我從java1.3開始學習java，后來主要用1.4，再后來1.5和1.6中的很多新特性，都停留在“知道”的狀態，比如nio，雖然據說可以提升性能，但并沒有真正深入使用和測試過，工作操作文件的情況不多，所以關注也不多，即便用到，也還是習慣性的用java.io。今天看到的這篇文章，雖然測試手段非常簡單，所得結論也難免有些片面 ，但依然說明，在順序訪問的時候，NIO的性能相對java.io有很大的提升。
也許應該update一下自己的知識了，否則就要OUT，或者早已經OUT了。
下次操作文件或者寫socket要用NIO了。
---
轉自：http://links.techwebnewsletters.com/ctt?kn=28&m=34038811&r=MzI1Mjc3MDAzOAS2&b=0&j=NTc5NjM4MTAS1&mt=1&rt=0
以下為翻譯的內容：
最近我在工作中用到了java i/o相關功能。因為對java.io的了解更多（畢竟面世較早），所以一開始我使用的是java.io包下的類，后來為了測試一下是不是能夠通過NIO提高文件操作性能，于是轉向了java.nio。我得到的結論讓我感到有些震驚，下面是對比測試的一些細節：
1、在java.io的測試代碼中，我使用RandomAccessFile直接向文件寫數據，并搜索到特定的位置執行記錄的插入、讀取和刪除。
2、在java.nio的初步測試代碼中，使用FileChannel對象。NIO之所以比java.io更加高效，是因為NIO面向的是data chunks，而java.io基本上是面向byte的。
3、為了進一步挖掘NIO的能力，我又改用MappedByteBuffer執行測試，這個類是構建在操作系統的虛擬內存機制上的。根據java文檔所說，這個類在性能方面是最好的。
為了進行測試，我寫了一個模擬員工數據庫的小程序，員工數據的結構如下：
view plaincopy to clipboardprint?
class
Employee {
String last; // the key
String first;
int id;
int zip;
boolean employed;
String comments;
}
class
Employee {
String last; // the key
String first;
int id;
int zip;
boolean employed;
String comments;
}
員工數據寫入文件，并將last name作為索引key，日后可以通過這個key從文件中加載該員工對應的數據。無論使用IO、NIO還是MappedByteBuffers，首先都需要打開一個RandomAccessFile。以下代碼在用戶的home目錄下創建一個名為employee.ejb的文件，設置為可讀可寫，并初始化對應的Channel和MappedByteBuffer：
view plaincopy to clipboardprint?
String
userHome = System.getProperty("user.home");
StringBuffer pathname = new StringBuffer(userHome);
pathname.append(File.separator);
pathname.append("employees.ejb");
java.io.RandomAccessFile journal =
new RandomAccessFile(pathname.toString(), "rw");
//下面這一句是為了NIO
java.nio.channels.FileChannel channel = journal.getChannel();
//下面這兩句是為了使用MappedByteBuffer
journal.setLength(PAGE_SIZE);
MappedByteBuffer mbb =
channel.map(FileChannel.MapMode.READ_WRITE, 0, journal.length() );
String
userHome = System.getProperty("user.home");
StringBuffer pathname = new StringBuffer(userHome);
pathname.append(File.separator);
pathname.append("employees.ejb");
java.io.RandomAccessFile journal =
new RandomAccessFile(pathname.toString(), "rw");
//下面這一句是為了NIO
java.nio.channels.FileChannel channel = journal.getChannel();
//下面這兩句是為了使用MappedByteBuffer
journal.setLength(PAGE_SIZE);
MappedByteBuffer mbb =
channel.map(FileChannel.MapMode.READ_WRITE, 0, journal.length() );
使用channel.map進行映射后，當該文件被追加了新的數據時，之前的MappedByteBuffer是看不到這些數據的。因為我們想測試讀和寫，所以當文件中追加寫入新的記錄后，需要重新做映射才能使得MappedByteBuffer讀取新數據。為了提高效率，降低重新映射的次數，每次空間不夠的時候，我們將文件擴張特定的大小（比如說1K）以防止每次追加新記錄都要重新映射。
下面是寫入員工記錄的對比測試：
使用java.io的代碼：
view plaincopy to clipboardprint?
public
boolean addRecord_IO(Employee emp) {
try {
byte[] last = emp.last.getBytes();
byte[] first = emp.first.getBytes();
byte[] comments = emp.comments.getBytes();
// Just hard-code the sizes for perfomance
int size = 0;
size += emp.last.length();
size += 4; // strlen - Integer
size += emp.first.length();
size += 4; // strlen - Integer
size += 4; // emp.id - Integer
size += 4; // emp.zip - Integer
size += 1; // emp.employed - byte
size += emp.comments.length();
size += 4; // strlen - Integer
long offset = getStorageLocation(size);
//
// Store the record by key and save the offset
//
if ( offset == -1 ) {
// We need to add to the end of the journal. Seek there
// now only if we're not already there
long currentPos = journal.getFilePointer();
long jounralLen = journal.length();
if ( jounralLen != currentPos )
journal.seek(jounralLen);
offset = jounralLen;
}else {
// Seek to the returned insertion point
journal.seek(offset);
}
// Fist write the header
journal.writeByte(1);
journal.writeInt(size);
// Next write the data
journal.writeInt(last.length);
journal.write(last);
journal.writeInt(first.length);
journal.write(first);
journal.writeInt(emp.id);
journal.writeInt(emp.zip);
if ( emp.employed )
journal.writeByte(1);
else
journal.writeByte(0);
journal.writeInt(comments.length);
journal.write(comments);
// Next, see if we need to append an empty record if we
inserted
// this new record at an empty location
if ( newEmptyRecordSize != -1 ) {
// Simply write a header
journal.writeByte(0); //inactive record
journal.writeLong(newEmptyRecordSize);
}
employeeIdx.put(emp.last, offset);
return true;
}
catch ( Exception e ) {
e.printStackTrace();
}
return false;
}
public
boolean addRecord_IO(Employee emp) {
try {
byte[] last = emp.last.getBytes();
byte[] first = emp.first.getBytes();
byte[] comments = emp.comments.getBytes();
// Just hard-code the sizes for perfomance
int size = 0;
size += emp.last.length();
size += 4; // strlen - Integer
size += emp.first.length();
size += 4; // strlen - Integer
size += 4; // emp.id - Integer
size += 4; // emp.zip - Integer
size += 1; // emp.employed - byte
size += emp.comments.length();
size += 4; // strlen - Integer
long offset = getStorageLocation(size);
//
// Store the record by key and save the offset
//
if ( offset == -1 ) {
// We need to add to the end of the journal. Seek there
// now only if we're not already there
long currentPos = journal.getFilePointer();
long jounralLen = journal.length();
if ( jounralLen != currentPos )
journal.seek(jounralLen);
offset = jounralLen;
}else {
// Seek to the returned insertion point
journal.seek(offset);
}
// Fist write the header
journal.writeByte(1);
journal.writeInt(size);
// Next write the data
journal.writeInt(last.length);
journal.write(last);
journal.writeInt(first.length);
journal.write(first);
journal.writeInt(emp.id);
journal.writeInt(emp.zip);
if ( emp.employed )
journal.writeByte(1);
else
journal.writeByte(0);
journal.writeInt(comments.length);
journal.write(comments);
// Next, see if we need to append an empty record if we inserted
// this new record at an empty location
if ( newEmptyRecordSize != -1 ) {
// Simply write a header
journal.writeByte(0); //inactive record
journal.writeLong(newEmptyRecordSize);
}
employeeIdx.put(emp.last, offset);
return true;
}
catch ( Exception e ) {
e.printStackTrace();
}
return false;
}
使用java.nio的代碼：
view plaincopy to clipboardprint?
public
boolean addRecord_NIO(Employee emp) {
try {
data.clear();
byte[] last = emp.last.getBytes();
byte[] first = emp.first.getBytes();
byte[] comments = emp.comments.getBytes();
data.putInt(last.length);
data.put(last);
data.putInt(first.length);
data.put(first);
data.putInt(emp.id);
data.putInt(emp.zip);
byte employed = 0;
if ( emp.employed )
employed = 1;
data.put(employed);
data.putInt(comments.length);
data.put(comments);
data.flip();
int dataLen = data.limit();
header.clear();
header.put((byte)1); // 1=active record
header.putInt(dataLen);
header.flip();
long headerLen = header.limit();
int length = (int)(headerLen + dataLen);
long offset = getStorageLocation((int)dataLen);
//
// Store the record by key and save the offset
//
if ( offset == -1 ) {
// We need to add to the end of the journal. Seek there
// now only if we're not already there
long currentPos = channel.position();
long jounralLen = channel.size();
if ( jounralLen != currentPos )
channel.position(jounralLen);
offset = jounralLen;
}
else {
// Seek to the returned insertion point
channel.position(offset);
}
// Fist write the header
long written = channel.write(srcs);
// Next, see if we need to append an empty record if we
inserted
// this new record at an empty location
if ( newEmptyRecordSize != -1 ) {
// Simply write a header
data.clear();
data.put((byte)0);
data.putInt(newEmptyRecordSize);
data.flip();
channel.write(data);
}
employeeIdx.put(emp.last, offset);
return true;
}
catch ( Exception e ) {
e.printStackTrace();
}
return false;
}
public
boolean addRecord_NIO(Employee emp) {
try {
data.clear();
byte[] last = emp.last.getBytes();
byte[] first = emp.first.getBytes();
byte[] comments = emp.comments.getBytes();
data.putInt(last.length);
data.put(last);
data.putInt(first.length);
data.put(first);
data.putInt(emp.id);
data.putInt(emp.zip);
byte employed = 0;
if ( emp.employed )
employed = 1;
data.put(employed);
data.putInt(comments.length);
data.put(comments);
data.flip();
int dataLen = data.limit();
header.clear();
header.put((byte)1); // 1=active record
header.putInt(dataLen);
header.flip();
long headerLen = header.limit();
int length = (int)(headerLen + dataLen);
long offset = getStorageLocation((int)dataLen);
//
// Store the record by key and save the offset
//
if ( offset == -1 ) {
// We need to add to the end of the journal. Seek there
// now only if we're not already there
long currentPos = channel.position();
long jounralLen = channel.size();
if ( jounralLen != currentPos )
channel.position(jounralLen);
offset = jounralLen;
}
else {
// Seek to the returned insertion point
channel.position(offset);
}
// Fist write the header
long written = channel.write(srcs);
// Next, see if we need to append an empty record if we inserted
// this new record at an empty location
if ( newEmptyRecordSize != -1 ) {
// Simply write a header
data.clear();
data.put((byte)0);
data.putInt(newEmptyRecordSize);
data.flip();
channel.write(data);
}
employeeIdx.put(emp.last, offset);
return true;
}
catch ( Exception e ) {
e.printStackTrace();
}
return false;
}
使用MappedByteBuffer的代碼如下:
view plaincopy to clipboardprint?
public
boolean addRecord_MBB(Employee emp) {
try {
byte[] last = emp.last.getBytes();
byte[] first = emp.first.getBytes();
byte[] comments = emp.comments.getBytes();
int datalen = last.length + first.length + comments.length + 12 +
9;
int headerlen = 5;
int length = headerlen + datalen;
//
// Store the record by key and save the offset
//
long offset = getStorageLocation(datalen);
if ( offset == -1 ) {
// We need to add to the end of the journal. Seek there
// now only if we're not already there
long currentPos = mbb.position();
long journalLen = channel.size();
if ( (currentPos+length) >= journalLen ) {
//log("GROWING FILE BY ANOTHER PAGE");
mbb.force();
journal.setLength(journalLen + PAGE_SIZE);
channel = journal.getChannel();
journalLen = channel.size();
mbb = channel.map(FileChannel.MapMode.READ_WRITE, 0,
journalLen);
currentPos = mbb.position();
}
if ( currentEnd != currentPos )
mbb.position(currentEnd);
offset = currentEnd;//journalLen;
}
else {
// Seek to the returned insertion point
mbb.position((int)offset);
}
// write header
mbb.put((byte)1); // 1=active record
mbb.putInt(datalen);
// write data
mbb.putInt(last.length);
mbb.put(last);
mbb.putInt(first.length);
mbb.put(first);
mbb.putInt(emp.id);
mbb.putInt(emp.zip);
byte employed = 0;
if ( emp.employed )
employed = 1;
mbb.put(employed);
mbb.putInt(comments.length);
mbb.put(comments);
currentEnd += length;
// Next, see if we need to append an empty record if we
inserted
// this new record at an empty location
if ( newEmptyRecordSize != -1 ) {
// Simply write a header
mbb.put((byte)0);
mbb.putInt(newEmptyRecordSize);
currentEnd += 5;
}
employeeIdx.put(emp.last, offset);
return true;
}
catch ( Exception e ) {
e.printStackTrace();
}
return false;
}
public
boolean addRecord_MBB(Employee emp) {
try {
byte[] last = emp.last.getBytes();
byte[] first = emp.first.getBytes();
byte[] comments = emp.comments.getBytes();
int datalen = last.length + first.length + comments.length + 12 + 9;
int headerlen = 5;
int length = headerlen + datalen;
//
// Store the record by key and save the offset
//
long offset = getStorageLocation(datalen);
if ( offset == -1 ) {
// We need to add to the end of the journal. Seek there
// now only if we're not already there
long currentPos = mbb.position();
long journalLen = channel.size();
if ( (currentPos+length) >= journalLen ) {
//log("GROWING FILE BY ANOTHER PAGE");
mbb.force();
journal.setLength(journalLen + PAGE_SIZE);
channel = journal.getChannel();
journalLen = channel.size();
mbb = channel.map(FileChannel.MapMode.READ_WRITE, 0, journalLen);
currentPos = mbb.position();
}
if ( currentEnd != currentPos )
mbb.position(currentEnd);
offset = currentEnd;//journalLen;
}
else {
// Seek to the returned insertion point
mbb.position((int)offset);
}
// write header
mbb.put((byte)1); // 1=active record
mbb.putInt(datalen);
// write data
mbb.putInt(last.length);
mbb.put(last);
mbb.putInt(first.length);
mbb.put(first);
mbb.putInt(emp.id);
mbb.putInt(emp.zip);
byte employed = 0;
if ( emp.employed )
employed = 1;
mbb.put(employed);
mbb.putInt(comments.length);
mbb.put(comments);
currentEnd += length;
// Next, see if we need to append an empty record if we inserted
// this new record at an empty location
if ( newEmptyRecordSize != -1 ) {
// Simply write a header
mbb.put((byte)0);
mbb.putInt(newEmptyRecordSize);
currentEnd += 5;
}
employeeIdx.put(emp.last, offset);
return true;
}
catch ( Exception e ) {
e.printStackTrace();
}
return false;
}
接下來，調用每種方法插入100,000條記錄, 耗時對比如下:
* With java.io: ~10,000 milliseconds
* With java.nio: ~2,000 milliseconds
* With MappedByteBuffer: ~970 milliseconds
使用NIO的性能改善效果非常明顯，使用MappedByteBuffer的性能，更是讓人吃驚。
使用三種方式讀取數據的性能對比如下：
* With java.io: ~6,900 milliseconds
* With java.nio: ~1,400 milliseconds
* With MappedByteBuffer: ~355 milliseconds
和寫入的時候情況差不多，NIO有很明顯的性能提升，而MappedByteBuffer則有驚人的高效率。從java.io遷移到nio并使用MappedByteBuffer，通常可以獲得10倍以上的性能提升。
源文檔 <http://blog.csdn.net/sean1203/archive/2010/01/06/5142464.aspx>