I recently set up a provisioned IOPS EBS volume on Amazon EC2. I decided to run CrystalDiskMark against it to see how it compared to a standard EBS volume in terms of performance. I wanted to know if provisioned IOPS is worth the extra money that you pay for it. Here are the results:
EBS Standard
Volume Size: 100GB
Server: c3.large
OS: Windows Server 2012
———————————————————————–
CrystalDiskMark 3.0.3 x64 (C) 2007-2013 hiyohiyo
Crystal Dew World : http://crystalmark.info/
———————————————————————–
* MB/s = 1,000,000 byte/s [SATA/300 = 300,000,000 byte/s]
Sequential Read : 59.218 MB/s
Sequential Write : 34.096 MB/s
Random Read 512KB : 59.173 MB/s
Random Write 512KB : 33.581 MB/s
Random Read 4KB (QD=1) : 7.792 MB/s [ 1902.2 IOPS]
Random Write 4KB (QD=1) : 3.165 MB/s [ 772.7 IOPS]
Random Read 4KB (QD=32) : 60.193 MB/s [ 14695.5 IOPS]
Random Write 4KB (QD=32) : 10.020 MB/s [ 2446.2 IOPS]
Test : 1000 MB [C: 46.3% (46.2/99.7 GB)] (x5)
Date : 2014/03/31 5:59:00
OS : Windows Server 2012 Server Standard (full installation) [6.2 Build 9200] (x64)
EBS Provisioned IOPS
Volume Size: 100GB
Provisioned IOPS: 3000
Server: c3.4xlarge (EBS-Optimized)
OS: Windows Server 2012
———————————————————————–
CrystalDiskMark 3.0.3 x64 (C) 2007-2013 hiyohiyo
Crystal Dew World : http://crystalmark.info/
———————————————————————–
* MB/s = 1,000,000 byte/s [SATA/300 = 300,000,000 byte/s]
Sequential Read : 50.116 MB/s
Sequential Write : 50.102 MB/s
Random Read 512KB : 50.121 MB/s
Random Write 512KB : 50.121 MB/s
Random Read 4KB (QD=1) : 12.634 MB/s [ 3084.6 IOPS]
Random Write 4KB (QD=1) : 8.028 MB/s [ 1959.9 IOPS]
Random Read 4KB (QD=32) : 12.534 MB/s [ 3060.1 IOPS]
Random Write 4KB (QD=32) : 12.534 MB/s [ 3060.2 IOPS]
Test : 1000 MB [C: 55.5% (55.3/99.7 GB)] (x5)
Date : 2014/03/31 5:57:08
OS : Windows Server 2012 Server Standard (full installation) [6.2 Build 9200] (x64)
Conclusion
Although EBS Provisioned IOPS does appear to almost always guarantee consistency in throughput, it also prevents you from attaining the maximum available speeds available through bursting.
Write performance on EBS is considerably slower than read performance. If you perform a lot of writes, Provisioned IOPS can really help.
If you are going to purchase a provisioned IOPS volume, provision 3000+ IOPS, or you’re likely to get better performance on average with EBS standard.
Unfortunately EBS Provisioned IOPS is not a magic bullet for increased IO performance, only consistency. No matter which type of EBS volume you choose, it’s going to be slow.
Leave a Reply