Instance types and network throughput

Network throughout is a very good metrics to know when running an Oracle Database on AWS or for any time you are moving data from different AWS services other AWS services. Maximum throughput for different Amazon EC2 instance types can be found here:
http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EBSOptimized.html

This 2013 AWS reInvent session also has details http://www.slideshare.net/AmazonWebServices/stg302-28617072 (Slide 19)

The m1.xlarge instance maximum network through put is 128 MBPS / 1K Mbps. The CC2.8xlarge, CR1.8xlarge, hi1.4xlarge, and cg1.xlarge (http://www.slideshare.net/AmazonWebServices/stg302-28617072) can provide up to 800 MPBS.

EC2 EBS optimized instances throughput

EC2 EBS-Optimized Instances are listed here:

http://docs.aws.amazon.com/AWSEC2/latest/UserGuide/EBSOptimized.html

It is important to understand the maximum throughput in MB/s and maximum 16 KB IOPS when architecting Oracle databases on AWS.  

Network throughput for an Oracle database : Smaller number of large PIOPS volumes or large number of smaller PIOPS volumes

Individual 4000 PIOP volumes at 16K  I/O's can theoretically transfer 64MB/s.  Smaller PIOP volumes (e.g 1000) can burst up to 40MB/s.   In this case, especially for bursty workloads, it may be more cost effective to use smaller PIOPS volumes (but more of them) in a raid0 configuration, than to use fewer higher rated PIOPS volumes.

Oracle Database size and network throughput : IOPS and network throughput

When configuring an Oracle Database on AWS EC2, you need to consider both storage (EBS) IOPS and network throughput.   With PIOPS, you can achieve up to 4K IOPS per EBS volumes.  However, the EC2 instances (assuming EBS optimized or 10 Gbps cluster compute) can be a potential bottle neck.  For example, the m2.2xlarge instance type has a maximum throughput of 0.5Gb.  Which means this instance type is limited to approximately 3750 * 16 KiB IOPS. Therefore, one 4K PIOPS volume would start to saturate the network.   Take into account the IOPS and instance network throughput when designing your Oracle Database on AWS EC2.

AWS EBS PIOPS : block size and IOPS

Having spent more time in the database world than in the web development world, I am accustomed to measuring (database) performance/through put in terms of IOPS or TPS.  The web/video/image world like to use MB/sec.  Why I am saying this? Because it relates to the conversation about getting a certain level of PIOPS (based upon a 16 KB block) on AWS EBS and how this effects MB/sec.  MB/sec, I am beginning to understand, and maybe move to the 'dark side', is the ultimate measure of disk 'performance'.  

Example: A 2000 Provisioned IOPS volume can handle:

•2000 16KB read/write per second, or 1000 32KB read/write per second, or 500 64KB read/write per second 

•You will get consistent 32 MB/sec throughput (with 16KB or higher IOs)

•Perform an index creation action and sends I/O of 32K, IOPS becomes 1000, you still get 32MB/sec throughput

•On best effort, you may get up to 40 MB/sec throughput 

So, you may be better off using a 64 KB block size but your PIOPS will show up as lower but your MB/sec could be better.

AWS S3 load throughput

Here is the S3 performance tips and tricks public blog entry with some good information:

http://aws.typepad.com/aws/2012/03/amazon-s3-performance-tips-tricks-seattle-hiring-event.html

The key learning is that S3 does continue to scale very well and there are customers running >100k TPS. However, S3 does not automatically create a partition scheme quickly nor can they use an API to manage this or use CloudWatch to monitor. 

·