在Java中的已排序(内存映射?)文件中进行二进制搜索

时间:2009-04-10 02:39:27

标签: java nio large-files binary-search memory-mapping

我正在努力将Perl程序移植到Java,并且随时学习Java。原始程序的核心组件是Perl module,它使用二进制搜索在+500 GB排序文本文件中执行字符串前缀查找 (基本上,“寻找”到文件中间的字节偏移,回溯到最近的换行,比较行前缀和搜索字符串,“搜索”到那个字节偏移的一半/两倍,重复直到找到...)

我已经尝试了几种数据库解决方案,但发现没有什么比这个大小的数据集的查找速度快。您知道任何实现此类功能的现有Java库吗?如果做不到这一点,你能否指出我在文本文件中进行随机访问读取的一些惯用示例代码?

或者,我不熟悉新的(?)Java I / O库,但它是一个内存映射500 GB文本文件的选项(我在64位机器上备用内存)并在内存映射字节数组上进行二进制搜索?我很想知道你必须分享的关于这个和类似问题的任何经验。

8 个答案:

答案 0 :(得分:29)

对于像这样的情况,我是Java MappedByteBuffers 粉丝。它非常快。下面是我为你准备的一个片段,它将缓冲区映射到文件,向中间搜索,然后向后搜索换行符。这应该足以让你去?

我在自己的应用程序中有类似的代码(搜索,读取,重复直到完成),进行基准测试 java.io在生产环境中对MappedByteBuffer进行了流式处理,并将结果发布在我的博客(Geekomatic posts tagged 'java.nio')上,包含原始数据,图表和所有内容。

两秒钟总结? 我基于MappedByteBuffer的实施速度提高了约275%。 YMMV。

要处理大于~2GB的文件,这是因为强制转换和.position(int pos)而导致的问题,我制作了由MappedByteBuffer数组支持的分页算法。您需要在64位系统上工作才能使用大于2-4GB的文件,因为MBB使用操作系统的虚拟内存系统来实现他们的魔力。

public class StusMagicLargeFileReader  {
    private static final long PAGE_SIZE = Integer.MAX_VALUE;
    private List<MappedByteBuffer> buffers = new ArrayList<MappedByteBuffer>();
    private final byte raw[] = new byte[1];

    public static void main(String[] args) throws IOException {
        File file = new File("/Users/stu/test.txt");
        FileChannel fc = (new FileInputStream(file)).getChannel(); 
        StusMagicLargeFileReader buffer = new StusMagicLargeFileReader(fc);
        long position = file.length() / 2;
        String candidate = buffer.getString(position--);
        while (position >=0 && !candidate.equals('\n')) 
            candidate = buffer.getString(position--);
        //have newline position or start of file...do other stuff    
    }
    StusMagicLargeFileReader(FileChannel channel) throws IOException {
        long start = 0, length = 0;
        for (long index = 0; start + length < channel.size(); index++) {
            if ((channel.size() / PAGE_SIZE) == index)
                length = (channel.size() - index *  PAGE_SIZE) ;
            else
                length = PAGE_SIZE;
            start = index * PAGE_SIZE;
            buffers.add(index, channel.map(READ_ONLY, start, length));
        }    
    }
    public String getString(long bytePosition) {
        int page  = (int) (bytePosition / PAGE_SIZE);
        int index = (int) (bytePosition % PAGE_SIZE);
        raw[0] = buffers.get(page).get(index);
        return new String(raw);
    }
}

答案 1 :(得分:3)

我有同样的问题。我试图在排序文件中找到以某些前缀开头的所有行。

这是我编写的一个方法,它主要是Python代码的端口:http://www.logarithmic.net/pfh/blog/01186620415

我测试了它但尚未彻底测试。但它不使用内存映射。

public static List<String> binarySearch(String filename, String string) {
    List<String> result = new ArrayList<String>();
    try {
        File file = new File(filename);
        RandomAccessFile raf = new RandomAccessFile(file, "r");

        long low = 0;
        long high = file.length();

        long p = -1;
        while (low < high) {
            long mid = (low + high) / 2;
            p = mid;
            while (p >= 0) {
                raf.seek(p);

                char c = (char) raf.readByte();
                //System.out.println(p + "\t" + c);
                if (c == '\n')
                    break;
                p--;
            }
            if (p < 0)
                raf.seek(0);
            String line = raf.readLine();
            //System.out.println("-- " + mid + " " + line);
            if (line.compareTo(string) < 0)
                low = mid + 1;
            else
                high = mid;
        }

        p = low;
        while (p >= 0) {
            raf.seek(p);
            if (((char) raf.readByte()) == '\n')
                break;
            p--;
        }

        if (p < 0)
            raf.seek(0);

        while (true) {
            String line = raf.readLine();
            if (line == null || !line.startsWith(string))
                break;
            result.add(line);
        }

        raf.close();
    } catch (IOException e) {
        System.out.println("IOException:");
        e.printStackTrace();
    }
    return result;
}

答案 2 :(得分:2)

我不知道任何具有该功能的库。但是,Java中外部二进制搜索的正确代码应与此类似:

class ExternalBinarySearch {
final RandomAccessFile file;
final Comparator<String> test; // tests the element given as search parameter with the line. Insert a PrefixComparator here
public ExternalBinarySearch(File f, Comparator<String> test) throws FileNotFoundException {
    this.file = new RandomAccessFile(f, "r");
    this.test = test;
}
public String search(String element) throws IOException {
    long l = file.length();
    return search(element, -1, l-1);
}
/**
 * Searches the given element in the range [low,high]. The low value of -1 is a special case to denote the beginning of a file.
 * In contrast to every other line, a line at the beginning of a file doesn't need a \n directly before the line
 */
private String search(String element, long low, long high) throws IOException {
    if(high - low < 1024) {
        // search directly
        long p = low;
        while(p < high) {
            String line = nextLine(p);
            int r = test.compare(line,element);
            if(r > 0) {
                return null;
            } else if (r < 0) {
                p += line.length();
            } else {
                return line;
            }
        }
        return null;
    } else {
        long m  = low + ((high - low) / 2);
        String line = nextLine(m);
        int r = test.compare(line, element);
        if(r > 0) {
            return search(element, low, m);
        } else if (r < 0) {
            return search(element, m, high);
        } else {
            return line;
        }
    }
}
private String nextLine(long low) throws IOException {
    if(low == -1) { // Beginning of file
        file.seek(0);           
    } else {
        file.seek(low);
    }
    int bufferLength = 65 * 1024;
    byte[] buffer = new byte[bufferLength];
    int r = file.read(buffer);
    int lineBeginIndex = -1;

    // search beginning of line
    if(low == -1) { //beginning of file
        lineBeginIndex = 0;
    } else {
        //normal mode
        for(int i = 0; i < 1024; i++) {
        if(buffer[i] == '\n') {
            lineBeginIndex = i + 1;
            break;
        }
        }
    }
    if(lineBeginIndex == -1) {
        // no line begins within next 1024 bytes
        return null;
    }
    int start = lineBeginIndex;
        for(int i = start; i < r; i++) {
            if(buffer[i] == '\n') {
                // Found end of line
                return new String(buffer, lineBeginIndex, i - lineBeginIndex + 1);
                return line.toString();
            }
        }
        throw new IllegalArgumentException("Line to long");
}
}

请注意:我特意编写了这个代码:角落案例的测试不够好,代码假设没有单行大于64K等。

我还认为构建线条开始的偏移量的索引可能是个好主意。对于500 GB文件,该索引应存储在索引文件中。您应该使用该索引获得一个不那么小的常数因子,因为不需要在每个步骤中搜索下一行。

我知道这不是问题,但是建立像(Patrica)Tries(在磁盘/ SSD上)的前缀树数据结构可能是一个好主意进行前缀搜索。

答案 3 :(得分:1)

这是您想要实现的一个简单示例。我可能首先索引文件,跟踪每个字符串的文件位置。我假设字符串由换行符(或回车符)分隔:

    RandomAccessFile file = new RandomAccessFile("filename.txt", "r");
    List<Long> indexList = new ArrayList();
    long pos = 0;
    while (file.readLine() != null)
    {
        Long linePos = new Long(pos);
        indexList.add(linePos);
        pos = file.getFilePointer();
    }
    int indexSize = indexList.size();
    Long[] indexArray = new Long[indexSize];
    indexList.toArray(indexArray);

最后一步是转换为数组,以便在进行大量查找时略微提高速度。我可能也会将Long[]转换为long[],但我没有在上面显示。最后是从给定索引位置读取字符串的代码:

    int i; // Initialize this appropriately for your algorithm.
    file.seek(indexArray[i]);
    String line = file.readLine();
            // At this point, line contains the string #i.

答案 4 :(得分:1)

如果您正在处理500GB文件,那么您可能希望使用比二进制搜索更快的查找方法 - 即基数排序,它实质上是散列的变体。这样做的最佳方法实际上取决于您的数据分布和查找类型,但如果您正在寻找字符串前缀,那么应该有一个很好的方法。

我发布了一个整数基数排序解决方案的例子,但你可以使用相同的想法 - 基本上通过将数据划分为桶来减少排序时间,然后使用O(1)查找来检索数据桶这是相关的。

Option Strict On
Option Explicit On

Module Module1

Private Const MAX_SIZE As Integer = 100000
Private m_input(MAX_SIZE) As Integer
Private m_table(MAX_SIZE) As List(Of Integer)
Private m_randomGen As New Random()
Private m_operations As Integer = 0

Private Sub generateData()
    ' fill with random numbers between 0 and MAX_SIZE - 1
    For i = 0 To MAX_SIZE - 1
        m_input(i) = m_randomGen.Next(0, MAX_SIZE - 1)
    Next

End Sub

Private Sub sortData()
    For i As Integer = 0 To MAX_SIZE - 1
        Dim x = m_input(i)
        If m_table(x) Is Nothing Then
            m_table(x) = New List(Of Integer)
        End If
        m_table(x).Add(x)
        ' clearly this is simply going to be MAX_SIZE -1
        m_operations = m_operations + 1
    Next
End Sub

 Private Sub printData(ByVal start As Integer, ByVal finish As Integer)
    If start < 0 Or start > MAX_SIZE - 1 Then
        Throw New Exception("printData - start out of range")
    End If
    If finish < 0 Or finish > MAX_SIZE - 1 Then
        Throw New Exception("printData - finish out of range")
    End If
    For i As Integer = start To finish
        If m_table(i) IsNot Nothing Then
            For Each x In m_table(i)
                Console.WriteLine(x)
            Next
        End If
    Next
End Sub

' run the entire sort, but just print out the first 100 for verification purposes
Private Sub test()
    m_operations = 0
    generateData()
    Console.WriteLine("Time started = " & Now.ToString())
    sortData()
    Console.WriteLine("Time finished = " & Now.ToString & " Number of operations = " & m_operations.ToString())
    ' print out a random 100 segment from the sorted array
    Dim start As Integer = m_randomGen.Next(0, MAX_SIZE - 101)
    printData(start, start + 100)
End Sub

Sub Main()
    test()
    Console.ReadLine()
End Sub

End Module

答案 5 :(得分:0)

我有类似的问题,所以我从这个帖子中提供的解决方案中创建了(Scala)库:

https://github.com/avast/BigMap

它包含在此已排序文件中对大文件和二进制搜索进行排序的实用程序...

答案 6 :(得分:0)

我发布了一个要点https://gist.github.com/mikee805/c6c2e6a35032a3ab74f643a1d0f8249c

这是相当完整的例子,基于我在堆栈溢出时发现的内容和一些博客希望其他人可以使用它

import static java.nio.file.Files.isWritable;
import static java.nio.file.StandardOpenOption.READ;
import static org.apache.commons.io.FileUtils.forceMkdir;
import static org.apache.commons.io.IOUtils.closeQuietly;
import static org.apache.commons.lang3.StringUtils.isBlank;
import static org.apache.commons.lang3.StringUtils.trimToNull;

import java.io.File;
import java.io.IOException;
import java.nio.Buffer;
import java.nio.MappedByteBuffer;
import java.nio.channels.FileChannel;
import java.nio.file.Path;

public class FileUtils {

    private FileUtils() {
    }

    private static boolean found(final String candidate, final String prefix) {
        return isBlank(candidate) || candidate.startsWith(prefix);
    }

    private static boolean before(final String candidate, final String prefix) {
        return prefix.compareTo(candidate.substring(0, prefix.length())) < 0;
    }

    public static MappedByteBuffer getMappedByteBuffer(final Path path) {
        FileChannel fileChannel = null;
        try {
            fileChannel = FileChannel.open(path, READ);
            return fileChannel.map(FileChannel.MapMode.READ_ONLY, 0, fileChannel.size()).load();
        } 
        catch (Exception e) {
            throw new RuntimeException(e);
        }
        finally {
            closeQuietly(fileChannel);
        }
    }

    public static String binarySearch(final String prefix, final MappedByteBuffer buffer) {
        if (buffer == null) {
            return null;
        }
        try {
            long low = 0;
            long high = buffer.limit();
            while (low < high) {
                int mid = (int) ((low + high) / 2);
                final String candidate = getLine(mid, buffer);
                if (found(candidate, prefix)) {
                    return trimToNull(candidate);
                } 
                else if (before(candidate, prefix)) {
                    high = mid;
                } 
                else {
                    low = mid + 1;
                }
            }
        } 
        catch (Exception e) {
            throw new RuntimeException(e);
        } 
        return null;
    }

    private static String getLine(int position, final MappedByteBuffer buffer) {
        // search backwards to the find the proceeding new line
        // then search forwards again until the next new line
        // return the string in between
        final StringBuilder stringBuilder = new StringBuilder();
        // walk it back
        char candidate = (char)buffer.get(position);
        while (position > 0 && candidate != '\n') {
            candidate = (char)buffer.get(--position);
        }
        // we either are at the beginning of the file or a new line
        if (position == 0) {
            // we are at the beginning at the first char
            candidate = (char)buffer.get(position);
            stringBuilder.append(candidate);
        }
        // there is/are char(s) after new line / first char
        if (isInBuffer(buffer, position)) {
            //first char after new line
            candidate = (char)buffer.get(++position);
            stringBuilder.append(candidate);
            //walk it forward
            while (isInBuffer(buffer, position) && candidate != ('\n')) {
                candidate = (char)buffer.get(++position);
                stringBuilder.append(candidate);
            }
        }
        return stringBuilder.toString();
    }

    private static boolean isInBuffer(final Buffer buffer, int position) {
        return position + 1 < buffer.limit();
    }

    public static File getOrCreateDirectory(final String dirName) { 
        final File directory = new File(dirName);
        try {
            forceMkdir(directory);
            isWritable(directory.toPath());
        } 
        catch (IOException e) {
            throw new RuntimeException(e);
        }
        return directory;
    }
}

答案 7 :(得分:-1)

如果您真的想尝试对文件进行内存映射,我在Java nio中找到了tutorial on how to use memory mapping