[HBASE-10676] Removing ThreadLocal of PrefetchedHeader in HFileBlock.FSReaderV2 make higher perforamce of scan

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 0.99.0
• Assignee: Unassigned
• Reporter: zhaojianbo
• Created: 05/Mar/14 03:38
• Updated: 17/Jun/22 05:10
• Resolved: 23/Nov/16 02:54

Description

PrefetchedHeader variable in HFileBlock.FSReaderV2 is used for avoiding backward seek operation as the comment said:

we will not incur a backward seek operation if we have already read this block's header as part of the previous read's look-ahead. And we also want to skip reading the header again if it has already been read.

But that is not the case. In the code of 0.98, prefetchedHeader is threadlocal for one storefile reader, and in the RegionScanner lifecycle，different rpc handlers will serve scan requests of the same scanner. Even though one handler of previous scan call prefetched the next block header, the other handlers of current scan call will still trigger a backward seek operation. The process is like this:

1. rs handler1 serves the scan call, reads block1 and prefetches the header of block2
2. rs handler2 serves the same scanner's next scan call, because rs handler2 doesn't know the header of block2 already prefetched by rs handler1, triggers a backward seek and reads block2, and prefetches the header of block3.

It is not the sequential read. So I think that the threadlocal is useless, and should be abandoned. I did the work, and evaluated the performance of one client, two client and four client scanning the same region with one storefile. The test environment is

1. A hdfs cluster with a namenode, a secondary namenode , a datanode in a machine
2. A hbase cluster with a zk, a master, a regionserver in the same machine
3. clients are also in the same machine.
   So all the data is local. The storefile is about 22.7GB from our online data, 18995949 kvs. Caching is set 1000. And setCacheBlocks(false)
   With the improvement, the client total scan time decreases 21% for the one client case, 11% for the two clients case. But the four clients case is almost the same. The details tests' data is the following:

   case	client	time(ms)
   original	1	306222
   new	1	241313
   original	2	416390
   new	2	369064
   original	4	555986
   new	4	562152

With some modification(see the comments below), the newest result is

case	client	time(ms)	case	client	time(ms)	case	client	time(ms)
original	1	306222	new with synchronized	1	239510	new with AtomicReference	1	241243
original	2	416390	new with synchronized	2	365367	new with AtomicReference	2	368952
original	4	555986	new with synchronized	4	540642	new with AtomicReference	4	545715
original	8	854029	new with synchronized	8	852137	new with AtomicReference	8	850401

Issue Links

• is related to: HBASE-17072 CPU usage starts to climb up to 90-100% when using G1GC; purge ThreadLocal usage