Periodic Reporting for period 1 - AERO (Accelerated EuRopean clOud)
Reporting period: 2023-01-01 to 2024-06-30
Cloud computing has become pervasive across the computing industry serving various domains ranging from HPC to Big Data and IoT. However, cloud service providers and hardware and software ecosystems are currently mainly driven by US companies. Hence, European companies that follow the established cloud-migration trend to exploit the inherent technological advancements, face critical challenges : a) strong dependencies to overseas technology providers, b) increased security and data privacy concerns, and c) limited flexibility in designing custom software and hardware that better satisfies the requirements of the European market and ecosystem. Recognizing these limitations of the current status-quo, the European Commission has launched the European Chip Act as a driver to achieve European sovereignty in chip design and computing infrastructure. At the center of this activity lies the EU Processor Initiative (EPI) that spearheads the development of the first EU made chip and compute units. Still, for EPI and its associated activities to be successful towards the goal of EU sovereignty in computing, tremendous work must be undertaken for the upbringing of the software ecosystem that CSPs offer. Software support that ranges from operating systems to compilers and runtimes, system software, and auxiliary software deployment services for cloud computing need to progress in a rapid pace alongside hardware development, in order to ensure that EU customers can migrate to the envisioned EU cloud seamlessly and safely without compromising any requirements or SLAs in terms of performance, energy efficiency, security, and usability.
AERO has the mission of enabling the future heterogeneous EU cloud infrastructure. Towards this, the project will develop - to a high TRL level - all components necessary to achieve out-of-the-box heterogeneous execution of the cloud ecosystem on the European processor. AERO will open-source all enabling technologies and release software prototypes to further accelerate development and adoption of the EU cloud via well-established interfaces that make heterogeneous execution accessible. The provision of these AERO technologies, will enable potential EU cloud users and customers to: a) utilize early prototypes to test existing and new codebases, b) optimize their codebase to fully utilize the benefits of heterogeneous hardware acceleration in the EU cloud, c) harness the existing security elements of the European processor as well as design and prototype next-generation post-quantum encryption modules, d) provide feedback to EPI and other relevant projects towards faster adoption, and e) contribute to the collective EU effort towards compute sovereignty.
More specifically, AERO aspires to:
1) Upbring and optimize the software ecosystem based on managed programming languages (mainly Java) – and their associated cloud-based infrastructure - on the heterogeneous EU cloud.
2) Upbring and optimize the software ecosystem based on native programming languages and runtimes (mainly SYCL , OpenCL, and DPC++/OneAPI) for the heterogeneous EU cloud.
3) Upbring and optimize the necessary OS, driver, and virtualization support infrastructure for EU cloud deployment.
4) Develop the necessary cloud-ready components for a typical state-of-the-art cloud deployment including management, orchestration, and monitoring.
5) Research and develop open interfaces for exploiting accelerated hardware components in the EU processor for performance and security.
6) Accelerate the adoption of the EU cloud ecosystem via upstreaming to open-source projects, communicating and disseminating AERO’s results to industry, academia, and standardization bodies.
Upon successful completion, AERO will complete the continuous path from EU flagship projects to a future heterogeneous EU cloud infrastructure. To achieve that, AERO combines pivotal expertise from EPI with world-leading EU-derived software technologies to pave the way for the envisioned highly performing, energy efficient and operational-ready EU cloud, catalyzing the creation and provision of a wide range of improved cloud services in various application domains (e.g. database acceleration, space exploration, automotive, etc.).
AERO has the mission of enabling the future heterogeneous EU cloud infrastructure. Towards this, the project will develop - to a high TRL level - all components necessary to achieve out-of-the-box heterogeneous execution of the cloud ecosystem on the European processor. AERO will open-source all enabling technologies and release software prototypes to further accelerate development and adoption of the EU cloud via well-established interfaces that make heterogeneous execution accessible. The provision of these AERO technologies, will enable potential EU cloud users and customers to: a) utilize early prototypes to test existing and new codebases, b) optimize their codebase to fully utilize the benefits of heterogeneous hardware acceleration in the EU cloud, c) harness the existing security elements of the European processor as well as design and prototype next-generation post-quantum encryption modules, d) provide feedback to EPI and other relevant projects towards faster adoption, and e) contribute to the collective EU effort towards compute sovereignty.
More specifically, AERO aspires to:
1) Upbring and optimize the software ecosystem based on managed programming languages (mainly Java) – and their associated cloud-based infrastructure - on the heterogeneous EU cloud.
2) Upbring and optimize the software ecosystem based on native programming languages and runtimes (mainly SYCL , OpenCL, and DPC++/OneAPI) for the heterogeneous EU cloud.
3) Upbring and optimize the necessary OS, driver, and virtualization support infrastructure for EU cloud deployment.
4) Develop the necessary cloud-ready components for a typical state-of-the-art cloud deployment including management, orchestration, and monitoring.
5) Research and develop open interfaces for exploiting accelerated hardware components in the EU processor for performance and security.
6) Accelerate the adoption of the EU cloud ecosystem via upstreaming to open-source projects, communicating and disseminating AERO’s results to industry, academia, and standardization bodies.
Upon successful completion, AERO will complete the continuous path from EU flagship projects to a future heterogeneous EU cloud infrastructure. To achieve that, AERO combines pivotal expertise from EPI with world-leading EU-derived software technologies to pave the way for the envisioned highly performing, energy efficient and operational-ready EU cloud, catalyzing the creation and provision of a wide range of improved cloud services in various application domains (e.g. database acceleration, space exploration, automotive, etc.).
In the first 18 months, the technical work packages focused on the following:
• UNIGE ported their main scientific algorithm, Period Search, to TornadoVM to leverage GPU acceleration, showing promising performance results. KTM, showed functional equivalency of their Vehicle Information Service pilot on an ARM platform with Neoverse N1 cores to execution on x86 platforms, and developed & tested two additional use cases. Finally, SED successfully concluded the development of User Defined Java Function (UDF) support for their distributed PostrgeSQL fork and initiated working on enabling GPU acceleration of core DB executor functionality and improving the database stack.
• VOSYS is porting VOSySmonitor to the ARM Neoverse architecture, ICCS extended the Linux kernel to generate/manage intermediate-sized translations on ARMv8/9-A systems, and FORTH developed QEMU support for RISC-V virtualization in Kubernetes. In addition, CPLAY extended and open sourced the oneAPI Construction Kit (OCK) with full support for OpenCL, and UNIPI optimised and released the CKKS benchmark for the Microsoft SEAL library, and designed a software acceleration library for ARM processors with vectorized computation capabilities like Rhea.
• UNIMAN & FORTH ported and are currently optimising TornadoVM & TeraHeap, respectively. CPLAY provided numerous contributions to the DPC++ toolchain towards adding AArch64 support as SYCL-targets, while RHAT developed further its Mandrel and Quarkus products, releasing the first fully supported build of Mandrel supporting AArch64.
• UBI & FORTH completed the porting of their MAESTRO & ExaFlow/Knot frameworks, respectively, and UBI extended MAESTRO to support GPU hardware acceleration. ICCS deployed the AERO Function-as-a-Service platform, which entails a state-of-practice software stack, a suite with realistic serverless workloads ported to AArch64 and a novel serverless load generator that can generate representative series of requests suitable for evaluating serverless prototypes.
• WP6 focused on securing access to Rhea alternative platforms and started working on setting up an appropriate continuous integration & validation system. In addition, all partners performed initial testing & evaluation of the developed components.
All milestones were successfully met & deliverables were submitted without any major delays.
• UNIGE ported their main scientific algorithm, Period Search, to TornadoVM to leverage GPU acceleration, showing promising performance results. KTM, showed functional equivalency of their Vehicle Information Service pilot on an ARM platform with Neoverse N1 cores to execution on x86 platforms, and developed & tested two additional use cases. Finally, SED successfully concluded the development of User Defined Java Function (UDF) support for their distributed PostrgeSQL fork and initiated working on enabling GPU acceleration of core DB executor functionality and improving the database stack.
• VOSYS is porting VOSySmonitor to the ARM Neoverse architecture, ICCS extended the Linux kernel to generate/manage intermediate-sized translations on ARMv8/9-A systems, and FORTH developed QEMU support for RISC-V virtualization in Kubernetes. In addition, CPLAY extended and open sourced the oneAPI Construction Kit (OCK) with full support for OpenCL, and UNIPI optimised and released the CKKS benchmark for the Microsoft SEAL library, and designed a software acceleration library for ARM processors with vectorized computation capabilities like Rhea.
• UNIMAN & FORTH ported and are currently optimising TornadoVM & TeraHeap, respectively. CPLAY provided numerous contributions to the DPC++ toolchain towards adding AArch64 support as SYCL-targets, while RHAT developed further its Mandrel and Quarkus products, releasing the first fully supported build of Mandrel supporting AArch64.
• UBI & FORTH completed the porting of their MAESTRO & ExaFlow/Knot frameworks, respectively, and UBI extended MAESTRO to support GPU hardware acceleration. ICCS deployed the AERO Function-as-a-Service platform, which entails a state-of-practice software stack, a suite with realistic serverless workloads ported to AArch64 and a novel serverless load generator that can generate representative series of requests suitable for evaluating serverless prototypes.
• WP6 focused on securing access to Rhea alternative platforms and started working on setting up an appropriate continuous integration & validation system. In addition, all partners performed initial testing & evaluation of the developed components.
All milestones were successfully met & deliverables were submitted without any major delays.
Initial results beyond the state of the art include:
• GPU accelerated space exploration scientific algorithms
• GPU accelerated user supplied Java functions in distributed PostgreSQL
• Linux kernel extensions for intermediate-sized translations on ARMv8/9-A systems
• A benchmark of the CKKS functions for x86_64 and AArch64 platforms
• TornadoVM extensions
• TeraHeap ported to AArch64
• SYCL support for AArch64 & RISC-V hardware
• Quarkus & Mandrel improved support for AArch64
• MAESTRO extensions for GPU acceleration, deployment to ARM and RISC-V devices, andserverless deployments
• Exaflow/Knot extensions to support AArch64 systems
• AWS Firecracker patches for deployment on latest AArch64 systems
• A novel serverless load generator
All developed code (except for VOSySmonitor and MAESTRO) is open source and currently released in the AERO GitHub.
• GPU accelerated space exploration scientific algorithms
• GPU accelerated user supplied Java functions in distributed PostgreSQL
• Linux kernel extensions for intermediate-sized translations on ARMv8/9-A systems
• A benchmark of the CKKS functions for x86_64 and AArch64 platforms
• TornadoVM extensions
• TeraHeap ported to AArch64
• SYCL support for AArch64 & RISC-V hardware
• Quarkus & Mandrel improved support for AArch64
• MAESTRO extensions for GPU acceleration, deployment to ARM and RISC-V devices, andserverless deployments
• Exaflow/Knot extensions to support AArch64 systems
• AWS Firecracker patches for deployment on latest AArch64 systems
• A novel serverless load generator
All developed code (except for VOSySmonitor and MAESTRO) is open source and currently released in the AERO GitHub.