I compared three workload configurations on a private cloud and on a public cloud over five years, using heavy workloads.
The detailed report was supposed to be linked from this post. It never was, so the figures are not on this page.
What the comparison showed:
Archive note, December 2017. Newer writing here is on VMware Cloud Foundation, Avi, and NSX.
Kubernetes manages containerized applications. These were other tools people reached for at the time.
Native clustering for Docker. It turns a pool of Docker hosts into one virtual Docker host. Because Swarm serves the standard Docker API, a tool that already talks to a Docker daemon can use Swarm to scale across hosts. Tools that fit that pattern included the Docker client, Dokku, Docker Compose, Docker Machine, and Jenkins. Docker Swarm overview
A platform for fine-grained resource sharing in the data center. Introduction
Manages a cluster of machines and runs applications on them. You declare what you want to run. Nomad decides where and how. Nomad introduction
A framework for orchestrating containers. Marathon introduction
Written in December 2017, and left as a record of that moment. Newer writing here is on VMware Cloud Foundation, Avi, and NSX.
A functioning cluster needed six components:
Each one could run as a normal Linux process, or inside a container.
The master ran the API server, the scheduler, and the controller manager. It could be set up as multi-master. The scheduler and the controller manager elect a leader. The API servers can sit behind a load balancer.
The API server exposes a REST interface to Kubernetes resources and is configurable. The scheduler places containers on nodes using policies, metrics, and resource requests, and it can be tuned with command-line flags. The controller manager reconciles the cluster’s actual state with the desired state from the API. It is a control loop, and it is configurable too.
Every worker ran kubelet, kube-proxy, and the Docker engine. kubelet talked to Docker and kept the containers that should be running actually running. kube-proxy handled network connectivity to those containers.
The post also pointed at rkt as an alternative to the Docker engine: rkt getting started guide.
The diagram was from a Kubernetes training slide by @wattsteve.
Written in December 2017. Newer writing here is on VMware Cloud Foundation, Avi, and NSX.
One processor with two or more independent cores. The cores read and execute ordinary CPU instructions (add, move, branch) at the same time, which helps programs that can run in parallel.
Examples from the original note: AMD A-Series, Athlon II, FX-Series, Phenom, and EPYC; IBM POWER4, POWER5, and POWER6; Intel Core 2 Duo, Core i3, Core i5, and Core i7.
A specialist multicore design for a high degree of parallel work, with a large number of simpler cores (tens, hundreds, or thousands). Used in embedded systems and high-performance computing.
Examples from the original note: the Sunway TaihuLight supercomputer, GPUs, Intel Xeon Phi, Tilera, and the Teraflops Research Chip.
Archive note, October 2017. Newer writing here is on VMware Cloud Foundation, Avi, and NSX.
A CPU is a general-purpose processor. It can in principle do any computation, but not always in the best way for that computation. You can do graphics on a CPU. A GPU built for the job will usually finish sooner.
A GPU is a special-purpose processor, tuned for the calculations computer graphics repeat, especially SIMD work.
Many actions happen at once on many pieces of data. Adding and multiplying at the same time to solve a problem with separate parts is the example used in the original note. The work may or may not be synchronized. MIMD is used when an algorithm splits into independent parts and each part goes to a different processor.
One identical action runs at the same time on many pieces of data: retrieve, calculate, or store. Fetching multiple files at once is the example used in the original note. Processors with their own local data execute the same instruction in step, and they communicate to shift work around. SIMD is used when a lot of computation is the same operation done in parallel.
Archive note, October 2017. Newer writing here is on VMware Cloud Foundation, Avi, and NSX.
