Putting Shared Memory to the Test
The University of Oslo, with funding help from PRACE and the Norwegian Research Council, is working to solve some of the world’s big data challenges. To achieve its goals means the university has had to find innovative ways to pull together its computational resources to suit the demands of data intensive computing.
At the heart of the effort, which will eventually tackle global warming, life sciences and physics problem, are contributions from IBM and partner Numascale. According to a statement this week, the NumaConnect technology will create a large shared memory system across the IBM x3755 servers during the course of a pilot project to check out Numascale’s performance, a pilot that could set the bar for the company—and open a new set of customer doors.
The statement also noted that IBM will be “providing the server infrastructure along with deep HPC competence across key functions and Gridcore, a joint partner of Numascale and IBM, will build the new advanced computing resource.”
At the high level, Numascale presents the opportunity to string together commodity servers to build a cohesive shared memory system that shares all processor, memory and I/O resources under one unified OS. In this case, the NumaConnect cards and fabric will package IBM’s boxes into a single image under the ccNUMA system. According to Numascale, this creates a more cost-effective approach to creating large shared memory systems and is also easier to manage than a gaggle of servers.
Numascale offers what it calls “shared memory at cluster prices” through its NumaConnect line, which lets users build scalable shard memory systems on top of commodity Opteron-based servers. They also provide their NumaConnect SMP Adapter and NumaChip for vendors, including IBM and HP to use in their systems.
As noted earlier, this prototype NumaConnect system could set the tone for the future of the company, at least in the European academic market. This system, which will be funded as a PRACE First Implementation Phase (1IP) project, could prove itself as a viable option for European exascale targets. While exascale targets are important as national goals, at the moment, this could also demonstrate the value of shared memory systems as a key to solving data-intensive computing challenges—not just those that are reliant on FLOPS.
According to the Petter Bjorstad from the Department of Informatics at the University of Bergen, “Systems based on standard high volume servers interconnected with the full-blown Numascale technology will be distributed for a wide range of compute intensive applications that can today only be efficiently executed on systems costing significantly more.”