question Re: HDFS Compression vs Performance in Archives of Support Questions (Read Only)

HDFS Compression vs Performance

michelsumbul — Sun, 24 Jan 2016 22:31:12 GMT

Hi,

I'm asking myself how to have a good idea of the impact on the performance to use compression in HDFS?

This question is important because, if I implement the compression, should I considere increasing the CPU need of 10%, 20 %, 30% for the same performance?

I know that I can win on performance because less IOPS will be needed but what's about CPU?

I would like also ask, what will be the impact on HBase performance?

Many thanks in advance!

Re: HDFS Compression vs Performance

nsabharwal — Sun, 24 Jan 2016 22:50:09 GMT

@Michel Sumbul Very good question ..Please see this to start with http://www.slideshare.net/Hadoop_Summit/kamat-singh-june27425pmroom210cv2

It was 3 years ago. http://hortonworks.com/blog/orcfile-in-hdp-2-better-compression-better-performance/

Re: HDFS Compression vs Performance

aervits — Mon, 25 Jan 2016 01:16:42 GMT

@Michel Sumbul in terms of HBase, your mileage can vary, it depends on your workloads, in some use cases I've seen a lot better performance with compression on and in some not. There are also multiple levels of compression in HBase, (per column family, you can compress rowkeys only or both).

Re: HDFS Compression vs Performance

michelsumbul — Mon, 25 Jan 2016 03:39:03 GMT

Thanks guys for the fast reply!

@nsabharwal : If I correctly understand the slides, I should expect a raise of the CPU usage between 5% and 60% depending of the compression algorythm. That can be really important!

@aervits : do you have some benchmarks, test results to have a idea?

Many thanks guys!

Michel

Re: HDFS Compression vs Performance

nsabharwal — Mon, 25 Jan 2016 07:07:52 GMT

@Michel Sumbul Based on the compression type...Yes

If I correctly understand the slides, I should expect a raise of the CPU usage between 5% and 60% depending of the compression algorythm. That can be really important!

HBASE - Link (Unofficial)

HBASE official guide - Production systems should use compression with their ColumnFamily definitions. See Appendix C, Compression In HBase for more information.

Re: HDFS Compression vs Performance

aervits — Mon, 25 Jan 2016 21:41:30 GMT

@Michel Sumbul I was able to reach 840k/sec reads on AWS, Centos7, XFS filesystem, 9 nodes, 12 7200 RPM drives in non-mapreduce mode. Same hardware write-only test resulted in 185k/sec. For mixed workload, I got 148k/sec writes and 270k/s reads. This is with snappy compression on.