September 17, 2013
Large shared memory systems are often a novelty and some, including the well-known SGI Altix UV1000 “Blacklight” system at PSCC, for instance, have received a great deal of attention due to their ability to address specific high performance computing workloads.
While Blacklight and similar large coherent shared memory systems are driven by hardware-based approaches to creating unified approaches, Cray veered off at the software fork, deciding to create similarly focused systems at the software level. This morning they announced two different pre-configured setups on their Cray CS300 systems that will make room for workloads that have a need for larger memory within a single operating system instance.
By tapping their longtime partner, virtualization-based shared memory system software vendor ScaleMP, the supercomputer maker says that they’re both able to broaden their cluster architectures to support larger memory applications—all without the risk of going it alone with a more investment-heavy hardware-based approach to creating shared memory systems. ScaleMP’s vSMP Foundation software snaps together commodity x86 servers to create a single virtual system, which provides an alternative to (what are usually more expensive) SMP systems.
More specifically, today Cray rolled out their CS300 SMP product, which is a shared memory parallel system that sports (upgrade aside) a basic 360 Xeon cores, 4.75 TB of memory and the ability to tap single or dual-rail FDR InfiniBand.
The other, a Cray CS300 LMS (a large memory system) manages these workloads via direct memory access without harnessing high core counts in high-RAM-demand environments that chug along on simpler dual and quad-socket systems. Cray says these stepped-down systems can scale from 4.75 TB to 8.75 TB of memory and harness 20-32 Xeon cores. These are standard air-cooled CS300s that have wrapped around ScaleMP’s vSMP Foundation software, which is at the core of the HPC system virtualization vendor’s business.
Cray’s Barry Bolding admitted that while there are certainly some HPC applications that can’t be broken up across conventional clusters, it’s a small number—perhaps around 10% at the most. Still, these workloads require large memory architectures, but the hardware-based approach that SGI, for example, takes can add significant expense and is not as simple to maintain (i.e. updates to the system required with new processor generations, etc.).
Interesting that a company known for its supercomputing hardware history would turn on its roots to favor software, but without a sizable known market, Bolding says the investments required to do what rival SGI has done with its Numalink technology are significant—and ScaleMP approach offers lower cost on all ends—and no real risk for Cray to add to its ranks of options for the CS300 line.
While Bolding said that creating their own hardware-based approach to large-memory systems isn’t out of the question (and has been an idea that’s been bandied about for some time already) this shouldn’t be seen as a definitive first step in that direction. While one can be certain Cray will assess the adoption and success of this addition to the CS300 line in their eventual evaluations of the hardware-shared memory field, Bolding says that there are advantages of the software-based take on shared memory—most notably, dramatically lower costs and, as mentioned previously, fewer maintenance hassles.
On the cost front, Bolding notes that the addition of ScaleMP’s shared memory software, which comes integrated and ready to roll from the factory, does not add significant cost. The systems range from around $200k for the large memory configuration and upwards from $300k for the SMP version. While Cray is not expecting this addition to shatter sales records, it does offer something to differentiate its CS300 portfolio—and to further test the shared memory waters.
In a conversation this morning with ScaleMP’s founder, president and CEO, Shai Fultheim, we talked about the value of the software-based approach to shared memory system creation. As Fultheim told us, their virtualization approach reduces overall system (CAPEX) and management complexities (OPEX) costs. Specifically, he says that their vSMP Foundation aggregates up to 128 x86 systems to create a single system with up to 32,768 cpus and up to 256 TB of shared memory.
Fultheim also noted that these approaches go beyond high performance computing environments. Big data, analytics and database-driven companies are looking to the benefits of the software-based paradigm of aggregating the common x86 systems into one single x86 virtualized system reach performance, management and efficiency targets.
ScaleMP has partnered with Cray in the past, beginning in 2009, via a joint solution for HPC customers to operate a shared-memory, deskside supercomputer that could scale up to 128 cores and 1TB of shared memory.
“Cray has always had a special relationship with the most demanding users, redefining the requirements for high-end systems. With this collaboration, Cray’s new large memory and shared memory systems will allow a broader technical computing audience to benefit from the ability to address larger workloads and get faster results,” said Fultheim.
10/30/2013 | Cray, DDN, Mellanox, NetApp, ScaleMP, Supermicro, Xyratex | Creating data is easy… the challenge is getting it to the right place to make use of it. This paper discusses fresh solutions that can directly increase I/O efficiency, and the applications of these solutions to current, and new technology infrastructures.
10/01/2013 | IBM | A new trend is developing in the HPC space that is also affecting enterprise computing productivity with the arrival of “ultra-dense” hyper-scale servers.
Ken Claffey, SVP and General Manager at Xyratex, presents ClusterStor at the Vendor Showdown at ISC13 in Leipzig, Germany.
Join HPCwire Editor Nicole Hemsoth and Dr. David Bader from Georgia Tech as they take center stage on opening night at Atlanta's first Big Data Kick Off Week, filmed in front of a live audience. Nicole and David look at the evolution of HPC, today's big data challenges, discuss real world solutions, and reveal their predictions. Exactly what does the future holds for HPC?