Nutanix Acropolis - облако на базе KVM под ключ, Максим Шапошников (Nutanix)

Editor's Notes

1. Purpose of the Slide: Talk about the big challenges enterprise customers face today. Establish a baseline that everyone can agree to. Key Points: - Datacenters have become increasingly complex over the years. Every part of the infrastructure lifecycle is complex, from buying and deploying to configuring, managing and scaling infrastructure. - As infrastructure became more complex, IT organizations also became more siloed. You needed storage experts to manage complex network storage and networking experts to manage enterprise network topologies. ITIL processes emerged to deal with the complexity. All this significantly slowed down the pace of IT deployment. Orgs have to trade off doing it right with doing it fast. - As demand for resources (compute, network, storage) goes up, organizations want to be able to add capacity incrementally and predictably. Scale-up (big iron) infrastructure makes it difficult to scale in small increments when needed.
2. Purpose: Talk about where and how web-scale IT originated, and what some of the common are between different web-scale data centers Key Points: Web companies like Google and Facebook started pushing the limits of existing infrastructure systems and processes in ways that traditional businesses did not. They needed infrastructure that could support their business requirements (rapid application development cycles, scale on demand, cost containment). They tried using existing infrastructure solutions, but quickly realized that legacy infra was a poor fit for their needs. Over time, these companies developed an alternate approach to IT that enabled them to get past limitations in infrastructure. Some common traits of web-scale IT: Infrastructure built from commodity server hardware pooled together using intelligent software. This allows customers to start small and scale one server at a time – true scale-out The software in the system is distributed across all the nodes. You don’t have central metadata servers or name nodes. You don’t see controller bottlenecks Embarrassingly parallel operations – everything in the system, including storage functions like deduplication and metadata management and system cleanup, is distributed across all nodes. There are no hotspots or bottlenecks, allowing for massive scale Compute and storage sit very close to each other. Data does not have to go back and forth between storage and compute over a network. Data has gravity, so co-locating storage and compute eliminates network bottlenecks and system slowdown Heavy automation eliminates the need for expensive, error-prone manual operations. You don’t
3. Hyper-converged on x86 servers Compute and storage integrated in a single server with no network in between Eliminate network bottlenecks, All intelligence in software No reliance on hardware for data services, performance, resilience, etc. Distributed everything Data is intelligently distributed in all the nodes in the cluster so that no single node becomes a data access hotspot Services are distributed across the cluster so that as new nodes are added the processing capacity for data services scales out No single point of failure Redundancy in data (through replication, erasure coding, etc.) allows for component failure without loss of data Redundancy of controllers and control path enables recovery from failure of individual controllers API-driven automation and rich analytics - Programmatic interface and data-driven decision making in the system
4. 15