SoftLayer tailors HPC clouds with GPU lining
Have Tesla, will crunch
Hosting and cloud computing service provider SoftLayer is getting into the modern hybrid supercomputer racket with the launch of GPU-enhanced server instances.
SoftLayer, which is based in Dallas, was founded in 2005 and merged with hosting rival The Planet before being bought up by GI Partners, in 2010. GI Partners is a private equity company co-headquartered in Menlo Park and London that has $6bn in assets, including data center and telecom operators and real estate. SoftLayer is the resulting entity, and it has all the goodness of all the different firms wrapped up into a single package.
The company is privately held, but said back in December that it had over 100,000 servers under management in its various data centers and that it had over 25,000 customers worldwide using its infrastructure cloud, which it characterizes as the largest in the world. About half of those customers are located outside of the United States, and thus the company said last October that it was investing $75m over the next two years to build up data centers in Amsterdam and its network points of presence (PoPs) in London, Frankfurt, and Amsterdam to reach that demand.
SoftLayer had $85m in revenues in the quarter, and these metrics put it at a portion of the size of cross-state and OpenStack-obsessed rival Rackspace Hosting, which is based in San Antonio and which had $264.6m in revenues, 78,717 servers, and 161,422 customers in the third quarter of 2011.
Both SoftLayer and Rackspace developed their own infrastructure control software, but Rackspace has thrown in with NASA to create the OpenStack project, hoping to forge an open source community that can go up against and beat the cloudy infrastructure stack from Amazon's Web Services behemoth. SoftLayer's homegrown Infrastructure Management System tool can provision bare metal and virtual servers, and can be accessed either through the customer portal or through an API stack programmatically. And SoftLayer has no interest in moving to OpenStack or anything else at this time, mainly because the IMS tool speaks bare metal as well as virty.
Perhaps equally important for SoftLayer is that it allows customers to tweak the underlying hardware in hosted or virty servers so customers get the right iron for their specific jobs.
"Whatever you want, we'll crack open the box and drop it in," Simon West, chief marketing officer at SoftLayer, tells El Reg. SoftLayer has about a dozen different default configurations of server nodes in its data centers, and keeps enough parts on hand in its chop shop so it can spin up any non-standard requests for 20 servers in about two hours. It might take six to eight hours to spin up 100 custom nodes. SoftLayer gets its servers and parts from Super Micro.
The new GPU-goosed service from SoftLayer is available on the hosted servers, not cloud infrastructure, because for the most part customers who are running supercomputing code don't want another layer of software abstraction burning up processor cycles, memory capacity, and I/O bandwidth. The HPC nodes are based on 3U SuperMicro bare-bones machines that are equipped with the two of the new Xeon E5-2600 processors from Intel.
At the moment, SoftLayer is supporting Intel's Xeon 2520, 2650, and 2690 processors in the HPC nodes, and you can add one or two of Nvidia's Tesla M2090 GPU coprocessors to the box. These are the top-end Tesla coprocessors (at least until future coprocessors based on the "Kepler" GPU ship later this year) and deliver 665 gigaflops of double-precision floating point number crunching each.
A dedicated server node with a single Xeon E5-2620 processor with 16GB of memory runs $500 a month, and adding a single GPU and a 500GB SATA disk only boosts the price to $879 per month. Considering that those GPUs cost several thousand dollars a pop, clearly renting them from SoftLayer – or any other service provider that can spin up a hybrid ceepie-geepie cloud – is a good idea unless you know you can keep the infrastructure busy.
SoftLayer has two private Ethernet networks that it uses internally within and across its data centers, which run at Gigabit or 10 Gigabit speeds. Customers wanting to deploy HPC instances can do so on either network, but obviously it costs more to do so on the portions of the SoftLayer data centers that have 10GE switches reaching out to the server nodes. Like Amazon's Cluster Compute Instances, InfiniBand is not supported on the SoftLayer backbone, but if enough customers asked for it, then it would be easy enough to carve out a chunk of capacity to support InfiniBand.
A base operating system license, which includes a Standard Edition of Windows Server 2008, or CentOS, Red Hat Enterprise, or Debian Linux, is included in the price. SUSE Linux Enterprise Server is not supported at the time, but you can, if you really want to, run the XenServer hypervisor from Citrix Systems on the HPC instances if you really want to. You can, of course, also load your own software stack onto the images, says West, provided you have paid for them. You have to install the job scheduling and cluster management software of your choice on the nodes.
SoftLayer operates 13 data centers around the world, in San Jose, Seattle, Dallas (which has six), Houston, Washington DC, Singapore, and Amsterdam. The HPC instances are available in San Jose, Seattle, Dallas, Singapore, and Amsterdam, starting today. ®