发送大于15GB的文件时,我遇到数据传输速率问题。我有3台服务器和一台客户端。从客户端向服务器发送文件时,我将文件拆分为块(每个块通常为256MB),并在2台服务器上复制每个块。复制发生在管道方法中。发送块时,每个块被切成较小的数据包(每个数据包通常为128 KB),发送到服务器,并在服务器端合并以存储在硬盘驱动器中。这里一切都很好。我测试了系统5GB到50GB的文件,增量为5GB。所有文件的平均写入速度约为600MB /秒。见下图。我在这里与HDFS进行比较。
从服务器读取相同文件时会出现问题。文件分布在多个服务器上。例如,我可以从server1读取block1,从server2读取block2,依此类推。直观地说,读取必须比写入更快,因为客户端并行读取3个服务器。阅读小于15GB {5GB,10GB,15GG}的文件时,性能约为1.1GB /秒。读取大于20GB {20GB,25GB,......,50GB}的文件时会出现问题。随着文件大小的增加,性能会下降。
上图显示了读取50GB文件的基准测试。每个黑点显示单独的块读取时间。如您所见,性能在第60-70块后开始下降。有趣的是,这种情况发生在所有大于15GB的文件中,在同一个地方(约65个区块)左右放慢速度。随着文件大小的增加,缓慢的部分占据主导地位,性能也越来越差。我觉得16GB左右有一些障碍。我看到的唯一提示可能有帮助,即3台服务器随机并行发送块直到65左右。所以阻止'转移是重叠的。之后,一台服务器以循环顺序一次发送。我可以从日志输出中看到这一点。这里仍然存在一些重叠,但没有65个区块之前那么多。
我正在为这个项目使用java 1.8,以及netty 4.1.8。作为tcp服务器。 操作系统是CentOS 7。 每台服务器有两个CPU(Intel(R)Xeon(R)CPU E5-2650 v3 @ 2.30GHz)= 40个核心 64GB RAM 10 GBit以太网。
我花了很多时间,找不到问题的根本原因。 问题可能来自Java VM,Netty,OS,OS TCP默认值或其他原因。
服务器端BlockSenderManager
@Override
public void run(){
while(nodeManager.isRunning()){
try
{
BlockRequest br = blockSenders.take();
if(br != null){
executor.execute(new BlockSender( br, this));
}
if(wait.take())
System.out.println(br.getBlockId()+" Delivered");
}
catch (InterruptedException e)
{
e.printStackTrace();
}
}
服务器端的BlockSender:
@Override
public void run()
{
FileInputStream fis = null;
try
{
java.io.File file = new java.io.File(path+"/" + blockRequest.getBlockId());
fis = new FileInputStream(file);
fSize = file.length();
long rem = fSize;
sendBlockInfo();
int bufSize;
if (fSize < (long) packetSize)
bufSize = (int) fSize;
else
bufSize = packetSize;
int read = 0, packetOrder = 1;
byte[] data;
if(bufSize <= rem)
data = new byte[bufSize];
else
data = new byte[(int)rem];
while ((read = (fis.read(data))) > 0)
{
if (read < 1)
break;
BlockPacket bp = new BlockPacket();
bp.setRequestId(blockRequest.getRequestId());
bp.setBlockId(blockRequest.getBlockId());
bp.setData(data);
bp.setPacketSeqNo(packetOrder);
if(read < bufSize)
{
bp.setIsLastPacket(true);
}
executor.execute(new Sender(bp));
packetOrder++;
if(rem > bufSize)
rem = rem - bufSize;
if(bufSize <= rem)
data = new byte[bufSize];
else
{
data = new byte[(int)rem];
}
}
fis.close();
executor.shutdown();
}
catch (FileNotFoundException e)
{
// TODO Auto-generated catch block
e.printStackTrace();
}
catch (IOException e)
{
// TODO Auto-generated catch block
e.printStackTrace();
}
}
public class Sender implements Runnable
{
private final BlockPacket bp;
private final FileBlock fb;
private DataClient dc;
public Sender(BlockPacket bp)
{
this.bp = bp;
this.fb = null;
dc = getDataClient(requestClient);
}
public Sender(FileBlock fb)
{
this.bp = null;
this.fb = fb;
dc = getDataClient(requestClient);
}
@Override
public void run()
{
if (dc != null)
{
if (bp != null)
{
dc.send(bp);
}
else if (fb != null)
{
dc.send(fb);
}
}
}
}
客户端的ReceivedPacketProcessor
public void processBlockPacket(BlockPacket bp)
{
ByteBuffer buf = ByteBuffer.wrap(bp.getData());
try
{
inChannel.write(buf);
}
catch (IOException e)
{
// TODO Auto-generated catch block
e.printStackTrace();
}
}
@Override
public void run()
{
try
{
aFile = new RandomAccessFile(path+"/"+fileName, "rw");
inChannel = aFile.getChannel();
//java.io.File f = new java.io.File(path+"/"+fileName);
//fop = new FileOutputStream(f);
String reqId = file.getFileID();
currentBlockId = reqId + "_" + currentBlockSeq;
while (true)
{
BlockPacket bp = null;
if (numberOfBlocks > 0)
{
try
{
bp = this.blockingQueue.take();
}
catch (InterruptedException e)
{
e.printStackTrace();
}
if (bp.getBlockId().equals(currentBlockId))
{
if (currentPacket == bp.getPacketSeqNo())
{
if(fileBlocks.containsKey(currentBlockId))
{
processBlockPacket(bp);
if(currentPacket < fileBlocks.get(currentBlockId).getNoOfPackets())
currentPacket++;
else
{
if (fileBlocks.get(currentBlockId).getPackets().size() < 1)
{
removeFileBlock(currentBlockId);
currentBlockSeq++;
currentBlockId = reqId + "_" + currentBlockSeq;
currentPacket = 1;
numberOfBlocks--;
}
}
}
else
{
tempList.add(bp);
}
for(int k =numberOfBlocks; k>0; k--)
{
if(fileBlocks.containsKey(currentBlockId))
{
int pCount = fileBlocks.get(currentBlockId).getNoOfPackets();
int i;
for (i = currentPacket; i <= pCount; i++)
{
if (fileBlocks.get(currentBlockId).getPackets().containsKey(i))
{
processBlockPacket(fileBlocks.get(currentBlockId).getPackets().remove(i));
currentPacket++;
}
else
{
break;
}
}
if(i <= pCount)
{
break;
}
else
{
if (fileBlocks.get(currentBlockId).getPackets().size() < 1)
{
removeFileBlock(currentBlockId);
currentBlockSeq++;
currentBlockId = reqId + "_" + currentBlockSeq;
currentPacket = 1;
numberOfBlocks--;
}
}
}
}
}
}
else
{
if(fileBlocks.containsKey(bp.getBlockId())){
fileBlocks.get(bp.getBlockId()).getPackets().put(bp.getPacketSeqNo(), bp);
}else{
tempList.add(bp);
}
}
}
else{
break;
}
for(int i=0; i<tempList.size(); i++){
if(fileBlocks.containsKey(tempList.get(i).getBlockId())){
BlockPacket temp = tempList.remove(i);
fileBlocks.get(temp.getBlockId()).getPackets().put(temp.getPacketSeqNo(), temp);
}
}
}
System.out.println("CLOSING FILE....");
this.isCompleted.put(true);
inChannel.force(true);
inChannel.close();
}
catch (FileNotFoundException e)
{
// TODO Auto-generated catch block
e.printStackTrace();
}
catch (IOException e)
{
// TODO Auto-generated catch block
e.printStackTrace();
}
catch (InterruptedException e)
{
// TODO Auto-generated catch block
e.printStackTrace();
}
}
启用-XX:+ PrintGCDetails,here is a sample log。
感谢任何评论/帮助。
答案 0 :(得分:0)
这是因为内存中的脏页面比例。由于传入数据速率高于本地IO刷新吞吐量,因此数据在内存中累积。达到最大允许脏页面比率后,接收器不接受更多数据。因此,在这种情况下,系统与本地IO绑定,而不是网络。因此,收益的减少发生在大约15GB。您可以在
中更改一些设置/etc/sysctl.conf中
如:
vm.dirty_background_ratio = 2
vm.dirty_ratio = 80
vm.dirty_expire_centisecs = 3000
vm.dirty_writeback_centisecs = 500
This可能是一本有用的阅读材料。
系统性能仍受本地IO和最大允许脏页率的限制。您可以增加脏页面比率,只能推迟缩短返回时间。如果文件/数据较大,它将再次达到这一点。新结果: