转自：http://blog.csdn.net/stark_summer/article/details/44174381
未实验
最近一个群友的boss让研究hbase，让hbase的入库速度达到5w+/s，这可愁死了，4台个人电脑组成的集群，多线程入库调了好久，速度也才1w左右，都没有达到理想的那种速度，然后就想到了这种方式，但是网上多是用mapreduce来实现入库，而现在的需求是实时入库，不生成文件了，所以就只能自己用代码实现了，但是网上查了很多资料都没有查到，最后在一个网友的指引下，看了源码，最后找到了生成Hfile的方式，实现了之后，发现单线程入库速度才达到1w4左右，和之前的多线程的全速差不多了，百思不得其解之时，调整了一下代码把列的Byte.toBytes(cols)这个方法调整出来只做一次，速度立马就到3w了，提升非常明显，这是我的电脑上的速度，估计在它的集群上能更快一点吧，下面把代码和大家分享一下。
String tableName = "taglog";
byte[] family = Bytes.toBytes("logs");
//配置文件设置
Configuration conf = HBaseConfiguration.create();
conf.set("hbase.master", "192.168.1.133:60000");
conf.set("hbase.zookeeper.quorum", "192.168.1.135");
//conf.set("zookeeper.znode.parent", "/hbase");
conf.set("hbase.metrics.showTableName", "false");
//conf.set("io.compression.codecs", "org.apache.hadoop.io.compress.SnappyCodec");
String outputdir = "hdfs://hadoop.Master:8020/user/SEA/hfiles/";
Path dir = new Path(outputdir);
Path familydir = new Path(outputdir, Bytes.toString(family));
FileSystem fs = familydir.getFileSystem(conf);
BloomType bloomType = BloomType.NONE;
final HFileDataBlockEncoder encoder = NoOpDataBlockEncoder.INSTANCE;
int blockSize = 64000;
Configuration tempConf = new Configuration(conf);
tempConf.set("hbase.metrics.showTableName", "false");
tempConf.setFloat(HConstants.HFILE_BLOCK_CACHE_SIZE_KEY, 1.0f);
//实例化HFile的Writer，StoreFile实际上只是HFile的轻量级的封装
StoreFile.Writer writer = new StoreFile.WriterBuilder(conf, new CacheConfig(tempConf),
fs, blockSize)
.withOutputDir(familydir)
.withCompression(Compression.Algorithm.NONE)
.withBloomType(bloomType).withComparator(KeyValue.COMPARATOR)
.withDataBlockEncoder(encoder).build();
long start = System.currentTimeMillis();
DecimalFormat df = new DecimalFormat("0000000");
KeyValue kv1 = null;
KeyValue kv2 = null;
KeyValue kv3 = null;
KeyValue kv4 = null;
KeyValue kv5 = null;
KeyValue kv6 = null;
KeyValue kv7 = null;
KeyValue kv8 = null;
//这个是耗时操作，只进行一次
byte[] cn = Bytes.toBytes("cn");
byte[] dt = Bytes.toBytes("dt");
byte[] ic = Bytes.toBytes("ic");
byte[] ifs = Bytes.toBytes("if");
byte[] ip = Bytes.toBytes("ip");
byte[] le = Bytes.toBytes("le");
byte[] mn = Bytes.toBytes("mn");
byte[] pi = Bytes.toBytes("pi");
int maxLength = 3000000;
for(int i=0;i<maxLength;i++){
String currentTime = ""+System.currentTimeMillis() + df.format(i);
long current = System.currentTimeMillis();
//rowkey和列都要按照字典序的方式顺序写入，否则会报错的
kv1 = new KeyValue(Bytes.toBytes(currentTime),
family, cn,current,KeyValue.Type.Put,Bytes.toBytes("3"));
kv2 = new KeyValue(Bytes.toBytes(currentTime),
family, dt,current,KeyValue.Type.Put,Bytes.toBytes("6"));
kv3 = new KeyValue(Bytes.toBytes(currentTime),
family, ic,current,KeyValue.Type.Put,Bytes.toBytes("8"));
kv4 = new KeyValue(Bytes.toBytes(currentTime),
family, ifs,current,KeyValue.Type.Put,Bytes.toBytes("7"));
kv5 = new KeyValue(Bytes.toBytes(currentTime),
family, ip,current,KeyValue.Type.Put,Bytes.toBytes("4"));
kv6 = new KeyValue(Bytes.toBytes(currentTime),
family, le,current,KeyValue.Type.Put,Bytes.toBytes("2"));
kv7 = new KeyValue(Bytes.toBytes(currentTime),
family, mn,current,KeyValue.Type.Put,Bytes.toBytes("5"));
kv8 = new KeyValue(Bytes.toBytes(currentTime),
family,pi,current,KeyValue.Type.Put,Bytes.toBytes("1"));
writer.append(kv1);
writer.append(kv2);
writer.append(kv3);
writer.append(kv4);
writer.append(kv5);
writer.append(kv6);
writer.append(kv7);
writer.append(kv8);
}
writer.close();
//把生成的HFile导入到hbase当中
HTable table = new HTable(conf,tableName);
LoadIncrementalHFiles loader = new LoadIncrementalHFiles(conf);
loader.doBulkLoad(dir, table);
最后再附上查看hfile的方式，查询正确的hfile和自己生成的hfile，方便查找问题。
hbase org.apache.hadoop.hbase.io.hfile.HFile -p -f hdfs://hadoop.Master:8020/user/SEA/hfiles/logs/51aa97b2a25446f89d5c870af92c9fc1
