Skip to main content

Percipient Storage for Exascale Data Centric Computing2

Periodic Reporting for period 1 - Sage2 (Percipient Storage for Exascale Data Centric Computing2)

Reporting period: 2018-09-01 to 2020-02-29

The landscape for High Performance Computing (HPC) is changing with the proliferation of enormous volumes of data created by scientific instruments and sensors, in addition to data from simulations. This data needs to be stored, processed and analysed, and existing storage system technologies in the realm of extreme computing need to be adapted to achieve reasonable efficiency for faster scientific innovation.

Sage2 (Percipient Storage for Exascale Data Centric Computing 2) hence intends to develop new storage technologies that help to manage the immense data and storage needs of a wide range of societally important use cases in the realms of weather & climate, bio-informatics, fundamental research in particle physics and astronomy, etc, and the needs of organisations that intend to exploit the power of data for AI and Deep Learning algorithms, for better predictive insights that will affect almost every aspect of human life.

The SAGE paradigm already provides a basic framework to address the extreme scale data aspects of High Performance Computing. Sage2 intends to validate a next generation storage system building on top of the already existing SAGE platform to address new storage requirements in the above mentioned areas.

The following are the new objectives extending the work of the original SAGE proposal:

1. Usage of compute node local storage resources: Sage2 will build a storage software system called Mero that can demonstrate the usage of special “byte addressable” Non Volatile Memories, that are mostly starting to be used within the computing systems of high end supercomputers. Mero reads and writes data as “objects” rather than files and is hence an “object storage” system.

2. Usage of new Non-Volatile Memories in a deep I/O stack: Sage2 intends to use novel Non- Volatile Memory technologies along side other types of storage technologies such as hard disks and flash based solid state memories – all part of a single storage system. With this, we can get very high performance for the applications with data being available at the right place for use at the right time.

3. Strengthening the ARM based ecosystem: Sage2 intends to include new processing technologies based on the ARM processor as part of its storage system.

4. Memory style access to storage: Sage2 intends to explore the usage of storage just like memory. Storage is so called “Block” addressable, but memory is accessed in “Bytes”. Providing flexibility for the use of storage like memory will greatly aid AI and Deep Learning applications.

5. Co-design with a new class of use cases: The Sage2 system intends to gather the requirements of new classes of extremely challenging and critically important and evolving applications such as Artificial Intelligence and Deep learning alongside scientific simulations that produce and consume a lot of data.

6. Building data management tools and programming capabilities: The Sage2 system will provide new data management tools & programming capabilities for applications, that can efficiently exploit the new storage system.
Availability of the basic SAGE system that can be used by external users – that is currently installed at Juelich Supercomputing Center in Germany. This has now been enhanced with new ARM based hardware and additional test capabilities (more “clients” to test applications).

Availability of data and storage requirements from chosen applications in the fields of AI and Deep Learning (including Cervical Cancer Diagnostics), Space Weather, Brain Image analysis, Particle Physics (Nuclear Fusion) and astronomy.

First design of the SAGE system based on all the requirements gathered, including the design of the Mero software and the software components that work “on top” of Mero to enable better use of the SAGE system by the applications, which forms the Mero ecosystem.

Designs of the “Global Memory Abstraction” techniques that help to work with Non Volatile Memories and enable the use of the storage system like Memory. We have also conducted a deep study of the methods needed to use Non Volatile Memories in ARM based environments. First implementations of the Global Memory Abstraction mechanisms are also available. We have also contributed to the linux open source community in this area.

We also have a head start in the dissemination of the project ideas and early results and have had formal collaborations initiated with other complimentary projects (a Memorandum of Understanding has been signed with the Maestro project and the EpigramHS project). We are working on a path towards open sourcing Mero with early adopters within Sage2 and other European R&D projects.
Sage2 progresses the state of the art in:

(a) smart usage of multiple tiers of storage including Non Volatile Memories through object storage software, as opposed to using parallel file system technologies typically used in such settings. The object storage software will have advanced capabilities such as in-storage computation and Quality of Service.

(b) usage of Non Volatile Memories in ARM based platforms, and in the usage of ARM based processing for object storage systems within HPC.

(c) usage of block based data storage technologies as byte addressable entities through the mechanisms of Global Memory Abstraction hence blurring the line between memory and storage.

(d) In the areas of applications (both classical HPC and AI/DL), in enabling the usage of novel features of SAGE such as object storage, its associated API, and advanced features such as in-storage computation.

(e) AI tools such as TensorFlow, visualization tools in HPC, HPC schedulers such as SLURM, distributed storage technologies such as dCache – all used with object storage systems. We also develop novel software tools such as high speed object transfer interfaces and I/O containers for working better with object stores.

Sage2 hence aims to achieve the impacts in the area of High Performance Extreme Scale/Exascale Computing, and, AI by providing novel methods of data storage, processing and retrieval helping scientists and innovators to better utilize High Performance Computing for scientific and industrial innovation. Sage2 also aims to promote the usage of the new technologies developed within the project for wider adoption by the HPC and the enterprise computing community, with plans being considered to fully open source Mero, which will be finalised before the project ends. Sage2 will also promote the further usage of low power processing technology, ARM in High Performance Computing environments – especially within data storage. Sage2 will also influence the EuroHPC roadmaps and provide inputs to future Strategic Research and Innovation agendas that drive the same through Sage2 partners that are actively part of organisations such as ETP4HPC. Sage2 will also impact the High Energy Physics community, medical communities (Cancer diagnostics, medical Imaging, etc), radio astronomy, and communities that increasingly rely on supercomputers to run AI and Deep Learning tasks. Sage2 enables commercial partners, Seagate, ATOS and Kitware to achieve impacts in their markets – namely data storage, computing systems, and visualition and performance tools respectively.
SAGE Architecture in Sage2 Project
Prototype at Juelich Supercomputing Centre