What is Pod? What is a cluster?

What we do on the Internet (that is, all our activities today) cannot be separated from Pod (that is, cluster).When we watch popular TV programs online, order jeans or chat with grandma through Zoom, we become part of this cluster. At this moment, you can read this article thanks to Pod.So,

what is Pod? What is a cluster?

Pod or cluster, in fact, is a group of computers linked by high-speed network and become a whole.Computer architects must have borrowed (at least unconsciously) this natural term. Like modern computer clusters, pea pods and the Super Dolphin Group show the energy that many individuals have when they unite as a whole.The Root of pea pods and Super DolphinsThe connection is deeper. Botanists point out that pea pods can not only protect and nourish every bean, but also redistribute the resources of damaged seeds to the thriving seeds. Similarly, the load balancer will move the job from the failed compute node to the working node.This dynamic relationship is similar to that of dolphins.Denise Herzing, an experienced marine biologist, often sees the same groups of dolphins (a family group of about 20 dolphins) every day when she works on the Bahamas coast. Once, she saw a large number of dolphins."Many years ago, outside the Baja Peninsula, I saw a super dolphin group. This is very exciting and a bit overwhelming, because as a researcher, I just want to observe a small group of animals carefully, not a thousand animals scattered in a large area. " The scientist who founded the wild dolphin project said.Super dolphins are very important to dolphins. She pointed out that "equipped with 1000 pairs of ears, the super dolphins will form a huge sensing system, listening to the movements of predators like a super sensor, thus protecting the individual dolphins that travel back and forth." In addition, she said, this is similar to the cluster used in today's cloud computing services.Data centers with multiple clusters or PODs are distributed in multiple buildings and run as a single system.Pod prototype of early data centerGreg Pfister, a former IBM technician and cluster expert, said that when many companies computerized their accounting systems in the early 1960s, they naturally connected multiple computers together so that they could have spare computer equipment in case of a computer failure."I'm pretty sure NCR, MetLife and many others have done this." Pfister said. He is the author of In Search of Clusters, which is regarded by some as the "Bible" in the field of clusters.In May 1983, Digital Equipment Corp integrated several popular 32-

bit VAX minicomputers into a system called VAXcluster. Each computer has its own system running independently, but they share other resources and can provide a single system image for IT users.

In the late 1990s, the emergence of low-cost

PC processors, Ethernet and Linux gave birth to at least eight major research projects aimed at building clusters. NASA designed a cluster (consisting of 16 PC motherboards connected by two 10 Mbs networks) and named it Beowulf, imagining that it would eliminate the supercomputers and massively parallel systems at that time.Cluster networks need speed.Researchers have found that as long as high-speed

networks are used to eliminate bottlenecks, clusters can be built quickly and high performance can be provided at low cost.Another project in the late 1990s was Berkeley's Workstation Network (NoW), which connected dozens of Sparc workstations through the fastest interconnection technology at that time. They also created a picture of a group of small fish eating a big fish to show their achievements.The researchers behind the Berkeley NoW project speculate that many clusters of small systems will outperform a single large computer.Eric Brewer, one of the researchers, thinks that clustering is very suitable for emerging Internet applications, so he uses the NoW system consisting of 100 servers as a search engine."Once upon a time, we were running the best search engine in the world in Berkeley Park." David Patterson, a veteran of many computer research projects in NoW and Berkeley, said.The work was very successful–Inktomi, co-founded by Brewer, was an early search engine based on a cluster inspired by NoW, which contained 1000 systems. Inktomi has many competitors, including a startup called Google, which is rooted in Stanford University.Patterson, the co-author

of a popular computing textbook, said: "They used PCs to build network clusters, and developed a business model that is conducive to their own development and really improve the search quality–everyone

knows the rest."Today, cluster or Pod is the foundation of most global top 500 supercomputers and almost all cloud computing services. Most of them use NVIDIA GPU, but we are getting better and better.Pod and Cluster: Debate on TermsAlthough computer architects refer to these systems as clusters, some network experts prefer the word Pod. They converted this biological term into a technical abbreviation, pointing out that POD stands for the "delivery point" of computing services.The word Pod was very popular in the early stage of cloud computing. Service providers are competing to build a larger system of warehouse scale. They usually order the pre-configured

system of the whole "shipping container" (also known as Pod) and can plug them together like Lego bricks.

Recently, the Kubernetes group adopted the word Pod. They define the software Pod as "a container or a few containers that are closely coupled and share resources".The word Pod has also been adopted in aerospace and consumer electronics industries, which may be to bring the warmth of organisms to their concepts. The most iconic examples include the iPod (the predecessor of the iPhone) and the one-man

spaceship in the movie 2001: A Space Odyssey.When AI meets the clusterIn 2012, cloud computing service providers ushered in the AI Big Bang, and a powerful new computing form was born. They are competing to build large GPU clusters, which can handle huge data sets to train and run neural networks with the help of their internal accelerator core clusters.In order to help promote AI to various enterprise data centers, NVIDIA integrated the GPU cluster on InfiniBand network into NVIDIA DGX system. The reference architecture allows users to easily expand from a single DGX system to NVIDIA DGX POD or even supercomputer-level NVIDIA DGX SuperPOD.For example, the Cambridge-1 AI supercomputer in Britain is built on DGX SuperPOD, which is dedicated to promoting the development of life science and medical health. It is one of many AI-ready

clusters and PODs promoted in an unprecedented way. Just like AI itself, they are developing rapidly in all walks of life in various forms and scales.