microsoft Explorer x:\shared ) In my experience it also take a considerably amount of time.

You can change the protocol to something like HTTP, only to fetch the file names. This way you can retrieve the list of files over http ( 10k lines should't be too much ) and let the server deal with file listing. This would be very fast, since it will run with local resources ( those in the server )

Then when you have the list, you can process them one by exactly the way you're doing right now.

The keypoint is to have an aid mechanism in the other side of the node.

Is this feasible?

Today:

File [] content = new File("X:\\remote\\dir").listFiles();

for ( File f : content ) {
    process( f );
}

Proposed:

String [] content = fetchViaHttpTheListNameOf("x:\\remote\\dir");

for ( String fileName : content ) {
    process( new File( fileName ) );
}

The http server could be a very small small and simple file.

If this is the way you have it right now, what you're doing is to fetch all the 10k files information to your client machine ( I don't know how much of that info ) when you only need the file name for later processing.

If the processing is very fast right now it may be slowed down a bit. This is because the information prefetched is no longer available.

Give it a try.

Grati

that implies a lot of things though: a webserver, implementing some kind of security that SMB/etc already have, and so on. – Fredkin

Not really. Is like have a LAN website isn't? I think this is an already solved issue. Although now that you mention it, I have no idea how to setup a website that could be visible only in the LAN. I'm pretty sure a .exe that server files and uses Windows authentication already exists. – Grati

Isn't IIS support this kind of functionality?... I think it can authenticate window users very easily. But I agree, It is not transparent. However I find it far easier / faster than wrap native code via JNI and still using SMB as the protocol. – Grati

Answer 4

A non-portable solution would be to make native calls to the operating system and stream the results.

For Linux

You can look at something like readdir. You can walk the directory structure like a linked list and return results in batches or individually.

For Windows

In windows the behavior would be fairly similar using FindFirstFile and FindNextFile apis.

Huba

for now, would agree that you might want to wrap FFF/FNF as suggested here. – Fredkin

Bear in mind that bridging the native barrier will have performance impacts itself. If you take this approach, you might want to consider batching several FNF results per JNI call (or possibly use a DirectByteBuffer for the data transfer) – Cryptonymous

+1 for suggesting FFF/FNF. I have posted a working example below. – Bo

Answer 5

I doubt the problem is relate to the bug report you referenced. The issue there is "only" memory usage, but not necessarily speed. If you have enough memory the bug is not relevant for your problem.

You should measure whether your problem is memory related or not. Turn on your Garbage Collector log and use for example gcviewer to analyze your memory usage.

I suspect that it has to do with the SMB protocol causing the problem. You can try to write a test in another language and see if it's faster, or you can try to get the list of filenames through some other method, such as described here in another post.

Vaules

Answer 6

If you need to eventually process all files, then having Iterable over String[] won't give you any advantage, as you'll still have to go and fetch the whole list of files.

Gunnery

Answer 7

If you're on Java 1.5 or 1.6, shelling out "dir" commands and parsing the standard output stream on Windows is a perfectly acceptable approach. I've used this approach in the past for processing network drives and it has generally been a lot faster than waiting for the native java.io.File listFiles() method to return.

Of course, a JNI call should be faster and potentially safer than shelling out "dir" commands. The following JNI code can be used to retrieve a list of files/directories using the Windows API. This function can be easily refactored into a new class so the caller can retrieve file paths incrementally (i.e. get one path at a time). For example, you can refactor the code so that FindFirstFileW is called in a constructor and have a seperate method to call FindNextFileW.

JNIEXPORT jstring JNICALL Java_javaxt_io_File_GetFiles(JNIEnv *env, jclass, jstring directory)
{
    HANDLE hFind;
    try {

      //Convert jstring to wstring
        const jchar *_directory = env->GetStringChars(directory, 0);
        jsize x = env->GetStringLength(directory);
        wstring path;  //L"C:\\temp\\*";
        path.assign(_directory, _directory + x);
        env->ReleaseStringChars(directory, _directory);

        if (x<2){
            jclass exceptionClass = env->FindClass("java/lang/Exception");
            env->ThrowNew(exceptionClass, "Invalid path, less than 2 characters long.");
        }

        wstringstream ss;
        BOOL bContinue = TRUE;
        WIN32_FIND_DATAW data;
        hFind = FindFirstFileW(path.c_str(), &data);
        if (INVALID_HANDLE_VALUE == hFind){
            jclass exceptionClass = env->FindClass("java/lang/Exception");
            env->ThrowNew(exceptionClass, "FindFirstFileW returned invalid handle.");
        }


        //HANDLE hStdOut = GetStdHandle(STD_OUTPUT_HANDLE);
        //DWORD dwBytesWritten;


        // If we have no error, loop thru the files in this dir
        while (hFind && bContinue){

          /*
          //Debug Print Statment. DO NOT DELETE! cout and wcout do not print unicode correctly.
            WriteConsole(hStdOut, data.cFileName, (DWORD)_tcslen(data.cFileName), &dwBytesWritten, NULL);
            WriteConsole(hStdOut, L"\n", 1, &dwBytesWritten, NULL);
            */

          //Check if this entry is a directory
            if (data.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY){
                // Make sure this dir is not . or ..
                if (wstring(data.cFileName) != L"." &&
                    wstring(data.cFileName) != L"..")
                {   
                    ss << wstring(data.cFileName) << L"\\" << L"\n";
                }
            }
            else{
                ss << wstring(data.cFileName) << L"\n";
            }
            bContinue = FindNextFileW(hFind, &data);
        }   
        FindClose(hFind); // Free the dir structure



        wstring cstr = ss.str();
        int len = cstr.size();
        //WriteConsole(hStdOut, cstr.c_str(), len, &dwBytesWritten, NULL);
        //WriteConsole(hStdOut, L"\n", 1, &dwBytesWritten, NULL);
        jchar* raw = new jchar[len];
        memcpy(raw, cstr.c_str(), len*sizeof(wchar_t));
        jstring result = env->NewString(raw, len);
        delete[] raw;
        return result;
    }
    catch(...){
        FindClose(hFind);
        jclass exceptionClass = env->FindClass("java/lang/Exception");
        env->ThrowNew(exceptionClass, "Exception occured.");
    }

    return NULL;
}

Credit: https://sites.google.com/site/jozsefbekes/Home/windows-programming/miscellaneous-functions

Even with this approach, there are still efficiencies to be gained. If you serialize the path to a java.io.File, there is a huge performance hit - especially if the path represents a file on a network drive. I have no idea what Sun/Oracle is doing under the hood but if you need additional file attributes other than the file path (e.g. size, mod date, etc), I have found that the following JNI function is much faster than instantiating a java.io.File object on a network the path.

JNIEXPORT jlongArray JNICALL Java_javaxt_io_File_GetFileAttributesEx(JNIEnv *env, jclass, jstring filename)
{   

  //Convert jstring to wstring
    const jchar *_filename = env->GetStringChars(filename, 0);
    jsize len = env->GetStringLength(filename);
    wstring path;
    path.assign(_filename, _filename + len);
    env->ReleaseStringChars(filename, _filename);


  //Get attributes
    WIN32_FILE_ATTRIBUTE_DATA fileAttrs;
    BOOL result = GetFileAttributesExW(path.c_str(), GetFileExInfoStandard, &fileAttrs);
    if (!result) {
        jclass exceptionClass = env->FindClass("java/lang/Exception");
        env->ThrowNew(exceptionClass, "Exception Occurred");
    }

  //Create an array to store the WIN32_FILE_ATTRIBUTE_DATA
    jlong buffer[6];
    buffer[0] = fileAttrs.dwFileAttributes;
    buffer[1] = date2int(fileAttrs.ftCreationTime);
    buffer[2] = date2int(fileAttrs.ftLastAccessTime);
    buffer[3] = date2int(fileAttrs.ftLastWriteTime);
    buffer[4] = fileAttrs.nFileSizeHigh;
    buffer[5] = fileAttrs.nFileSizeLow;

    jlongArray jLongArray = env->NewLongArray(6);
    env->SetLongArrayRegion(jLongArray, 0, 6, buffer);
    return jLongArray;
}

You can find a full working example of this JNI-based approach in the javaxt-core library. In my tests using Java 1.6.0_38 with a Windows host hitting a Windows share, I have found this JNI approach approximately 10x faster then calling java.io.File listFiles() or shelling out "dir" commands.

Bo

Answer 8

I wonder why there are 10k files in a directory. Some file systems do not work well with so many files. There are specifics limitations for file systems like max amount of files per directory and max amount of levels of subdirectory.

I solve a similar problem with an iterator solution.

I needed to walk across huge directorys and several levels of directory tree recursively.

I try FileUtils.iterateFiles() of Apache commons io. But it implement the iterator by adding all the files in a List and then returning List.iterator(). It's very bad for memory.

So I prefer to write something like this:

private static class SequentialIterator implements Iterator<File> {
    private DirectoryStack dir = null;
    private File current = null;
    private long limit;
    private FileFilter filter = null;

    public SequentialIterator(String path, long limit, FileFilter ff) {
        current = new File(path);
        this.limit = limit;
        filter = ff;
        dir = DirectoryStack.getNewStack(current);
    }

    public boolean hasNext() {
        while(walkOver());
        return isMore && (limit > count || limit < 0) && dir.getCurrent() != null;
    }

    private long count = 0;

    public File next() {
        File aux = dir.getCurrent();
        dir.advancePostition();
        count++;
        return aux;
    }

    private boolean walkOver() {
        if (dir.isOutOfDirListRange()) {
            if (dir.isCantGoParent()) {
                isMore = false;
                return false;
            } else {
                dir.goToParent();
                dir.advancePostition();
                return true;
            }
        } else {
            if (dir.isCurrentDirectory()) {
                if (dir.isDirectoryEmpty()) {
                    dir.advancePostition();
                } else {
                    dir.goIntoDir();
                }
                return true;
            } else {
                if (filter.accept(dir.getCurrent())) {
                    return false;
                } else {
                    dir.advancePostition();
                    return true;
                }
            }
        }
    }

    private boolean isMore = true;

    public void remove() {
        throw new UnsupportedOperationException();
    }

}

Note that the iterator stop by an amount of files iterateds and it has a FileFilter also.

And DirectoryStack is:

public class DirectoryStack {
    private class Element{
        private File files[] = null;
        private int currentPointer;
        public Element(File current) {
            currentPointer = 0;
            if (current.exists()) {
                if(current.isDirectory()){
                    files = current.listFiles();
                    Set<File> set = new TreeSet<File>();
                    for (int i = 0; i < files.length; i++) {
                        File file = files[i];
                        set.add(file);
                    }
                    set.toArray(files);
                }else{
                    throw new IllegalArgumentException("File current must be directory");
                }
            } else {
                throw new IllegalArgumentException("File current not exist");
            }

        }
        public String toString(){
            return "current="+getCurrent().toString();
        }
        public int getCurrentPointer() {
            return currentPointer;
        }
        public void setCurrentPointer(int currentPointer) {
            this.currentPointer = currentPointer;
        }
        public File[] getFiles() {
            return files;
        }
        public File getCurrent(){
            File ret = null;
            try{
                ret = getFiles()[getCurrentPointer()];
            }catch (Exception e){
            }
            return ret;
        }
        public boolean isDirectoryEmpty(){
            return !(getFiles().length>0);
        }
        public Element advancePointer(){
            setCurrentPointer(getCurrentPointer()+1);
            return this;
        }
    }
    private DirectoryStack(File first){
        getStack().push(new Element(first));
    }
    public static DirectoryStack getNewStack(File first){
        return new DirectoryStack(first);
    }
    public String toString(){
        String ret = "stack:\n";
        int i = 0;
        for (Element elem : stack) {
            ret += "nivel " + i++ + elem.toString()+"\n";
        }
        return ret;
    }
    private Stack<Element> stack=null;
    private Stack<Element> getStack(){
        if(stack==null){
            stack = new Stack<Element>();
        }
        return stack;
    }
    public File getCurrent(){
        return getStack().peek().getCurrent();
    }
    public boolean isDirectoryEmpty(){
        return getStack().peek().isDirectoryEmpty();
    }
    public DirectoryStack downLevel(){
        getStack().pop();
        return this;
    }
    public DirectoryStack goToParent(){
        return downLevel();
    }
    public DirectoryStack goIntoDir(){
        return upLevel();
    }
    public DirectoryStack upLevel(){
        if(isCurrentNotNull())
            getStack().push(new Element(getCurrent()));
        return this;
    }
    public DirectoryStack advancePostition(){
        getStack().peek().advancePointer();
        return this;
    }
    public File[] peekDirectory(){
        return getStack().peek().getFiles();
    }
    public boolean isLastFileOfDirectory(){
        return getStack().peek().getFiles().length <= getStack().peek().getCurrentPointer();
    }
    public boolean gotMoreLevels() {
        return getStack().size()>0;
    }
    public boolean gotMoreInCurrentLevel() {
        return getStack().peek().getFiles().length > getStack().peek().getCurrentPointer()+1;
    }
    public boolean isRoot() {
        return !(getStack().size()>1);
    }
    public boolean isCurrentNotNull() {
        if(!getStack().isEmpty()){
            int currentPointer = getStack().peek().getCurrentPointer();
            int maxFiles = getStack().peek().getFiles().length;
            return currentPointer < maxFiles;
        }else{
            return false;
        }
    }
    public boolean isCurrentDirectory() {
        return getStack().peek().getCurrent().isDirectory();
    }
    public boolean isLastFromDirList() {
        return getStack().peek().getCurrentPointer() == (getStack().peek().getFiles().length-1);
    }
    public boolean isCantGoParent() {
        return !(getStack().size()>1);
    }
    public boolean isOutOfDirListRange() {
        return getStack().peek().getFiles().length <= getStack().peek().getCurrentPointer();
    }

}

Forb

? You use listFiles() anyways, which would end up with the same performance. You code does a whole lot of nothing. listFiles().iterator() would have been the same. – Bassorilievo

If you want to walk recursively through 5 levels of subdirectory with 10.000 files per directory listFiles.iterator() it kill your memory. For several files in the same directory it's the same. – Forb

Answer 9

Using an Iterable doesn't imply that the Files will be streamed to you. In fact its usually the opposite. So an array is typically faster than an Iterable.

Carrell

Answer 10

Are you sure it's due to Java, not just a general problem with having 10k entries in one directory, particularly over the network?

Have you tried writing a proof-of-concept program to do the same thing in C using the win32 findfirst/findnext functions to see whether it's any faster?

I don't know the ins and outs of SMB, but I strongly suspect that it needs a round trip for every file in the list - which is not going to be fast, particularly over a network with moderate latency.

Having 10k strings in an array sounds like something which should not tax the modern Java VM too much either.

Romie