Advanced Computing in the Age of AI | Friday, March 29, 2024

Fat and Fast C3 Server Slices Unveiled for AWS Cloud 

Rackspace Hosting just announced its own high-performance, flash-based server instances on its public cloud, and Amazon Web Services has answered with EC2 compute cloud instances that have the latest Intel Xeon processors, big gobs of memory, and flash storage to support the most demanding applications.

The new C3 instance types on the EC2 compute cloud are based on two-socket servers using Intel's Xeon E5-2680 v2 processors. These chips, which debuted with the rest of the E5-2600 v2 family back in September, have ten cores that run at 2.8 GHz and can Turbo Boost up to 3.6 GHz on a single core if there is thermal headroom. And AWS is letting Turbo Boost run as well as giving customers access to the AVX vector math units on the chips, too. As it has been doing for certain kinds of instances, AWS is also talking about what particular processor is underneath the C3 instances so companies can tune up their applications.

Werner Vogels, CTO at Amazon, introduced the new C3 instances during his keynote at the re:Invent conference that AWS is hosting in Las Vegas this week, saying that it was the "absolute highest performance instance" that AWS has. Vogels added that the inter-instance latencies have been "brought down much lower" and it allows customers to run the highest performance HPC applications.

Here is how the new instances stack up:

aws-ec2-c3-instances

If you are not familiar with how Amazon carves up processing capacity on its cloud, an EC2 Compute Unit (ECU) 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, which was when Amazon first launched its public cloud. That two-socket Xeon E5-2680 v2 server has 20 cores and 40 threads, so it looks like a lot of the processing capacity – but not all of it – is being burned up in the c3.8xlarge instance type. HyperThreading is available on these processors and is turned on, and each virtual CPU is running on one of those virtual threads.

Amazon did not specify what it has done to lower the latency between the nodes in the C3 instances, but presumably it is using a more recent implementation of 10 Gb/sec Ethernet switching between the nodes; it is possible that it has moved to 40 Gb/sec Ethernet, but unlikely that the company has moved to InfiniBand for networking. All the company said is that the networking now has "higher performance (in terms of packets per second), much lower latency, and lower jitter."

Just for fun, and to show how much oomph these new C3 instances have, AWS fired up a cluster of C3 instances with 26,496 cores and ran the Linpack parallel Fortran benchmark test on it; it delivered 484.2 teraflops of sustained Linpack performance. A virtual cluster with 8,192 cores was also tested and delivered 163.9 teraflops sustained on the Linpack test. The machines are not limited to Linpack, of course. Any kind of workload that demands high CPU performance, a reasonable amount of memory, and high I/O from local storage is going to benefit. In fact, it is hard to imagine a workload that would not benefit from such a configuration.

The C3 instances are available now in the US East (Northern Virginia), US West (Oregon), EU (Ireland), Asia/Pacific (Singapore), Asia/Pacific (Tokyo), and Asia/Pacific (Sydney) regions of the AWS cloud, and reserved, on-demand, and spot pricing are all available on them.

The Performance Cloud Server instances that Rackspace announced last week have larger amounts of memory and flash storage for a given number of virtual CPUs than the C3 instances, but they are based on the earlier "Sandy Bridge-EP" Xeon E5-2600 v1 processors. Rackspace has not said which Xeon E5 it is using, so it is not possible to gauge relative performance.  Rackspace is charging less than AWS for skinny slices of its high-performance cloudy servers, but considerably more for larger slices of the same machines. What would be really interesting is to see a bakeoff between these two instance types on a variety of benchmarks.

EnterpriseAI