HBase 0.20.0 Performance Evaluation

HBase‐0.20.0 Performance Evaluation
Anty Rao and Schubert Zhang, August 21, 2009
{ant.rao, schubert.zhang}@gmail.com，http://cloudepr.blogspot.com/

Background

We have been using HBase for around a year in our development and projects, from 0.17.x to
0.19.x. We and all in the community know the critical performance and reliability issues of these
releases.

Now, the great news is that HBase‐0.20.0 will be released soon. Jonathan Gray from Streamy,
Ryan Rawson from StumbleUpon, Michael Stack from Powerset/Microsoft, Jean‐Daniel Cryans
from OpenPlaces, and other contributors had done a great job to redesign and rewrite many
codes to promote HBase. The two presentations [1] [2] provide more details of this release.

The primary themes of HBase‐0.20.0:
− Performance
 Real‐time and Unjavafy software implementations.
 HFile, based on BigTable’s SSTable. New file format limits index size.
 New API
 New Scanners
 New Block Cache
 Compression (LZO, GZ)
 Almost a RegionServer rewrite
− ZooKeeper integration, multiple masters (partly, 0.21 will rewrite Master with better ZK
integration)
Then, we will get a bran‐new, high performance (Random Access, Scan, Insert, …), and stronger
HBase. HBase‐0.20.0 shall be a great milestone, and we should say thanks to all developers.

Following items are very important for us:
− Insert performance: We have very big datasets, which are generated fast.
− Scan performance: For data analysis in MapReduce framework.
− Random Access performance: Provide real‐time services.
− Less memory and I/O overheads: We are neither Google nor Yahoo!, we cannot operate
big cluster with hundreds or thousands of machines, but we really have big data.
− The HFile: Same as SSTable. It should be a common and effective storage element.

Testbed Setup

Cluster:
− 4 slaves + 1 master
− Machine: 4 CPU cores (2.0G), 2x500GB 7200RPM SATA disks, 8GB RAM.
− Linux: RedHat 5.1 (2.6.18‐53.el5), ext3, no RAID

1

− 1Gbps network, all nodes under the same switch.
− Hadoop‐0.20.0 (1GB heap), HBase‐0.20.0 RC2 (4GB heap), Zookeeper‐3.2.0

By now, HBase‐0.20.0 RC2 is available for evaluation:
http://people.apache.org/~stack/hbase‐0.20.0‐candidate‐2/. Refer to [2] and [4] for installation.

Hadoop‐0.20.0 configuration important parameters:
core‐site.xml:
parameter value notes
io.file.buffer.size 65536 Irrelevant to this evaluation. We like use this size to improve
file I/O. [5]
io.seqfile.compress.blocksize 327680 Irrelevant to this evaluation. We like use this size to compress
data block.

hdfs‐site.xml
dfs.namenode.handler.count 20 [5]
dfs.datanode.handler.count 20 [5] [6]
dfs.datanode.max.xcievers 3072 Under heavy read load, you may see lots of DFSClient
complains about no live nodes hold a particular block. [6]
dfs.replication 2 Our cluster is very small. 2 replicas are enough.

mapred‐site.xml
io.sort.factor 25 [5]
io.sort.mb 250 10 * io.sort.factor [5]
mapred.tasktracker.map.tasks.maximum 3 Since our cluster is very small, we
set 3 to avoid client side bottleneck.
mapred.child.java.opts ‐Xmx512m JVM GC option
‐XX:+UseConcMarkSweepGC
mapred.reduce.parallel.copies 20
mapred.job.tracker.handler.count 10 [5]

hadoop‐env.sh
HADOOP_CLASSPATH ${HADOOP_HOME}/../hbase‐0.20.0/hbase‐0.20.0.jar: Use HBase jar and
${HADOOP_HOME}/../hbase‐0.20.0/conf:${HADOOP_ configurations. Use
HOME}/../hbase‐0.20.0/lib/zookeeper‐r785019‐hbas Zookeeper jar.
e‐1329.jar
HADOOP_OPTS "‐server ‐XX:+UseConcMarkSweepGC" JVM GC option

HBase‐0.20.0 configuration important parameters:
hbase‐site.xml

2

hbase.cluster.distributed true Fully‐distributed with unmanaged ZooKeeper Quorum
hbase.regionserver.handler.count 20

hbase‐env.sh
HBASE_CLASSPATH ${HBASE_HOME}/../hadoop‐0.20.0/conf Use hadoop configurations.
HBASE_HEAPSIZE 4000 Give HBase enough heap size
HBASE_OPTS "‐server ‐XX:+UseConcMarkSweepGC" JVM GC option
HBASE_MANAGES_ZK false Refers to hbase.cluster.distributed

Performance Evaluation Programs

We modified the class org.apache.hadoop.hbase.PerformanceEvaluation, since the code has
following problems: The patch is available at http://issues.apache.org/jira/browse/HBASE‐1778.
− It is not updated according to hadoop‐0.20.0.
− The approach to split maps is not strict. Need to provide correct InputSplit and
InputFormat classes. Current code uses TextInputFormat and FileSplit, it is not
reasonable.

The evaluation programs use MapReduce to do parallel operations against an HBase table.
− Total rows: 4,194,280.
− Row size: 1000 bytes for value, and 10 bytes for rowkey. Then we have ~4GB of data.
− Sequential row ranges: 40. (It’s also used to define the total number of MapTasks in
each evaluation.)
− Rows of each sequential range: 104,857
The principle is same as the evaluation programs described in Section 7, Performance Evaluation,
of the Google BigTable paper [3], pages 8‐10. Since we have only 4 nodes to work as clients, we
set mapred.tasktracker.map.tasks.maximum=3 to avoid client side bottleneck.

Performance Evaluation

Report‐1: Normal, autoFlush=false, writeToWAL=true
Google paper
Eventual Total
rows/s ms/row (rows/s in
Experiment Elapsed row/s throughput
per node per node single node
Time(s) (MB/s)
cluster)
random reads 948 4424 1106 0.90 4.47 1212
random writes(i) 390 10755 2689 0.37 10.86 8850
random writes 370 11366 2834 0.35 11.45 8850
sequential reads 193 21732 5433 0.18 21.95 4425
sequential writes(i) 359 11683 2921 0.34 11.80 8547
sequential writes 204 20560 5140 0.19 20.77 8547
scans 68 61681 15420 0.06 62.30 15385

3

Report‐2: Normal, autoFlush=true, writeToWAL=true
Google paper
Eventual Total
Time(s) (MB/s)
cluster)
random writes 600 6990 1748 0.57 7.06 8850

Random writes (i) and sequential write (i) are operations against a new table, i.e. initial writes.
Random writes and sequential writes are operations against an existing table that is already
distributed on all region servers, i.e. distributed writes. Since there is only one region server is
accessed at the beginning of inserting, the performance is not so good until the regions are split
across all region servers. The difference between initial writes and distributed writes for random
writes is not so obvious, but that is distinct for sequential writes.

In Google’s BigTable paper [3], the performance of random writes and sequential writes is very
close. But in our result, sequential writes are faster than random writes, it seems the client side
write‐buffer (12MB, autoFlush=false) taking effect, since it can reduce the number of RPC
packages. And it may imply that further RPC optimizations can gain better performance. If we set
autoFlush true, they will be close too, and less than the number of random writes in report‐1. The
report‐2 shows our inference is correct.

Random reads are far slower than all other operations. Each random read involves the transfer of
a 64KB HFlie block from HDFS to a region server, out of which only a single 1000‐byte value is
used. So the real throughput is approximately 1106*64KB=70MB/s of data read from HDFS, it is
not low. [3] [8]

The average time to random read a row is sub‐ms here (0.90ms) on average per node, that seems
about as good as we can get on our hardware. We're already showing 10X better performance
than a disk seeking (10ms). It should be major contributed by the new BlockCache and new HFile
implementations. Any other improvements will have to come from HDFS optimizations, RPC
optimizations, and of course we can always get better performance by loading up with more RAM
for the file‐system cache. Try 16GB or more RAM, we might get greater performance. But
remember, we're serving out of memory and not disk seeking. Adding more memory (and region
server heap) should help the numbers across the board. The BigTable paper [3] shows 1212
random reads per second on a single node. That's sub‐ms for random access, it’s clearly not
actually doing disk seeks for most gets. (Thanks for good comments from Jonathan Gray and
Michael Stack.) [9]

The performance of sequential reads and scans is very good here, even better than the Google
paper’s [3]. We believe this performance will greatly support HBase for data analysis
(MapReduce). In Google’s BigTable paper [3], the performance of writes will be better than reads,
includes sequential reads. But in our test result, the sequential reads are better than writes.
Maybe there are rooms to improve the performance of writes in the future.

4

And remember, the dataset size in our tests is not big enough (only average 1GB per node), so the
hit ratio of BlockCache is very high (>40%). If a region server serves large dataset (e.g. 1TB), the
power of BlockCache would be downgraded.

Bloom filters can reduce the unnecessary search in HFiles, which can speed up reads, especially
for large datasets when there are multiple files in an HBase region.

We can also consider RAID0 on multiple disks, and mount local file system with noatime and
nodiratime options, for performance improvements.

Performance Evaluation (none WAL)

In some use cases, such as bulk loading a large dataset into an HBase table, the overhead of the
Write‐Ahead‐Logs (commit‐logs) are considerable, since the bulk inserting causes the logs get
rotated often and produce many disk I/O. Here we consider to disable WAL in such use cases, but
the risk is data loss when region server crash. Here I cite a post of Jean‐Daniel Cryans on HBase
mailing list [7].
“As you may know, HDFS still does not support appends. That means that the write ahead logs or WAL
that HBase uses are only helpful if synced on disk. That means that you lose some data during a region
server crash or a kill ‐9. In 0.19 the logs could be opened forever if they had under 100000 edits. Now in 0.20
we fixed that by capping the WAL to ~62MB and we also rotate the logs after 1 hour. This is all good because
it means far less data loss until we are able to append to files in HDFS.
Now to why this may slow down your import, the job I was talking about had huge rows so the logs
got rotated much more often whereas in 0.19 only the number of rows triggered a log rotation. Not writing
to the WAL has the advantage of using far less disk IO but, as you can guess, it means huge data loss in the
case of a region server crash. But, in many cases, a RS crash still means that you must restart your job
because log splitting can take more than 10 minutes so many tasks times out (I am currently working on that
for 0.21 to make it really faster btw).”

So, here we call put.setWriteToWAL(false) to disable WAL, and expect get better writing
performance. This table is the evaluation result.

Report‐3: NonWAL (autoFlush=false, writeToWAL=false)
Google paper
Eventual Total
Time(s) (MB/s)
cluster)
random reads 1001 4190 1048 0.95 4.23 1212
random writes(i) 260 16132 4033 0.25 16.29 8850
random writes 194 21620 5405 0.19 21.84 8850
sequential reads 187 22429 5607 0.18 22.65 4425
scans 62 67650 16912 0.06 68.33 15385

5

We can see the performance of sequential writes and random writes are far better (~double)
than the normal case (with WAL). So, we can consider using this method to solve some bulk
loading problems which need high performance for inserting. But we suggest calling admin.flush()
to flush the data in memstores to HDFS, immediately after each bulk loading job, to persist data
and avoid loss as much as possible. The above evaluation does not include the time of flush.

Conclusions

Compares to the metrics in Google’s BigTable paper [3], the write performance is still not so good,
but this result is much better than any previous HBase release, especially for the random reads.
We even got better result than the paper [3] on sequential reads and scans. This result gives us
more confidence.

HBase‐0.20.0 should be good to be used:
− as a data store to be inserted/loaded large datasets fast
− to store large datasets to be analyzed by MapReduce jobs
− to provide real‐time query services

We have made a comparison of the performance to run MapReduce jobs which sequentially
retrieve or scan data from HBase tables, and from HDFS files. The latter is trebly faster. And the
gap is even bigger for sequential writes.

HBase‐0.20.0 is not the final word on performance and features. We are looking forward to and
researching following features, and need to read code detail.
− Other RPC‐related improvements
− Other Java‐related improvements
− New master implementation
− Bloom Filter
− Bulk‐load [10]

In the world of big data, best performance usually comes from tow primary schemes:
(1) Reducing disk I/O and disk seek.
If we come back to review the BigTable paper [3] Section 6, Refinements, we can find the
goals of almost all those refinements are related to reducing disk I/O and disk seek, such as
Locality Group, Compression, Caching (Scan Cache and Block Cache), Bloom filters, Group
commit, etc.
(2) Sequential data access.
When implementing our applications and organize/store our big data, we should always try
our best to write and read data in sequential mode. Maybe sometimes we cannot find a
generalized access scheme like that in the traditional database world, to access data in
various views, but this may be the trait of big data world.

6

References:
[1] Ryan Rawson’s Presentation on NOSQL.
http://blog.oskarsson.nu/2009/06/nosql‐debrief.html
[2] HBase goes Realtime, The HBase presentation at HadoopSummit2009 by Jonathan Gray
and Jean‐Daniel Cryans
http://wiki.apache.org/hadoop‐data/attachments/HBase(2f)HBasePresentations/attach
ments/HBase_Goes_Realtime.pdf
[3] Google paper, Bigtable: A Distributed Storage System for Structured Data
http://labs.google.com/papers/bigtable.html
[4] HBase‐0.20.0‐RC2 Documentation,
http://people.apache.org/~stack/hbase‐0.20.0‐candidate‐2/docs/
[5] Cloudera, Hadoop Configuration Parameters.
http://www.cloudera.com/blog/category/mapreduce/page/2/
[6] HBase Troubleshooting, http://wiki.apache.org/hadoop/Hbase/Troubleshooting
[7] Jean‐Daniel Cryans’s post on HBase mailing list, on Aug 12, 2009: Tip when migrating
your data loading MR jobs from 0.19 to 0.20.
[8] ACM Queue, Adam Jacobs, 1010data Inc., The Pathologies of Big Data,
http://queue.acm.org/detail.cfm?id=1563874
[9] Our another evaluation report:
http://docloud.blogspot.com/2009/08/hbase‐0200‐performance‐evaluation.html
[10] Yahoo! Research, Efficient Bulk Insertion into a Distributed Ordered Table，
http://research.yahoo.com/files/bulkload.pdf

7

HBase 0.20.0 Performance Evaluation

More Related Content

What's hot (20)

Similar to HBase 0.20.0 Performance Evaluation (20)

More from Schubert Zhang (20)

Recently uploaded (20)

HBase 0.20.0 Performance Evaluation