Advanced Computing in the Age of AI | Friday, September 30, 2022

Amazon Cranks Up I/O On New EC2 Cloud Slices 

Cloud computing juggernaut Amazon Web Services has forged a new class of server instances for its EC2 compute cloud that feature the kind of high I/O that it says are better suited to transactional applications and NoSQL databases like Cassandra and MongoDB. AWS is also slashing prices on its S3 and EBS storage services, which is sure to cause much consternation among its many competitors.

The instance type, called I2, is the next generation of high I/O instances on the EC2 compute cloud. As it has been doing since last fall, Amazon is being specific about the underlying processors on the servers that support these I2 instances, and in this case AWS is using two-socket servers based on Intel's latest "Ivy Bridge-EP" Xeon E5-2600 v2 processors. Specifically, the nodes are based on the ten-core E5-2670 v2, which runs at 2.5 GHz and which is the middle chip of the five ten-core variants of the Xeon E5s. Each virtual CPU on the I2 instances is a HyperThread on the processor, which has a total of 20 cores and 40 threads across the two-socket machine. Here are the feeds and speeds for the I2 instances:



In the table above, ECU is short for EC2 Compute Unit, and as the Amazon documentation explains, this is roughly equivalent to the processing capacity of a 1 GHz to 1.2 GHz Opteron or Xeon core from around 2007 or a 1.7 GHz Xeon chip from 2006 when AWS first launched. This being a high I/O instance, the networking and the I/O operations per second coming off the solid state disks are more important than compute capacity. The largest i2.8xlarge instance is capable of doing over 365,000 random reads and 315,000 random writes use 4 KB files, according to AWS; the smaller slices do proportionally less based on their number of SSDs and virtual CPUs. The flash drives also have the trim command activated, which allows the server file systems and operating system to keep track of what pages in the flash drives are empty and more efficiently write data to the flash. Trim is supported in Free BSD Unix, Linux, and Windows Server 2008 R2 and 2012.

The new I2 instances have faster processors than the prior generation of HI1 instances, also shown in the table, and there are four different I2 options rather than one. They also have twice the virtual memory per virtual CPU and SSD capacity that scales up with CPU and memory. The I2 instances have fatter SSDs as well as a larger number of drives compared to general purpose M3 instances. The prices scale up proportionately, too.

The I2 instances also have what AWS calls Enhanced Networking, which became available on the C3 instances that were launched in November 2013. The exact specs of this networking have not been revealed – AWS says merely that it supports Single Root I/O Virtualization (SR-IOV) as well as offering higher network performance in terms of packets per second handled, lower latency, and lower jitter compared to standard EC2 instances. Our guess is that AWS has shifted to 10 Gb/sec Ethernet adapters on the servers and switches linking them from Gigabit Ethernet on old servers and switches in its fleet of machines.

The I2 instances are available through on demand and reserved pricing, but you cannot get them on the spot market. Servers in the US East (N. Virginia), US West (Oregon), US West (California), EU (Ireland), Asia Pacific (Singapore), Asia/Pacific (Tokyo), and Asia/Pacific (Sydney) regions have been configured to supply the I2 instances.

In a separate announcement, AWS has cut prices on its S3 object storage by between 6 and 22 percent across all of its regions. Pricing on the standard edition of its Elastic Block Storage has been cut by 50 percent. EBS price cuts are also across all regions and include both the capacity fees (GB/month) as well as I/O fees (which are sold in 1 million I/O request blocks).

Add a Comment