Today's data centers are tasked with processing colossal amounts of data every day. Are they equipped to keep up? In this video, Fungible founders Pradeep Sindhu and Bertrand Serlet discuss how data centers have changed over the years and what next-generation data center architecture should look like.

To learn more about Fungible, visit: [ Ссылка ]

Follow us on Twitter: [ Ссылка ]
Follow us on LinkedIn: [ Ссылка ]

#datacenter #technology #cloud #network #data #datacentric #datacenterarchitecture #datacentricsw #softwarearchitecture #fungible_inc


The data flood began in the 1940s with the realization that all information could be represented digitally stored indefinitely transmitted over long distances and manipulated by general-purpose programmable computers. Today we understand clearly the path from raw data to insights and from insights to economic value.
The advent of machine learning has accelerated the explosion of big data. Machine learning benefited greatly from the existence of large stores of data, but once machine learning became successful in particular application areas it gave people the incentive to gather even more data. It is not an exaggeration to say that most applications today are data-centric. Unfortunately, the infrastructure for information processing is still largely compute-centric. In fact,– in fact, the CPU sits at the heart of most if not all server building blocks.
As Moore's Law slowed down in the early 2000s, infrastructure transformed to using servers connected by a network in a so-called, "scale-out configuration." As Moore's Law slowed even further, people started to connect specialized engines like GPUs to CPUs inside of servers to increase the efficiency of applications like machine learning. Today with Moore's Law nearly flatlining, this compute-centric approach is no longer viable.
We need a data-centric approach to infrastructure. In today's data-centric world, we need a new programmable engine we call the data processing unit, or DPU. This data engine, is a new category of microprocessors and it provides two fundamentally new capabilities. The first one is that it can perform data-centric computations at least an order of magnitude more efficiently than general-purpose CPUs. Data-centric actually means four things: first, the workload arrives in the form of packets. Second, that these packets belong to many, many computations. Third, the ratio of IO to arithmetic is high, and fourth, that the computations are stateful. The second fundamental capability that the DPU provides is the ability to connect together into a single fabric large numbers of storage and compute servers in such a way that their resources can be fully pooled across the entire data center. So Moore's Law has been weakening for silicon, but of course people have more and more data everyday and they want not only to process that data, but to get deeper insights on that data and to get them in real-time. That data and to get them in real-time.
The classical architecture for the software is inadequate – people have been doing the software the same way for fifty years. The way it works, is that you have some tasks to do, some computations, and you break it down into little pieces, and smaller pieces, all the way down to instructions of your processor and then you pump data one piece at a time through that computation tree. This is a very sequential process and people have been trying for ways to improve on that down into little pieces, and smaller pieces, all the way down to instructions of your processor and then you pump data one piece at a time through that computation tree. This is a very sequential process and people have been trying for ways to improve on that and they've gone via multi-thread programming to more parallel architecture, the software is a mess it's very buggy – it's very difficult to do this. So, there's been a modern way to do this – and that's based on microservices. You cut your task into horizontal subtasks that are implemented by microservices, and then you instantiate your microservice on multiple nodes, so you get the scalability, the "scale-out" scalability that way. Then, you stitch together a number of microservices to get the computation that you want done. This architecture is what we call data-centric. To cope with the data demands, we believe all next-gen data centers need to be data-centric. So, Pradeep and I created Fungible, to bring together data-centric hardware and data-centric software to power next-gen data centers.

свернуть