The Machine and Exaflops for the Big Data era
In May, HPE demonstrated an important stage of its research program The Machine, a prototype of a computer with 160 TB of shared memory. The program is aimed at developing a new memory-oriented computing architecture, rather than a processor (memory-driven computing). And in June, it was announced that, on the basis of this architecture, a reference model of an exaflop supercomputer for the US Department of Energy would be created. We tell in more detail about these revolutionary news.
Hewlett Packard Enterprise received a research grant from the US Department of Energy to create a reference model of an exaflop supercomputer that will allow creating unattainable mathematical models and simulations for use in science, medicine, design and other technologies.
To achieve exaflops performance by 2022–2023. It is necessary to increase the speed, energy efficiency and density of high-performance computing systems by 10 times compared with the fastest modern supercomputers. To implement low latency exaflops computations, the reference model created by HPE will have to eliminate these problems and remove the limitations on memory size, memory scalability and bandwidth inherent in today's high-performance computing architecture.
')
The HPE reference design model is based on the concept of Memory-Driven Computing (memory-oriented computing). This architecture is a computing platform whose central element is memory, and not a processor, which makes it possible to obtain previously unavailable gains in performance and efficiency. HPE's Memory-Driven Computing architecture is a large-scale set of technologies that are being developed by Hewlett Packard Labs as part of The Machine research project. On May 16, 2017, HPE introduced a new version of the prototype created during the implementation of this project and which became the world's largest computer with shared memory.
The fundamental technologies underlying the developed reference architecture of the exaflops computing supercomputer include a new memory factory and data transfer using low-energy photonics. The memory factory is the ideal technological basis for a wide range of high-performance computing and tasks aimed at processing a significant amount of data, including big data and analytics. HPE also continues to explore various options for non-volatile memory that can be connected to a memory factory, significantly increasing the reliability and efficiency of exaflop systems.
More details about the premises of The Machine project, innovative technologies related to the development of the project (memristors, non-volatile memory, photonics, system on chip), the logical and engineering structure of The Machine can be found in the report by Alexander Starygin, director of the HPE technical solutions development department in Russia, at the HPE Digitize conference held on July 5 in Moscow.
HPE collaborates with technology partners to develop world-class open architectures based on open industry standards. The central element of this approach is the use of the innovative Gen-Z protocol , which determines the interaction of integrated circuits with each other.
Gen-Z protocol provides data exchange between integrated circuits on the basis of memory semantics, which allows to organize reliable interaction of many devices, including central and graphics processors, programmable logic integrated circuits (FPGA), volatile and non-volatile memory, intersystem elements and many other devices; however, they all use a single address space. This makes it possible to create memory-oriented systems to dramatically increase application performance and energy efficiency. Gen-Z protocol allows partners to expand cooperation in the development of supercomputers, which use the most advanced technologies, developing in an open and competitive ecosystem.
The memory-oriented computational architecture makes it possible to remove the problem inherent in the traditional architecture: the inefficient interaction of memory subsystems, data storage systems and processors. Due to this, the time to perform complex tasks is radically reduced: from several days to several hours, from hours to minutes, and from minutes to seconds, allowing you to get a meaningful result in real time.
For example, the computational task — a Monte Carlo simulation, which is often used by banks and traders to predict the development of stock markets or currency exchange rates, takes almost two hours to complete at the current modern computing complex. The first tests show that with the use of a memory-oriented architecture, this time is reduced to just over a second, on the same amount of data and with the same complexity, which gives acceleration several thousand times.
Not long ago, the German Center for Neurodegenerative Diseases (DZNE) signed a cooperation agreement with HPE to apply the new computing architecture in its scientific and medical research. DZNE generates huge amounts of data, for example, obtained during magnetic resonance imaging (MRI) of the brain or information about the genome. In-depth analysis of these data often takes one to two weeks, and only after that doctors can determine the next step of treatment and / or start a new cycle of tests / studies. The use of a new memory-oriented computing architecture will not only speed up the analysis of the data obtained, but also use images of much higher accuracy.
At the HPE Discover 2017 conference in Las Vegas, the company announced the creation of The Machine User Group - an open community of developers interested in programming for the memory-driven computing environment, using a massive pool of non-volatile memory and specialized processors dedicated to a specific task (SoC). A set of tools is already available for developers on GitHub , and in the future, User Group members will have access to The Machine emulation environment.
Exaflop supercomputer for the US Department of Energy
Hewlett Packard Enterprise received a research grant from the US Department of Energy to create a reference model of an exaflop supercomputer that will allow creating unattainable mathematical models and simulations for use in science, medicine, design and other technologies.
To achieve exaflops performance by 2022–2023. It is necessary to increase the speed, energy efficiency and density of high-performance computing systems by 10 times compared with the fastest modern supercomputers. To implement low latency exaflops computations, the reference model created by HPE will have to eliminate these problems and remove the limitations on memory size, memory scalability and bandwidth inherent in today's high-performance computing architecture.
')
The HPE reference design model is based on the concept of Memory-Driven Computing (memory-oriented computing). This architecture is a computing platform whose central element is memory, and not a processor, which makes it possible to obtain previously unavailable gains in performance and efficiency. HPE's Memory-Driven Computing architecture is a large-scale set of technologies that are being developed by Hewlett Packard Labs as part of The Machine research project. On May 16, 2017, HPE introduced a new version of the prototype created during the implementation of this project and which became the world's largest computer with shared memory.
The fundamental technologies underlying the developed reference architecture of the exaflops computing supercomputer include a new memory factory and data transfer using low-energy photonics. The memory factory is the ideal technological basis for a wide range of high-performance computing and tasks aimed at processing a significant amount of data, including big data and analytics. HPE also continues to explore various options for non-volatile memory that can be connected to a memory factory, significantly increasing the reliability and efficiency of exaflop systems.
More details about the premises of The Machine project, innovative technologies related to the development of the project (memristors, non-volatile memory, photonics, system on chip), the logical and engineering structure of The Machine can be found in the report by Alexander Starygin, director of the HPE technical solutions development department in Russia, at the HPE Digitize conference held on July 5 in Moscow.
Gen-Z protocol
HPE collaborates with technology partners to develop world-class open architectures based on open industry standards. The central element of this approach is the use of the innovative Gen-Z protocol , which determines the interaction of integrated circuits with each other.
Gen-Z protocol provides data exchange between integrated circuits on the basis of memory semantics, which allows to organize reliable interaction of many devices, including central and graphics processors, programmable logic integrated circuits (FPGA), volatile and non-volatile memory, intersystem elements and many other devices; however, they all use a single address space. This makes it possible to create memory-oriented systems to dramatically increase application performance and energy efficiency. Gen-Z protocol allows partners to expand cooperation in the development of supercomputers, which use the most advanced technologies, developing in an open and competitive ecosystem.
Practical application of memory-driven computing
The memory-oriented computational architecture makes it possible to remove the problem inherent in the traditional architecture: the inefficient interaction of memory subsystems, data storage systems and processors. Due to this, the time to perform complex tasks is radically reduced: from several days to several hours, from hours to minutes, and from minutes to seconds, allowing you to get a meaningful result in real time.
For example, the computational task — a Monte Carlo simulation, which is often used by banks and traders to predict the development of stock markets or currency exchange rates, takes almost two hours to complete at the current modern computing complex. The first tests show that with the use of a memory-oriented architecture, this time is reduced to just over a second, on the same amount of data and with the same complexity, which gives acceleration several thousand times.
Not long ago, the German Center for Neurodegenerative Diseases (DZNE) signed a cooperation agreement with HPE to apply the new computing architecture in its scientific and medical research. DZNE generates huge amounts of data, for example, obtained during magnetic resonance imaging (MRI) of the brain or information about the genome. In-depth analysis of these data often takes one to two weeks, and only after that doctors can determine the next step of treatment and / or start a new cycle of tests / studies. The use of a new memory-oriented computing architecture will not only speed up the analysis of the data obtained, but also use images of much higher accuracy.
The Machine User Group
At the HPE Discover 2017 conference in Las Vegas, the company announced the creation of The Machine User Group - an open community of developers interested in programming for the memory-driven computing environment, using a massive pool of non-volatile memory and specialized processors dedicated to a specific task (SoC). A set of tools is already available for developers on GitHub , and in the future, User Group members will have access to The Machine emulation environment.
Source: https://habr.com/ru/post/332732/
All Articles