H2020 European ExaNeSt and ExaNoDe Shows Successful Results
Virtual Open Systems enriching experience through HPC research projects
Success story
Recently in the 2019 second quarter, Virtual Open Systems has showcased with its partners the results of ambitious exascale H2020 FET projects, ExaNeSt and ExaNoDe, targeting the development of new technologies, both hardware and software, for the next generation HPC. They are part of the European effort towards the exascale computing, aiming to design and prototyping HPC compute nodes capable to perform at least one exaFLOPS. It's an European effort which also includes the development of software, programming models and the surrounding ecosystem to ease the exploitation of such a huge computing power.
Challenge
ExaNeSt has addressed primarily the problem of the complex HPC interconnection networks, storage, and cooling system, while ExaNoDe has looked more deeply at enhancing the architecture of the compute node itself, targeting the design and manufacturing of a chiplet (complex SoCs with accelerators and application processor functionalities integrated into a single 3D substrate). Both projects have brought considerable innovations also in software technologies, with ExaNeSt focusing on algorithms optimization and storage file systems, ExaNoDe on programming models.
Contribution
In both projects, Virtual Open Systems has extended its experience in the virtualization domain. Specifically, in ExaNoDe a new solution for virtual machine incremental and periodic checkpointing has been designed and developed, which is the basis to provide a reliable cloud infrastructure to handle different user workloads in the same physical computing node. In both projects, the Virtual Open Systems technology VOSySremote has been extended to provide support for accelerators APIs which are relevant for HPC use cases, like for instance RDMA and OpenCL API. It allows to virtualize the hardware resources used by such libraries, providing an additional isolation and management layer that can help in securing and isolating the different workloads running in the same compute node. At last, in ExaNeSt, Virtual Open Systems integrated its software networking switch, VOSySwitch, into Openstack and enhanced its core functionalities to support multiple cores/threads.