Amazon gooses HPC cloud with Xeon E5s
Passes the Linpack speed test
SC11 Amazon Web Services, the cloud computing unit of the online retailing giant, has been peddling special HPC instances of its EC2 cloud for more than a year now.
This week, in conjunction with the SC11 supercomputing conference in Seattle, the company announced that it has beefed up its HPC cloud with the new Xeon E5 processors from Intel, and has put the new iron through the Linpack paces to get it ranked on the latest Top 500 supercomputers list.
Amazon first came out with its dedicated HPC instances back in July 2010 after recognizing that HPC users needed faster 10 Gigabit Ethernet networking between physically adjacent nodes, and also to have the precise specifications of the nodes available to them so they could properly tune their workloads and get balanced performance across identical nodes. In this case, Amazon networked a bunch of two-socket x86 servers using Intel's quad-core Xeon X5570 chips, which spin at 2.93GHz.
When virtualized, this physical server node, which was called a Cluster Compute Instance, presented 33.5 EC2 compute units and 23GB of virtual memory to HPC applications. Initially, AMD restricted the cluster size to eight nodes, or 64 cores and about 750 gigaflops of peak theoretical performance. (That's a little more than you get in an Nvidia Tesla GPU.) Amazon told El Reg back then that it had tested an 880 server subcluster internally and delivered 41.82 teraflops of oomph sustained on the Linpack test.
A year ago, at the SC10 conference, Amazon added two Nvidia M2050 fanless GPU coprocessors to each physical server, each capable of pushing 515 gigaflops of peak double-precision floating math. The idea was to give customers who might want to build ceepie-geepie hybrid supercomputers a place to play around with the technology first before making substantial investments in hardware. Or, in some cases, to allow researchers who never will have any interest in owning iron a chance to play with CPU-GPU clusters.
This week, Amazon is announcing Cluster Compute 2 (CC2) instances running atop the EC2 cloud, which sport two eight-core Xeon E5 processors per socket. With HyperThreading turned on, there are 32 threads in the CC2 instance and a much larger 88 EC2 compute units, almost three times the capacity of the original HPC instances (now called CC1) that Amazon was peddling only 16 months ago.
This CC2 instance – which is technically known as the cc2.8xlarge instance in Amazon-speak – presents 60.5GB of virtual memory and 3.37TB of disk storage to HPC applications; its cores run at 2.6GHz and it is using 10GE networking between the nodes, as before.
These CC2 instances cost $2.40 per hour if you buy them on demand (the same price as the CC1 instances 16 months ago), with lower prices if you reserve them ahead or buy them on the spot market. Amazon has also dropped the price of the CC1 instance to $1.30 per hour if the old iron can work for you. That's a 45.8 per cent reduction in price, by the way – something you can't get from real hardware vendors, rest assured. The CC2 instances can run Linux or Windows Server 2008 R2 instances that have pre-packaged to run atop Amazon's variant of the Xen hypervisor for x86 platforms.
Just to show that cloud computing is just as real for HPC as is a cluster of physical servers residing in your data center, Amazon took 1,064 of the CC2 instances and ran the Linpack benchmark test on it.
This cluster had 17,024 Xeon E5 cores running at 2.6GHz and had a peak theoretical performance of 354.1 teraflops. On the Linpack test, the cloudy cluster delivered 240.1 teraflops of sustained performance. It would cost you $2,554 to rent such a CC2 setup for an hour; with reserved and spot pricing, it could cost a lot less. Buying such a cluster would easily cost many millions of dollars, even with steep discounts. By the way, this Amazon virtual HPC cluster ranked number 42 on the Top 500 list.
As was the case with the CC1 instances, the CC2 instances are only available in Amazon's Northern Virginia data center at the moment. You can fire up a two-node CC2 cluster right now, but if you want more nodes that than, you have to contact AWS to get permission. Amazon says that it will be adding the CC2 HPC instances in its other data centers in 2012. ®