Details
-
New Feature
-
Status: Closed
-
Major
-
Resolution: Fixed
-
None
-
None
-
Reviewed
-
Support reusing MemStoreChunk when MSLAB is enabled, thereby decrease the cost of YGC. It is off by default, turn it on by configuring "hbase.hregion.memstore.chunkpool.maxsize" (between 0.0 and 1.0)
Description
Usage:
Disable pool(default): configure "hbase.hregion.memstore.chunkpool.maxsize" as 0
Enable pool: configure "hbase.hregion.memstore.chunkpool.maxsize" as a percentage of global memstore size (between 0.0 and 1.0, recommend to set it as the gap between min and max sizes of memstore, e.g. 0.5)
Background：
When we use mslab,we will copy the keyvalue together in a structure called MemStoreLAB$Chunk, therefore we could decrease the heap fragment.
Problem：
When one chunk is full, we will create a new chunk, and then the old chunk will be reclaimed by JVM if no reference to it.
Mostly the chunk object would be promoted when doing Young GC, cause increasing the cost of YGC
When a Chunk object has no reference? It should meet two following condition:
1.Memstore which this chunk belongs to is flushed
2.No scanner is opening on the memstore which this chunk belongs to
Solution:
1.Create a chunk pool to manage the no-reference chunks, instead of being reclaimed by JVM
2.When a Chunk has no reference, put it back to the pool
3.The pool has a max capacity, it will skip the chunks when achieve the max size
4.When we need new Chunk to store KeyValue, get it from the pool if exists, else create new one by pool, so we could be able to reuse the old chunks
Test results:
Environment:
hbase-version:0.94
-Xms4G -Xmx4G -Xmn2G
Row size=50 bytes, Value size=1024 bytes
50 concurrent theads per client, insert 10,000,000 rows
Before:
Avg write request per second:12953
After testing, final result of jstat -gcutil :
YGC YGCT FGC FGCT GCT
747 36.503 48 2.492 38.995
After:
Avg write request per second:14025
After testing, final result of jstat -gcutil :
YGC YGCT FGC FGCT GCT
711 20.344 4 0.284 20.628
Improvement: YGC 40+%; WPS 5%+
review board :
https://reviews.apache.org/r/10056/