Project: Scalable, cost and Energy Efficient DC/HPC technology with SMP capabilities for big-data challenges.

The project will focus on the development of a game-changing end-product for next-generation Data Centres (DCs) as well as the current High Performance market segment. In particular it targets three key components:_x000D__x000D_- Specialised networking hardware that allows off-the-shelf servers to be used to build cost-efficient COframe class shared memory computers. _x000D_- Scale-out DC efficiency management software for managing next generation DCs, with a particular focus on energy efficiency. _x000D_- System architecture COtaining preferred performance metrics for production environments in large HPC/DC systems._x000D__x000D_The project aims to demonstrate that not only can large shared memory systems be built at a fraction of the cost of bespoke COframe and SMP systems but they can also be deployed, managed and reconfigured in much the same manner as existing server farms and clusters._x000D__x000D_The primary driver for the initiation of this pan-European collaborative R&D project and the business opportunity for the consortium stems from the increased data-intensive (Big-data) problems that the scientific and industrial sectors are facing right now. As the data explosion is upon us, data-intensive methodologies such as pattern matching, scenario development, behavioural prediction and anomaly detection are developing in parallel to max the value of the collected data. These high-end data-intensive methodologies place huge demands on system-wide capabilities - memory size, memory bandwidth and memory latency. DCs addresses these demands in a sub-optimal manner by deploying two separate administrative units containing; i) commodity stand-alone servers, and ii) clustered servers or high-end COframes (SMP) for HPC offering. HPC and more general high-end computing is today performed on either SMPs, or on cluster systems comprising a large quantity of servers connected by high-speed interconnects. For solving the increased Big-data problems, cluster systems suffer from a low degree of resource utilisation, high administration costs and complex programming models when compared to SMPs. SMP systems, on the other hand, suffer from high purchase costs, typically 30x more than clusters. This situation does not meet the future demands for easy access to efficient and versatile problem solving._x000D__x000D_To achieve the CO objective we have established a pan-European consortium, made up of three companies that are experts in their respective fields of HW and SW. We will complement each other´s development activities perfectly with the technological know-how and execution capability that is needed to develop the most cost-effective and power-efficient solution for big-data problem solving in HPC facilities and next generation DCs. _x000D__x000D_Numascale (NO) has sprung out from the leading Norwegian computing and hardware company Dolphin Interconnect with historical roots in the former ground-breaking computing company Norsk Data. By leveraging on the findings from Dolphin they have over the past three years developed a unique shared memory concept for connecting commodity servers to provide HPC capacity that counters the memory issues found in cluster systems. When realised, this concept will surpass existing interconnect technology eg. InfiniBand by wide margins as the performance level will be close to SMP levels. In this project, Numascale will exploit the developed shared memory concept to research and develop a fully functioning interconnect chip that dramatically exceeds current market offerings._x000D__x000D_Concurrent Thinking (UK) has a long heritage in providing management and COtenance software for large-scale computing infrastructures. Many of its staff has over the past eight years been responsible for the design, implementation and support of hundreds of HPC cluster installations- including several Top 500 systems. In this project Concurrent-Thinking will draw on this wealth of experience in HPC and DCs to develop a highly scalable monitoring and management architecture able to orchestrate millions of monitored metrics simultaneously for optimised energy and resource utilisation. _x000D__x000D_Founded in 2002, Gridcore (SE, DE, US) has wide experience in integrating and delivering DC/HPC infrastructure to end users through large-scale corporate implementation and HPC solutions for scientific and technical computing. Gridcore is the developers and operator of the leading HPC on-demand service GOMPUTE. Gridcore has deep knowledge of industrial usage of HPC in a wide range of industries and will COtain the initial user and market front in this project._x000D__x000D_The commercial objectives of this project are to remove not only the costs associated with the acquisition of SMP systems, but demonstrate that such systems can be managed as a dynamic subsystem within the next-generation energy efficient DC with little or no additional running costs. The market opportunity for the consortium is significant and estimates indicate a total aggregated profit of €93M, 5 years after project end

Acronym SEED (Reference Number: 6988)
Duration 01/03/2012 - 30/06/2014
Project Topic We will be leveraging expertise in specialised interconnects, management software and HPC service provision to develop a disruptive solution aimed at Big Data applications. This enables Exabyte-capable COframe-class systems to be built from commodity servers at a fraction of current costs.
Project Results
(after finalisation)
Target 1) Optimizing the chip architecture for bandwidth and latency towards the goal of providing higher processing capacity and throughput of the systems embedding the Numascale hardware. _x000D__x000D_Achievements:_x000D_The first prototype card is complete and in advanced testing._x000D__x000D_A second generation card has been made utilizing a FPGA chip with larger capacity. This card has increased internal parallel capacities and additional micro engines that enable distributing tasks between the engines. This card also handles higher speeds at the host computer interface and more fabric lanes._x000D__x000D_The FPGA cards are complete implementations that can be refined to be a shippable product. The CO technology is contained in an FPGA an the final refinements are done by FPGA coding._x000D__x000D_Numascale expects to launch products based on the project results in 2015._x000D__x000D_The goals in the product plan are fully achieved by these results by finalization of the project. Numascale put more effort into the project that estimated in the project plan but also got better results than planned._x000D__x000D_Target 2) Define an architecture and method for accessing larger data areas the 256TB (defined by the CPU architecture) with a goal of being able to effectively handle up to 16EB (10^18 Bytes)_x000D__x000D_Achievement:_x000D_The activity was completed and the architectural decisions have been taken._x000D__x000D_No hardware has been implemented in the current FPGA design. As long as Numascale is working on FPGAs the implementation can be delayed until the demand is present or an ASIC implementation project is started._x000D__x000D_The goals of the product plan are achieved. Since a practical implementation has been postponed the effort in this was lower than planned. _x000D__x000D_Target 3) Sustainability and performance benchmarks for production environments_x000D__x000D_Achievements:_x000D_The project provided the P Gridcore with experience on SMP systems and its computing advantages and limitations._x000D__x000D_Final users had access at different steps of the project for direct evaluation of the Numascale architecture._x000D__x000D_Some of the most used applications by the end uses have been run on the new SMP platform, comparing its performance mostly against distributed memory computers such as Linux clusters._x000D__x000D_The technical goals of this task has been achieved, and a lot of very valuable information has been gathered. Several bottlenecks in the hardware has been isolated and the experience has been fed back to Target 1 activities and handled there. Ps and users have not always been pleased with the performance of the prototypes but Numascale has got valuable information for improving its products._x000D__x000D_Target 4) Proof of concept_x000D_Achievements:_x000D_The concept is proven for a number of important application areas._x000D__x000D_Some key industry benchmarks and user codes have been run on the Numascale/Gridcore System. _x000D__x000D_In collaboration with Statoil good results have been achieved on a code that is promising to be very important for the oil and gas industry. _x000D__x000D_In collaboration with a racing team we have shown good results on a STAR-CCM+ application. _x000D__x000D_Other codes that are on the work-list are: ANSYS Fluent, ANSYS Polyflow and OpenFOAM. Numascale has tested some important Big Data related SW packages including Apache Spark, R in an integrated environment and MonetDB. Numascale will continue this work after the end of the project and introduce a Scale-Up and Scale-Out capability based on Spark and Hadoop._x000D__x000D_Researchers from RWTH Aachen continued testing the TrajSearch code and the code has been now been run on up to 1024 cores with linear scaling. _x000D__x000D_The Concurrent software runs on Numascale's internal systems and are verified to be productive on the platform._x000D_
Network Eurostars
Call Eurostars Cut-Off 7

Project partner

Number Name Role Country
3 Concurrent Thinking Ltd Partner United Kingdom
3 Gridcore Aktiebolag Partner Sweden
3 Numascale AS Coordinator Norway