.

.

Average latency on HBase 0.95 by Hadoop version by caching solution on 100% cached and local data

.

Hadoop VersionBlock cacheOS cacheOS cache + SCR*OS cache + SCR + no CRCPure OS*** cache+SCR+no CRCHBase CRC over OS cache+SCRCliff Click'd block cache

.

1.0.4 / 1.1.2**91418971489853XXX

.

CDH4.2.089414011247855987935974

.

.

.

Data: 100% local. Hit ratio on the caches: 100%

.

Setup is: 10 nodes, where 1 runs all the masters and the tests

.

Using EC2 m1.xlarge instances

.

Enough memory is provided so that there's no churn (about 5GB of data per node)

.

8 threads are random reading following a zipfian distribution via YCSB

.

Latency is represented in microseconds

.

Compression isn't used, in fact the configs are the bare minimum to make the tests run

.

All the tests are run on pre-filled caches, there's no "load time" penalty

.

* Note that SCR in 2.0 is HDFS-347, and in Hadoop 1.0 is HDFS-2246

.

** The cluster runs on 1.0, the lib HBase is using is 1.1

.

*** Pure OS means that the block cache is disabled and doesn't hold meta blocks