Intel thinks integration as it moves toward exascale
- 15 February, 2013 21:25
Intel's aim is to lead the way in creating the first exaflop supercomputer, so it is making architectural improvements to its processors with the hope to reach that goal between 2018 and 2020.
The chip maker is integrating network and storage controllers into processors so supercomputers can serve up results faster. The integration will speed up communication links between processors and network and storage systems, and create a path to build an exaflop supercomputer, which will be around 50 times faster than today's fastest computer.
"It's not so much when we get there," said Joe Yaworski, fabric product marketing manager at Intel. "Along the way, if we solve each one of these issues, the overall HPC and cloud marketplaces will benefit."
With current chips and components, exaflop computers would be too expensive, big and power hungry. Intel is developing fabric technology to bring data communications closer to the processor, which will improve bandwidth and reduce the number of components in supercomputers. Fabric virtualizes I/O and ties together storage and networking in data centers, and an integrated controller will provide a wider pipe to reduce latency.
Companies such as Google, Facebook and Amazon are buying servers in large volumes and are consistently looking to reduce energy costs while improving system performance. Depending on system topology, fabrics enable intelligent data movement between processors, servers, memory, storage and appliances. By reorganizing traffic patterns, data can be moved in an energy-efficient way.
High-performance computers have fabric based on Infiniband, Ethernet and proprietary interconnects like Cray's Aries, while microservers have fabric based on Ethernet and PCI-Express. Advanced Micro Devices in its SeaMicro cloud servers offers Freedom Fabric, which supports standard network storage protocols.
In addition to improving bandwidth, the integration of fabric controllers will reduce power consumption and density of servers, Yaworski said.
"Fabrics are starting to become the next bottleneck. To address those bottlenecks ... you need to drive the fabric closer and closer to the processor," Yaworski said..
Moving forward, Intel will bring together expertise from multiple internal groups to offer an integrated fabric product in the "not-so-distant future," Yaworski said.
Virtually any networking technology like Ethernet or Infiniband can be applied to an HPC cluster, and external components create localized network connections, said Dean McCarron, principal analyst at Mercury Research.
Integrating the controller reduces the number of components on a motherboard. It also creates building blocks for large clusters in which fewer connections need to be established.
"You're not going off-chip for that first step to talk to the network controller. You can do it all on-chip," McCarron said.
Intel has been offering networking components for years, but its fabric business gained prominence when the company acquired Ethernet switching specialist Fulcrum Microsystems, the Infiniband assets from Qlogic early last year, and HPC interconnect assets from Cray in April last year.
The acquisitions were considered shrew d, giving Intel a range of products that are key to supercomputers, including processors, interconnects, networking and storage. Intel already offers the Xeon Phi co-processor, which has up to 60 cores, for supercomputers and that works alongside Xeon server CPUs.
Intel is initially focusing on integrating Infiniband, a high-speed interconnect used in supercomputers as an alternative to Ethernet and Fibre Channel. Intel recently updated its True Switch line of products, which is the first refresh of products originally acquired from Qlogic.
The new True Switch line brings the controller closer to the processor while improving performance via software-based fabric management and an additional card to provide more communication lanes. The new product can extend to 10,000 ports, and even more through specialized hardware, Yaworski said. The improvements can help speed up tasks such as computational fluid dynamics, which is key in designing aircraft.
True Switch is designed for the Xeon E5-2600 processor lineup code-named Romley, which is based on the Sandy Bridge microarchitecture. The True Switch products includes specialized chips that connect the host to the fabric, and also networking equipment such as adapters, switches, software and cables.
The company has an added advantage of researching and manufacturing its products, which helps in effective integration at the processor level, Yaworski said.
"We are uniquely positioned to make this connection between the fabric and the processor," Yaworski said.