Roughly a year and a half after its launch, the SAGE exascale storage project led by Seagate has delivered a substantive interim report – Data Storage for Extreme Scale. It outlines technical details of progress to date and architectural plans moving forward. Of particular note is progress on co-design for use cases and applications expected to benefit most from exascale. There’s also been a fair amount of work to be able to accommodate big data and traditional HPC workflows in the same environment.
“We’ve tried to give ourselves lofty goals,” said Malcolm Muggeridge, senior engineering director at Seagate based in the U.K. who is leading the initiative. “We would like to become the platform of choice in exascale for storage solutions and will have the technology addressing that space in the 2022 timeframe. The main piece of work that has been completed [so far] is co-design activities.”
You may recall that SAGE (StorAGe for Exascale Data Centric Computing (SAGE) system aims to implement a Big Data/Extreme Computing (BDEC) and High Performance Data Analytics (HPDA) capable infrastructure suitable for Extreme scales – including Exascale and beyond. SAGE is one of 15 projects recently funded under Horizon 2020. Direct funding is actually through the European Technology Platforms (ETP) organization – “industry-led stakeholder groups recognized by the European Commission as key actors in driving innovation, knowledge transfer and European competitiveness. ETPs develop research and innovation agendas and roadmaps for action at EU and national level to be supported by both private and public funding.”
The new white paper is a fairly extensive document that follows a nine-month formal project review last June and includes work completed since. Among the topics covered are: platform requirements; systems architecture; platform components; and ecosystem elements. Launched in September of 2015, SAGE tackles eight research areas: “the study of the 1) application use cases co-designing solutions to address 2) Percipient Storage Methods, 3) Advanced Object Storage, and 4) tools for I/O optimization, supporting 5) next generation storage media and developing a supporting ecosystem of 6) Extreme Data Management, 7) Programming techniques and 8) Extreme Data Analysis tools.”
According to the report, the SAGE storage system will be capable of efficiently storing and retrieving immense volumes of data at extreme scales, with the added functionality of “percipience” or the ability to accept and perform user defined computations integral to the storage system. SAGE will be built around the Mero object storage software platform and its supporting ecosystem of tools and techniques, that will work together to provide the required functionalities and scaling desired by extreme scale workflows.
One important goal is accommodating new storage technologies, such as non-volatile RAM (NVRAM). Leveraging object storage to assist ‘in-memory, closer-to-memory” computing is another. In an earlier interview Sai Narasimhamurthy, Seagate research staff engineer responsible for coordinating the technical work, told HPCwire that the stack would “have memory at the top, various NVRAM technologies in the middle, of course you have your flash technology as well as part of the stack, and then you have scratch disks and then archival disks.”
“You could have an object, or a piece of it, lying in high speed memory, a piece of it in NVRAM, and a piece of the object lying in scratch based upon the usage profile of the object,” explained Narasimhamurthy. “The view of the object is transparent to the application, it’s just I0 to an object, but on the back end you could have various types of layout which could be very interesting because you could optimize your layout for performance or for resiliency. You could do all sorts of things.”
Clearly there are big goals for the project. Co-design is a critical early element in defining functional requirements, emphasized Muggeridge, “We have carefully selected use cases that reflect these data-centric applications. The use cases provide specific inputs that are designed to fine tune/modify the framework for the SAGE architecture.”
Muggeridge noted there is range of requirements drivers. The report calls out: inputs from the BDEC community and the US Department of Energy labs; data needs for big science, as exemplified by the Square Kilometer Array and the Human Brain Project; and Extreme scale I/O requirements drafted by the ETP; and extreme scale data needs highlighted by the HPDA community. The information was gathered mostly through workshops.
Top-level objectives have also been established and are largely familiar. One calls for the ability “to store and retrieve extreme volumes of data approaching orders of ~Exabyte for a given problem”. Another is the ability to manage workflows that include data from simulations and instruments. Not surprisingly, data IO rates, data integrity, data analytics, among other capabilities are being targeted. Indeed the first part of the project has been largely ‘definitional’ with a roll out of demonstrations planned for the next year.
Use of co-design principles to inform these objectives is a distinguishing feature of the project. SAGE has selected several use cases (applications) and spelled out in detail the parameters being measured. Use cases “cover a broad range of domains, including data from some of the world’s largest scientific experiments (including one of the world’s largest nuclear fusion facilities and one of the largest synchrotrons in Europe), aside from extremely data-centric HPC codes.” Below is a table with the uses cases selected.
So far, SAGE has gathered the first formal list of inputs from all of the specified use cases. “This phase included gathering inputs on formal I/O characterization, SAGE architecture analysis, data retention characterization and data scaling analysis, which was an analytical study of how data and I/O requirements of the use cases would scale on a future basis.”
The SAGE system is built on multiple tiers of storage device hardware technology (see figure below). SAGE does not require a specific type of storage device technology, but typically it would include at least one NVRAM tier (Intel 3DxPoint technology is a strong contender at the moment), at least one flash tier and at least one disk tier. Together, these tiers are housed in standard form-factor enclosures and provide their own compute capability, enabled by standard x86 embedded processing components. Moving up the system stack, compute capability increases for faster, lower latency devices.
Mero, the object storage software first developed by Xyratex and now being extended by Seagate, is layered on top of this hardware stack, providing fundamental management of object I/O and storage across tiers. Essentially, Mero forms the core of the SAGE system. Mero is presented to users through the Clovis API. Everything above Clovis forms the SAGE ecosystem components.
Much remains to be done but it seems as if SAGE is making steady progress. Demonstrations, some at the Julich Supercomputing Centre, are expected over the next year or so. This latest paper is best read in full for current technical details of SAGE plans.
Link to new SAGE paper (Data Storage for Extreme Scale): http://sagestorage.eu/sites/default/files/Sage%20White%20Paper%20v1.0.pdf