Cooling
Thanks to its higher thermal capacity per volume unit, about 3500 times larger than air, water based cooling is a must to maximize efficiency. Every Aurora module features an efficient and simple heat removal mechanism: liquid flowing across a plate which is thermally coupled with electronic devices. Aurora allows significant savings on the energy bill, with its flexible liquid cooling system: flow rates and coolant inlet temperature can be varied greatly, allowing economical operation with different infrastructures, and in different climates.
Energy savings can then be carried out also by making appropriate infrastructure choices:
- use of a single coolant circuit, which is simpler, and has no energy losses due to heat exchangers
- choice of efficient chillers, with appropriate sizing for operation in the "sweet spot" of performance
- Use of dry coolers, to benefit from conductive free cooling, maximizing the length of time when chillers are off, depending on the climate of the location where Aurora is installed. The efficacy of free cooling can vary, since it works using a temperature difference between the outside ambient, and outlet coolant.
The internal infrastructure is often no more than pipes and valves. It is made up of fewer parts than an ordinary air cooled system, which usually features also a liquid circuit per se.
Power Conversion
Power conversion in Aurora systems takes place in two separate steps. 400-200VAC is converted to 48VDC by high (93%) efficiency ACDC converters which share the liquid cooling infrastructure with the computer system. Chassis receive a 48VDC supply, which has an advantage in safety terms: Aurora is a low voltage equipment, hence safety regulatory tests are less strict than for VAC powered computers. 48VDC are in turn stepped down to 12VDC, reaching all Aurora boards, by a PSU within each chassis (DCDC Tray) with 93% efficiency. Failsafe operation is considered at each step: ACDC converters are redundant and ganged, while DCDC converters within each PSU board are n+1 redundant.
Noiseless Operation
Aurora has no fans, no mechanical moving parts within it. Local data storage is based on SSD units, also enhancing reliability and maximizing density. Aurora installations cannot be confined to data rooms. Aurora is easy to maintain and service, and looks good, no need to hide it .
Density
Packaging density is something more than mere miniaturization. By making systems denser and more compact, building infrastructure costs can be effectively reduced, in two ways: by making the data room no different than an ordinary room, and by reducing the sheer real estate required by the computing equipment. This of course has long term implications, by reducing the entire level of floor occupancy for a given system size and number of on site users.
FPGA-Based Reconfigurable Computing
The FPGA devices present on the AURORA Node cards allow in-system reconfigurability and update capabilities, with very low power demands (5W/Gflops). The on board FPGA provides DSP blocks specialized in operations such as floating and fixed point arithmetics and dynamic shift. Available DSP blocks permit to efficiently implement in hardware special functions such as:FIR and IIR filters, FFT functions, Data compression, SAR processing. Using all DSP blocks available, the on board FPGA can offer up to 80Gflops. Reconfigurable computing is a good fit for AURORA low node count installations, enhancing the system's versatility.
On Board Memory
AURORA features DDR3 memory chips directly soldered on the main node card. This has a number of advantages: longer and more reliable operation due to the more effective cooling, better signal integrity due to the direct electrical connection with the CPU, without the interposition of a connector, and finally better overall performance, owing to the direct connection CPU-memory chips with no registers or buffers in between.
Synchronization
OS Jitter is appearing always more often as a performance limitation in HPC installations, hindering CPU performance as systems grow in size. The effect is OS interference, primarily due to scheduling of daemon processes, and handling of asynchronous events such as interrupts, which can reduce performance. AURORA relies on three different levels of synchronization networks, in order to overcome the effects of lack of timing alignment between distributed parts of an application. AURORA sync networks permits use of collective policies of context switching at application level, Os kernel alignment. They also allow efficient debugging due to fast propagation of exceptions and traps. Deterministic propagation delay time of sync objects such as barriers, semaphores, global conditions, permits efficient checkpointing taking full advantage of newly introduced features such as in new generation schedulers. On the 3D Torus network low level synchronization is also possible for improved data transfer rate (deadlock avoidance, reduced jitter/noise, SIMD communication primitives)