Virtually every component in a server will have to change if we are to reach exascale computing. So says Ed Turkel, HP’s manager of Business Development, Service Providers and HPC Business. It’s not simply enough to scale out existing systems; for instance, HP’s largest HPC system, located at the Tokyo Institute of Technology and currently #5 on the Top 100 list, has 1,440 servers, each with two Xeon CPUs and two or three Nvidia GPUs and consumes 1.4MW. This delivers peak performance of 2.4 petaflops. Now, if it were simply possible to scale it up by 400 times to get into the exaflop range, it would require 576,000 servers and consume roughly 550MW. It would also be quite large at 400 times the current 160m2.
Clearly, adds Turkel, we need some new architectural advances. We must migrate to low-energy processors and for this he points to HP’s Project Moonshot, a new server infrastructure which addresses low-energy requirements. A traditional server is built around multiple processors, each with separate dedicated management, storage, fabric, cooling and power elements. Moonshot, in contrast, shares that infrastructure to decouple scale from complexity. In essence, you move from tens of servers per rack that share basically nothing to thousands of servers per rack that share almost everything. In addition, it takes advantage of the low-power embedded processors typically seen in portable devices.
The first actual hardware manifestation of Project Moonshot is the HP Redstone Server Development Platform, the start of a line of HP sever development platforms that feature extreme low-energy sever processors. The Redstone server in a 4U SL6500 chassis with shared power and cooling can hold 2,880 servers.
Although Moonshot is processor agnostic, this first Redstone platform is based on the EnergyCore ARM processors from Calxeda. That company was founded in 2008 and raised $48 million in one of the most significant fabless semiconductor start-up financings in recent history, while earning recognition by MIT as one of the 50 most innovative companies in the world. The company’s ARM-based EnergyCore processor consumes as little as 1.5W per SoC (system on a chip). When equipped with four cores, an on-chip fabric switch and management engine plus 4GB DRAM, a complete server node consumes only 5W (0.5W when idle).
An integrated EnergyCore fabric switch within every SoC provides up to five 10GB lanes for connecting thousands of EnergyCore server nodes into clusters capable of handling distributed applications at extreme scale. Every EnergyCore’s management engine, running on a separate, dedicated processor, performs real-time power optimisations by individually tuning 15 different power domains within the chip, idling or turning off entire subsystems when possible, yet instantly resuming operation within a matter of clock cycles.
The potential savings possible with Project Moonshot are quite impressive. HP took a system intended for applications that require ‘light scale out’ where fetching and delivering data is more important than computational power. Compared to a traditional x86 system, the equivalent Moonshot system using the Redstone Server consumed 89 per cent less energy, required 94 per cent less space, cost 63 per cent less and is 97 per cent less complex.
This HP server system is designed for testing and proof of concept, and the initial HP Redstone platform is expected to be available in limited volumes in the first half of this year.