Comments on: EC2/EBS single and RAID volumes IO benchmark

By: Ozgur Akan

Ozgur Akan — Tue, 20 Dec 2011 16:50:32 +0000

Hi,

@Vadim, I believe things must have improved on AWS/EBS side so like Laszio posted, maybe there is not much performance difference any more with different RAID setups.Do you plan to repeat the test?

Also it would be interesting to compare these with RDS.

best wishes,
Oz

By: László Bácsi

László Bácsi — Fri, 07 Oct 2011 10:41:07 +0000

@Vadim, we were considering a RAID setup on EC2 and I found this post. I wanted to confirm your findings before doing something like this in production. I created a new large instance and a RAID0 array of 4 EBS volumes (4GB each). I used the same configuration as you did. I also attached a single EBS volume for reference. I used XFS on both and I run the same benchmark as you did.

I was surprised to find *no performance improvement* whatsoever over the single EBS volume. All the results were almost the same except for rndwr with 8 threads for the 256Mb file (which showed a 1.6x improvement with raid0).

Do you have any idea why this might be? I’ve uploaded my results to CloudApp: http://cl.ly/0W3q3e3V432X2e3z3n1Y/sysbench_ebs_raid0_4.zip and http://cl.ly/3G2v280d2a2r3N3r3w1N/sysbench_ebs_single_drive.zip

By: Aaron Brown

Aaron Brown — Sun, 24 Jul 2011 18:41:45 +0000

RAIDed EBS volumes can be snapshotted just as with non-RAIDed volumes. You simply have to either unmount the md volume or use XFS and freeze the volume briefly for the duration of the snapshot. ec2-consistent-snapshot from Alestic supports the XFS freeze/unfreeze methodology with a FLUSH TABLES WITH READ LOCK or MySQL shutdown.

By: Wes Shull

Wes Shull — Thu, 11 Nov 2010 04:31:19 +0000

@Dave Rose:

EBS volumes can only be accessed by instances in the same availability zone, so the answer to your question is “yes”, by necessity.

By: Dave Rose

Dave Rose — Tue, 04 May 2010 21:35:18 +0000

Were all your EBS volumes in the same availability zone?

By: Marlon

Marlon — Sat, 01 May 2010 07:13:35 +0000

http://aws.amazon.com/ec2/instance-types/ says:

“While some resources like CPU, memory and instance storage are dedicated to a particular instance, other resources like the network and the disk subsystem are shared among instances.”

While I have no experience with EC2, I assume this to cause quite a bit of unrealiable I/O behavior.

By: Marc Slemko

Marc Slemko — Thu, 13 Aug 2009 23:49:41 +0000

The issue we ran into trying to get performance out of EBS volumes with software RAID is that they are not all consistent over time. We were ending up with a slow volume, and which volume was slow might change over time. If you care about reads, then in theory a smart RAID1 setup with the right kernel could fix that by balancing reads based on performance, but if you just have a dumb round robin read setup then you are hamstrung by the slowest disk. If you care about writes, it is pretty tough no matter what you do.

By: Thorsten von Eicken

Thorsten von Eicken — Thu, 13 Aug 2009 07:52:59 +0000

Very nice and interesting results. Something that looks fishy to me is that with 1 volume you seem to be getting the best performance with a single thread, but with two striped volumes you get the best performance with 16 threads. If one volume “can’t handle” more than one thread effectively, you should see the best performance on 2 volumes with 2 threads or perhaps 4, but not with 16. So something is going here…

Amazon is pretty clear about the fact that each EBS volume has redundancy built in and that adding additional redundancy is not particularly recommended (at that point other failure modes dominate anyway). Instead, it is recommended to spend the extra effort on doing frequent snapshots (which have their own set of performance effects).

How many distinct instances did you fire up and run the tests, btw?

By: Mark Callaghan

Mark Callaghan — Fri, 07 Aug 2009 13:29:07 +0000

@Morgan – the first write to a disk block on EC2 has a performance penalty. I have not read the explanation, but Amazon mentions it at http://docs.amazonwebservices.com/AWSEC2/latest/DeveloperGuide/index.html?instance-storage.html

By: Vadim

Vadim — Fri, 07 Aug 2009 04:22:41 +0000

Jean-Emmanuel OrfÃ¨vre ,

Sorry, typo. fixed