For the second time in five years, Appro has been tapped to provide the National Nuclear Security Administration (NNSA) with HPC capacity clusters for the agency's Advanced Simulation and Computing (ASC) and stockpile stewardship programs. The Tri-Lab Linux Capacity Cluster 2 (TLCC2) award is a two-year contract that will have the cluster-maker delivering HPC systems across three of the Department of Energy's national labs. The deal is worth tens of millions of dollars to Appro and represents the biggest contract in the company's 20-year history.

Over the past several years, Appro has been on a hot streak with the NNSA labs. The Peloton contract with Lawrence Livermore Lab back in 2006 was the company's first big win in the NNSA arena. That was back in the day when 50 to 100 teraflops of capacity was considered monster-sized HPC. Peloton was a prelude to the initial Tri-Lab Linux Capacity Cluster (TLCC) award for Appro in 2007, which was the first time the NNSA decided to use a single cluster architecture, software stack, and system vendor for all three NNSA weapons labs -- Lawrence Livermore (LLNL), Los Alamos (LANL) and Sandia (SNL) National Laboratories. That contract had Appro delivering 426 teraflops of capacity spread across nine clusters, and sending $26.1 million into the company's revenue stream over a two-year period.

The sequel to the 2007 contract, the TLCC2 award announced today, will deliver something north of 3 petaflops of aggregate capacity to the three DOE labs. But if the contract exercises all its options, the amount will double to around 6 petaflops -- that according to Appro VP John Lee, who heads up the company's Operations & Advanced Technology Solutions Group. The total spend for the TLCC2 work will hinge on ASC funding in fiscal year 2012, but will end up being a good deal larger than the $26.1 million garnered for 2007 contract, says Lee.* "This is the largest single contract in Appro has ever been awarded," he told HPCwire.

The mission of the new clusters is the same as in the original TLCC contract: to provide a capacity HPC infrastructure for computer simulations in support of the nation's aging nuclear deterrent. The capacity systems are meant to be the workhorses that support the more rarified software running on the labs' capability supercomputers. Here we're referring to machines like Sequoia, the 20-petaflop Blue Gene/Q system to be deployed at Lawrence Livermore in 2012. Sequoia and machines of that ilk are designed to run the big scaled-out codes for the nuclear weapons models and uncertainty quantification simulations, both of which require highly tuned supercomputing technology. Meanwhile, the capacity clusters will be employed for algorithm development as well as to provide a general-purpose compute pool for basic science codes.

HPC capacity for this procurement is specified in scalable units, a concept the labs came up with in 2007 to define a unit of compute infrastructure than can be added in a modular fashion. One scalable unit delivers 50 peak teraflops of hardware (a scalable unit in the original Appro contract was just 20 teraflops), where a cluster may be anywhere from one to 18 units. The idea behind this approach is to simplify procurement, deployment and management of the systems across the three labs.

The Peloton and original Tri-Lab contract had Appro delivering AMD Opteron-based servers glued together with Mellanox DDR InfiniBand. This time around though, the labs will be getting Intel Xeon-based servers and QLogic QDR. Specifically, all the clusters will be based on a new version of Appro's GreenBlade system using the upcoming eight-core Sandy Bridge-EP Xeon CPUs, with two processors and 32 GB of memory per node.

The company is also putting these same GreenBlade servers in the upcoming Gordon supercomputer at the San Diego Supercomputer Center (SDSC), says Appro's Lee. According to him, Gordon and the new Tri-Lab clusters will be the first two Sandy Bridge-equipped HPC server deployments in the world.

As an aside, Appro will also be offering a cluster products with the upcoming AMD "Interlagos" (Opteron 6200 series) CPUs. Those system will be especially suited to applications that can spread easily across lots of cores and memory in SMP fashion -- up to 16 per cores/processor and, theoretically at least, up to four processors per node, along with the associated memory.

But for HPC applications that are sensitive to memory performance, as in the Tri-Lab workloads, the customers will tend to favor the new Xeons, says Lee. Both Interlagos and Sandy Bridge EP offer four memory channels, however in the case of the Opteron, a 16-core Interlagos processor is packaged as two 8-core dies stuffed into the same socket. So each chip has direct access to only two channels. If cores from one chip wants to access memory attached to its companion chip, it has to go through the HyperTransport bus. With the monolithic Sandy Bridge die, all four memory channels are directly accessible to all the cores in the socket.

The new Tri-Lab clusters will be outfitted with QLogic QDR InfiniBand hardware, ditching the Mellanox parts in the TLCC and Peloton systems. In this case, the labs are favoring QLogic gear based on impressive scalability and performance results on some of their existing QLogic-equipped systems, in particular, the 23-thousand-core Sierra cluster at Lawrence Livermore.

If the contract timeframe had been a little later, the labs might have been tempted to go with Mellanox and their new FDR (Fourteen Data Rate) InfiniBand solutions, but the adapters, switches and related software will just be hitting the streets in the third quarter of this year. Since the initial TLCC2 systems are scheduled to be installed at all three labs before the end of September 2011, that would have made FDR InfiniBand an iffy proposition. However, the new GreenBlades will support PCIe 3.0, so the systems could conceivably be upgraded to FDR and eventually EDR adapters down the road.

Sandy Bridge-EP availability also adds some risk to the deployment schedule. Supposedly the new Xeons should be rolling out of the fabs in Q3 this year. But if they're not in full production, Appro and the labs are hoping that they'll be able to get enough of the CPUs to begin deliveries. Multiple systems are slated to be deployed at each site, with the largest machine, a 900 teraflop cluster, initially going to Lawrence Livermore

They'll be some GPUs in the mix as well. All three labs have expressed an interest in accelerators for some of these clusters. Initially though, only Los Alamos will be installing such a system, in this case, a 324-node cluster, equipped with 648 of the latest NVIDIA Tesla M2090 GPUs. That's just for starters; Los Alamos is also hoping to purchase a GPU-cluster about twice that size. The three labs are also interested in Intel's forthcoming "Knights Corner" accelerators, but they are not in the deliverables on any planned systems at this point, says Lee.

*The NNSA has subsequently announced that the initial contract award will be for $39 million, with up to $89 million possible if all options are exercised.