Periodic Reporting for period 2 - OPERA (lOw Power heterogeneous architecture for nExt generation of smaRt infrastructure and platforms in industrial and societal Applications)
Okres sprawozdawczy: 2017-06-01 do 2018-11-30
The aim of OPERA is to create a cooperative, secure, reliable, customized, and low power computing platform that is able to address the challenges imposed by the future convergence of datacentre computing, embedded devices and sensors. To this end OPERA is developing a new generation of high-density servers. These servers are the basic “bricks” for implementing a scalable Low Power datacentre. At the base of these modules there are heterogeneous architectures that use ARM, Intel, POWER8 processors and FPGA accelerators for optimized functions and computation offloading. These devices form a mix of processing elements and accelerators designed for achieving significantly better energy efficiency at the cost of flexibility. Initially an integration of the CAPI technology has been planned. Due to Altera acquisition by Intel in the end of 2015, IBM decides to stop working with this FPGA brand and switched completely to Xilinx as their main provider for the work they are doing with CAPI. Remediation was to use RDMA board from Mellanox to interconnect the multiple architectures used in OPERA VDI use case. The OPERA project aims also at exploiting high-speed optical links to provide interconnections between the accelerators and the external system with no performance limitations.
These technologies form the substrate used by OPERA to implement an efficient workload decomposition system that automatically and dynamically executes tasks on the most suitable processing element. The result of the integration of this workload decomposition system is to make optimal use of computing resources and application in a heterogeneous architecture so that it is possible to guarantee high performance computing capabilities in a scalable (cloud) energy efficient infrastructure perspective.
The project has developed a highly parallel embedded system based on an ultra-low power System-on-Chip (SoC) which can be deployed on various application domains. A new version of the ULP platform has been recently added to the OPERA project portfolio. This new system is based on a new STMicroelectronics R&D SoC, named Orlando, specifically designed for the acceleration of convolutional neural networks. This new device allows the implementation of a new class of applications that contributes to the improvement of the state of the art and the results of the project in terms of efficiency and applications performances.
The strong integration of sensors, radio communication interfaces, and computing resources allows spreading such small intelligent devices in several contexts. The availability of reconfigurable wireless communication interfaces guarantees that these devices can be flexibly adapted to different contexts (e.g. smart cameras can dynamically switch between Wi-Fi and 3G connections depending on the quality of the signal optimizing the power consumption dynamically). Highly parallel on-board processing elements provide enough computing power for pre-processing raw data that can be later sent to a remote facility (scalable Low Power datacentre) for subsequent analysis. OPERA leverages these capabilities to deliver real-world applications where to test the proposed platform and that can easily be extended to other application scenarios
These technologies form the substrate used by OPERA to implement an efficient workload decomposition system that automatically and dynamically executes tasks on the most suitable processing element. The result of the integration of this workload decomposition system is to make optimal use of computing resources and application in a heterogeneous architecture so that it is possible to guarantee high performance computing capabilities in a scalable (cloud) energy efficient infrastructure perspective.
The project has developed a highly parallel embedded system based on an ultra-low power System-on-Chip (SoC) which can be deployed on various application domains. A new version of the ULP platform has been recently added to the OPERA project portfolio. This new system is based on a new STMicroelectronics R&D SoC, named Orlando, specifically designed for the acceleration of convolutional neural networks. This new device allows the implementation of a new class of applications that contributes to the improvement of the state of the art and the results of the project in terms of efficiency and applications performances.
The strong integration of sensors, radio communication interfaces, and computing resources allows spreading such small intelligent devices in several contexts. The availability of reconfigurable wireless communication interfaces guarantees that these devices can be flexibly adapted to different contexts (e.g. smart cameras can dynamically switch between Wi-Fi and 3G connections depending on the quality of the signal optimizing the power consumption dynamically). Highly parallel on-board processing elements provide enough computing power for pre-processing raw data that can be later sent to a remote facility (scalable Low Power datacentre) for subsequent analysis. OPERA leverages these capabilities to deliver real-world applications where to test the proposed platform and that can easily be extended to other application scenarios
The OPERA architecture, in its full implementation, has been realized with a local processing node, based on hardware accelerators and programmable DSPs, heterogeneous in its nature. This node is based on a ULP computing device. The applications on the ULP architecture has been optimized for the highest performance on the embedded platform. Through a configurable wireless connection it transfers to the server part of the computation, optimizing the overall execution. The server, based on the OPERA low power datacentre, accelerates specific functions with programmed FPGA cards. From the software perspective, the entire system implements a state-of-the-art CNN architecture, with different levels of accuracy. In the truck use case, also, the image processing application is enhanced accessing to hardware accelerators. For the objective related to new generation servers we demonstrated through the validation in the three use cases, that low power processors together with accelerators is a more efficient way to design new architectures.
On the OPERA architecture, executed in the use cases, different metrics for the power efficiency evaluation have been adopted. The energy cost of the state of the art implementation and usage (network, people involved, displays for video streaming) has been compared to the consumption of OPERA solution, not only in term of electric power consumption, but also regarding energy saving for reduced need of human control and continuous monitoring. The optimization of the power consumption has been realized through the implementation of the workload management, the heterogeneous computing, and the configurable wireless channel.
The workload management has been demonstrated through the VDI use case. Several technologies, including the containerization of legacy workloads for integration with cloud management systems and post-copy container migration for implementing the dynamic policies of the power aware scheduler have been implemented.
On the OPERA architecture, executed in the use cases, different metrics for the power efficiency evaluation have been adopted. The energy cost of the state of the art implementation and usage (network, people involved, displays for video streaming) has been compared to the consumption of OPERA solution, not only in term of electric power consumption, but also regarding energy saving for reduced need of human control and continuous monitoring. The optimization of the power consumption has been realized through the implementation of the workload management, the heterogeneous computing, and the configurable wireless channel.
The workload management has been demonstrated through the VDI use case. Several technologies, including the containerization of legacy workloads for integration with cloud management systems and post-copy container migration for implementing the dynamic policies of the power aware scheduler have been implemented.
The OPERA project largely contributed to spread heterogeneous and low power technologies over the computing spectrum, ranging from single devices tailored for cyber-physical systems to modular high-performance high-density servers to FPGA accelerators. Such broad spectrum of technologies put in place and successfully integrated by OPERA, allowed to greatly reinforce the position of Europe as leader in low-power computing technologies. Looking at the data center side, OPERA had a great impact on promoting the adoption of low-power high-performance servers to largely improve the energy efficiency, as well as to support modern applications. When compared to the state-of-the-art server 2013 technology, OPERA solutions showed an impressive improvement in terms of energy efficiency and performance. The integration of hardware and software solutions allowed to greatly increase the energy savings of servers, especially in a data center with multiple nodes. This is of worth for many use cases and applications, also not directly taken into consideration by OPERA project. In all the cases where the connection to the power grid is not possible, OPERA demonstrated the effectiveness of its energy-aware server technology. Looking to the case of large data centers, the same solutions provided the capability of managing large base of users with a reduced number of physical servers.