Heterogeneous I/O for Scale
Heterogeneous I/O for Scale
We are working to develop a proof of concept for an FPGA-based I/O intermediary. The potential impact would be to change the way data ingestion happens when using remote storage locations. In view of the enormous amounts of data to be employed in the future for data analytics, it is crucial to efficiently manage the flow in order to harness the computational power provided by high-performance computing (HPC) facilities.
One of the common aspects of all data-intensive applications is the streaming of recorded data from remote storage locations. This often imposes constraints on the network and forces a compute node to introduce complex logic to perform aggressive caching in order to remove latency. Moreover, this substantially increases the memory footprint of the running application on a compute node. This project, abbreviated to ‘HIOS’, aims to provide a scalable solution for such data-intensive workloads by introducing heterogeneous I/O units directly on the compute clusters. This makes it possible to offload the aggressive caching functionality onto these heterogeneous units. By removing this complicated logic from compute nodes, the memory footprint decreases for data-intensive applications. Furthermore, the project will investigate the possibility of including additional logic, coding/decoding, serialisation, I/O specifics, directly onto such units.
An integral part of the project will be the ability to integrate the units developed directly with current HPC facilities. One of the main outcomes of the project will be the reduced time required to extract insights from large quantities of acquired information, which, in turn, directly impacts society and scientific discoveries.
HIOS is one of the 170 breakthrough projects receiving funding through the ATTRACT initiative. ATTRACT, which is part of the European Union’s Horizon 2020 programme, is financing breakthrough ideas in the fields of detection and imaging.
During the past year, the first phase of the ATTRACT project, we performed testing and validation of the TCP/IP network stack. This targeted an FPGA and could be used to integrate with Extoll’s existing solution. At the same time, a range of compression/decompression algorithms (e.g. LZMA, LZ4, zstd, Deflate) were explored in order to incorporate not only network-related functionality, but also data-processing primitives.
The first phase of the project was concluded at the end of 2020. Discussions are currently underway regarding how to take this work forwards.
- V. Khristenko, M. Girone, M. Nuessle, D Frey, U. Bruening, J. Schonbohm, HIOS: Heterogenous I/O for Scale. Published on ATTRACT, 2020. cern.ch/go/qH7N