INTRO TO CLOUD COMPUTING - First Half

What is multitenancy?

• One or more instances run as guests of a physical host system, these guests are called tenants

• For multitenancy that defines many tenants running on the same host

• Performance issues might happen (noisy neighbors)

How does hardware virtualization work

• Uses a hypervisor to create and manage VM instances to act as dedicated computers

What is a cloud API

Provided by cloud providers to provision resources automatically

What is an m5?

• General purpose instance type (the m is general purpose and the 5 is the generation)

What is c5, [i3, d2], [r4, x1], [p1, g3, f1]

• Compute optimized

• Storage optimized

• Memory optimized

  • Accelerated computing (GPUs)

Difference between vertical and horizontal scalability

• Horizontal: adds more machines to run a server faster

  • Vertical: Increasing the size of an instance

Describe instance storage and persistent storage

• AKA ephemeral drives

• Volatile and is destroyed with the instance

• For persistent you have three types:

  • File store - files over NFS (AWS = EFS)

  • Block store - such as blocks over iSCSI (AWS = EBS)

  • Object store - over an API - HTTP based (AWS = S3)

What is EC2

• Elastic compute cloud

  • Web service with a basic interface for launching app instances under several different OS

What is Kubernetes

It is kinda like Docker containers, but it automates those containers’ deployment, scaling, and general management

Reduces operational complexity

What are the 5 essential characteristics of cloud computing?

1) On demand self service

2) Broad network access

3) Resource pooling

4) Rapid Elasticity

5) Measured service

Describe the four cloud deployment models

• Private Cloud: Provisioned for exclusive use of a single org and can have multiple consumers that represent the business units

• Community cloud: Specific community of consumers that have shared concerns (i.e. compliance stuff). Could be owned by someone in there or a third party.

• Public cloud: For use by the general public - exists on premises of the cloud provider

  • Hybrid cloud: Two or more cloud infrastructures that remain unique entities

Define the following:

• Cloud Consumer

• Cloud Provider

• Cloud Auditor

• Cloud Broker

• Cloud Carrier

• Consumer - a person or org that maintains a business relationship with and uses services from a cloud provider

• Provider - person or org responsible for making a service available to interested parties

• Auditor - Third party to conduct independent assessments of cloud services

• Broker - manages the use, performance, and delivery of cloud services

  • Carrier - Intermediary that provides connectivity and transport of services from cloud providers to cloud consumers

Describe the User Responsibility graph

What is the typical communication pattern for a WSC?

• WSC is a Warehouse Scaled Computer

  • The typical pattern is to serialize the data to the protocol buffer, then execute an RPC (Remote Procedure Call) to pass the buffer to the callee, and then the caller de serializes the buffers received in response to the RPC

What is the difference between control flow and data flow

• Control flow is sequential, has program counters, loads instructions into register

  • Data flow - operations are carried out at the time when their input becomes available, BUT they are inefficient with cache (used by network routers)

What is Mesos

• Has master process that manages daemons running on all cluster nodes while other frameworks run the tasks on cluster nodes

• master dishes out fair sharing of resources

  • Has a scheduler which receives resource offers from the master and an executor on each machine to launch the tasks

What is a node cluster in the cloud?

A node is a machine and a cluster is a group of machines. Since the cloud is distributed you want to share that job across a bunch of machines for speed, so you spin up a cluster to do this.

Your cluster will have a master node and n worker nodes

What are the two sides of cluster management

• App developers - simple means to locate resources and then to control their use

• Service providers - system availability, reliability, and the resource utilization

What are the four types of cluster schedulers?

• Monolithic (Borg) - Big guy - uses priority and preemption, all resources

• Static Partition (Dryad) - Fixed subset of resources and is scheduler dependent policy wise

• Two Level (Mesos) - Dynamic resources, very strict fairness

• Shared-state (Omega) - All resources, also uses priority and preemption

What is the difference between Containers and Hardware VMs

1) A Hardware VM has its own full guest OS and has higher overhead

2) A container is more lightweight BUT you share your infrastructure. You still retain your little space though

Talk about the relationship between kernel and containers/VMs

A full VM has its own kernel so you can run custom modules and all that

A container has a shared kernel so if you wanted to extend the functionality of the kernel (device drivers etc) without restarting then you cannot do that

It’s also hard since you cannot analyze the kernel for performance stuff since it is shared, and the shared makes it less secure

What is the relationship between docker and Kubernetes

Docker is easy to use - think developers

Kubernetes is by Google and is performance based

What is weird about linux and containers

Linux dgaf about what a container is, BUT you can use user-space software (i.e. Docker) to create containers.

Even if you have four containers and all of them have some process with ID of 1, then they are actually different.

Windows knows how to do this stuff, or at least has it more down

Why is Kubernetes good

Multiple containers can be created to share the same namespaces as part of a Pod

Those containers in a Pod can communicate quickly

Namespace filters that view down so that containers really only care about that smaller area

What is Amazon’s Container Service Called?

ECS (Hosts regular EC2 instances)

Describe the Kubernetes Architecture

Master - makes the global decisions about scaling and all that

API server - users and tools talk to this as like a front door

etcd - key value store that holds your cluster state

Scheduler - helps put a pod onto a certain node

Controller Manager - makes sure the desired state is actually the state (if a pod crashes then spin up a new one)

What is a Kubernetes Pod?

Groups of containers scheduled on the same host

Units for scheduling and deployment and share fate and resources such as storage volumes

This helps you avoid running multiple applications like you might in a single Docker container

Horizontal would be changing the number of nodes, and vertical would be resizing the nodes

A hardware instance is also called an ______

Infrastructure as a Service

In cloud computing, guest/tenant is another name to _________

Instance

What is a hybrid cloud?

A combo of public and private clouds

What is orchestration

Also Kubernetes. Software that manages application deployment using containers

Which lowest layer is considered user responsibility in the service model SaaS, PaaS, and IaaS

S: User Specific Application Configuration

P: Application

IaaS: Operating System

What are the four deployment models in cloud and fog computing?

Private: A cloud or fog node used by a single organization

Community: Exclusive use for a specific community of consumers from organizations that have shared concerns

Public: For open use

Hybrid: Two or more of those

What is the difference between a cloud federation and a cloud provider?

Federation is essentially an environment where users in Org A can discover and invoke services in Org B

Service providers in Org B can validate credentials from Org A and make the proper access decisions

What are the two additional characteristics associated with fog computing but not with cloud computing

Predominance of wireless access

Support for mobility

List the three flavors of SIMD architectures discussed

Vector architectures

SIMD extensions for mobile and multimedia

GPUs

GPUs runs with CPUs. How are they treated? Does the OS run on a GPU?

CPU acts as a host connected with a GPU (device)

OS does not run on a GPU

What is a TPU?

Tensor Processing Unit

Not a multithreaded SIMD device and is specific for deep learning

What is SoC and what is the purpose?

System on a Chip and it is used for maximizing power efficiency

One of the desired properties of a distributed system is failure transparency. Give it a definition or description.

The concealment of faults

Software modularity may require information hiding. What is that?

The user of a module doesn’t need to know anything about the internal mechanism of the module to use it effectively

What is the difference between a client server system and a P2P system?

P2P is decentralized, dynamic, and unstructured

Supercomputers these days are more likely based on ____

A very large number of processors and cores

What is a computing grid and what is it to do with cloud computing?

A distributed system of a large number of geographically dispersed systems like an electric grid

What common procedures cause most of the WSC “tax”

RPC, serialization, deserialization, and compression

Give an example of an application that requires coarse grained parallelism

A passport application, since that unit of work is much large to process each form

Give an example of an application that requires fine grained parallelism

Matrix Multiplication since each computation is small

Dryad improves on MapReduce with ______ architecture

Data-Flow

Mesos handles fine grained parallelism with a Zookeeper that selects a ______

Master - that’s the Zookeeper’s job

In Borg, a task may be in one of three states: pending, running, and dead. What action transits the state from running to pending? What action turns dead to pending?

When you evict from running and submit from the dead

Which scheduler is considered as pessimistic in conflict resolution?

Mesos

used for fine grained for the most part

Pessimistic means it assumes conflicts will occur if multiple things compete for the same resources so it waits for a decision before offering them elsewhere

What QoS’s does Quasar try to guarantee?

Response time and execution time

Quasar is like smarter mesos almost

Clock rates of processor cores can be changed with DVFS, which stands for _________

Dynamic voltage and frequency scaling

In Heracles, memory bandwidth saturation may result in cores being removed from ____ tasks

BE

BE are low priority tasks

What are three revolutions named in the book?

Cloud Computing, DevOps, and containers

What is the devops movement

Bring software dev skills to operations

What is the difference between OS Virtualization and Hardware Virtualization

OS partitions the OS into instances (containers)

It differs from hardware because only one kernel is running

What kind of virtualization does WSL embody

Hardware

What is a pod in kubernetes

Group of containers that share a network namespace

What is a container orchestrator

Software designed to join together many different machines into a cluster

What does a Dockerfile do for a container?

Specifies what needs to go into the container image

What does port forwarding do when a container is run?

Converts a container port to a host port

Why do we need to register a Docker container?

Make it globally available

What is a hypervisor

AKA VMM and runs on the physical hardware and exports hardware level abstractions to guest OS

What is a VM

Isolated environment with access to a subset of physical resources of a system

What is a process VM

Created for individual process and destroyed once that process terminates

What is a traditional VM

One supported by a bare metal hypervisor that runs directly on the hardware

What is a hosted VM

One that runs on top of an existing OS

Some machine instruction may be nonvirtualizable, what are the two methods that a guest OS can run applications with a hypervisor

Binary translation: Hypervisor monitors the execution of guest OS and replaces nonvirtualizable instructions executed by a guest OS with other instructions

Paravirtualization: The guest OS can only use instructions that can be virtualized

VT-x provides Virtual Machine Control Structure (VMCS). Explain what and from where will be loaded by VMEntry and VMExit?

Entry - loads VMX non root from guest state in VMCS

Exit - loads VMX root from host state in VMCS

What is KVM

Kernel based Virtual Machine, virtualization infrastructure of Linux

What is QEMU and what is it doing in KVM?

Virtualizer and machine emulator. Emulates devices in KVM using dynamic binary translation

What is AMI

Amazon Machine Image

Extra copy of the original image without config dependent info like the hostname

What is the main purpose of a hypervisor

Creating VMs

What is another (older) name for a hypervisor?

Virtual Machine Manager

Why are Type 1 hypervisors also called bare metal or native ones

Executing directly on the processor

Why is KVm a Type 2 hypervisor?

This hypervisor is scheduled by the host kernel CPU scheduler and guests appear as processes on the host

Which component of KVM is eliminated in Nitro?

I/O Proxy

What is a Xen domain

VM

What is the main function function of Xen’s Dom0?

Administration and launching of new VMs/domains

How does hypercall demonstrate the para virtualization of Xen

Apps make system calls using hypercalls processed by Xen

What network communication feature is degraded by Xen guests but greatly enhanced by Xen 2.0

Send data rate

In nested virtualization, how sensitive instructions by a guest hypervisor are handled depends on the hardware support. How differently these instructions are handled with single level vs multi level/nested hardware virtualization support

Single: All traps handled by the host/bottom hypervisor

Multi-level: Each Hypervisor handles all traps caused by sensitive instructions of guest hypervisors running directly above of it

What is a VMBR? What can it support a malicious OS to do?

VM Based rootkit. Observe the data, the events, or the state of the target system (and others)

Can a virtual CPU run on more than one physical CPU at the same time?

Probably not

How do you find how many servers this connects for A and B 

For a it is 16 × 12 because you count the servers which are the circles.

For B it is 12 × 16 for the same reasons

How many switches does each use?

For A it is 12 + 2 because that is the rectangles

For B it is 16 + 2 for the same reasons

How many different path lengths a pair of servers need to go through to connect in each network

For A it is lengths (2, 4, 5), this is the circles to the big rectangles

For b it is lengths (2, 4) for the same reasons

What is the total number of wires used in each network?

For A it is 24 * 12 + 48. We have 24 on the middle rectangles, 12 total servers, and you have to include the number between the two top ones

For B it is 24 * 16

NVMe may give one queue to each process or thread for read, write and other commands, thus provides high [blank] performance

I/O

GFS eliminates caching at the client site. Caching increases the overhead for maintaining what among cached copies

Consistency

The Chubby cell in GFS may consist of k replicas, one of which is elected as the master. When it receives a write request, the master propagates the request to all replicas and waits for a reply from [blank] replicas before responding

Majority

When it receives a read request does the master respond with or without consulting the replicas

WithoutG

GFS’s Chubby cell uses a version of the Paxos algorithm to elect a master among the surviving what after a loss of the current master

ReplicasT

The Paxos algorithm is a protocol for distributed what

Consensus