Cloud Don McGregor Research Associate MOVES Institute [email protected].

Cloud

Don McGregor

Research Associate

MOVES Institute

[email protected]

2

Data Centers

• What sorts of things go on in a data center?

3

Data Center Problems

• “I want to set up a publicly visible web server”

• “I want to have 50 servers working in parallel on my social modeling simulation”

• “I’m a facebook competitor and will start with 500 webservers and 50 databases, then scale up from there”

4

Data Centers

• Really old solution: you buy some computers to do this task for your group. Say, ten desktops, crammed into a spare room

• A better solution: there’s a central data center on campus with AC, backup generator, fire suppression, power, access control, etc

• But this involves having your organization buy, rack, set up, and configure N boxes

5

Virtual Servers

• A key piece of technology is “virtual machines”. The idea is to have one piece of hardware run multiple operating systems at once

• Often seen with VMWare, Virtual Box, Xen, and others.

• The idea is to run multiple operating systems on one piece of hardware—say, two Linux instances, one windows instance, etc

6

One Box, Multiple OSes

7

Virtual Machines

• Most modern hardware is hardly taxed most of the time

• We can run multiple operating systems at the same time on one piece of hardware, just like we can run multiple programs at the same time within one OS

• So the idea is to buy racks full of powerful hardware and run multiple VMs on each piece of hardware

8

Virtual Machines

• This gives us all sorts of benefits: for example we can migrate a VM from one physical machine to another if we have hardware problems

• Lets us make better use of the hardware• Instead of buying and configuring a new web

server, call the IT department and tell them to spin up a new OS VM on existing hardware

• The virtual machine can be configured in an automated way in a few seconds

9

VM Configuration

• Start up an operating system on a virtual machine. But depending on the purpose of the VM, we might want to configure it differently

• Some have web servers, some have databases, some run your simulation, and different software needs to be installed on the VM for each purpose

• It’s not practical to do this by hand• There are tools to configure the VM automatically,

which is important if you want to start new VMs based on load

10

Data Centers

• But once you have this capability, you start asking yourself: why am I running a physical data center in the first place? It’s expensive, takes trained people, infrastructure is not cheap, it’s not a core competency for many businesses, and there are economies of scale to building a 10,000 CPU data center

• Companies have sprung up that provide data center services

11

Scale Out

12

Scale Out

13

Cloud Providers

• Amazon Web Services—preferred cloud provider for the Navy– Has approved infrastructure for ITAR issues

• Rackspace• Many providers for games

14

Data Center Scale Out

• You can make really big data centers, and make management of devices within those data centers very scalable

• The marginal costs of adding another 100 hosts to a big data center is much smaller than for most small data centers, as is the marginal cost of management

• http://opencompute.org/ for standards on data centers: CPUs, racks, storage, power supplies, etc

15

OpenStack

• OpenCompute deals with hardware and datacenter issues

• OpenStack deals with the software side of the house– How to handle multiple images (OS virtual

machines that can be started)– Storage – Networking– Management– Compute

16

OpenStack

• Very fast-moving project; major releases every six months

• Most hardcopy documentation is out of date as a result (eg, O’Reilly books, which are usually quite good)

• Supported by Rackspace, compatible with Amazon Web Services (aws)

17

Clouds

• Old model: Call IT department, have them spin up a new VM for your web server

• New model: call cloud provider, have them spin up new VM for your web server on their very cheap and scalable data center

• What are we losing here? – Control of data (maybe)– Central point of failure (network, power,

building) (maybe)– Latency (maybe)

18

Clouds

• Amazon Elastic Cloud (EC): call them, they spin up a VM for web server

• Call Rackspace, have them spin up 500 Linux compute nodes for two hours to run a cultural sim, along with one Windows node for a month

• You can get sophisticated and have the number of nodes spun up depend on the load: for example, under heavy load you have 200 web servers, then on the weekend dial back to 50 web servers

19

Cloud

• Platform as a Service (PaaS): the cloud vendor sells operating system VMs

• Software as a Service (SaaS): the cloud vendor sells an application (calendaring, email, etc.)

• Will your business survive several days of downtime?

20

Scalability

• Famously, Zynga used clouds for Farmville. How many people would play Farmville? They didn’t know, and they didn’t want to build a data center themselves

• Outsourced data center, number of hosts they use depends on load; as load goes up, more pre-configured hosts are spun up to handle it

21

Business Model Implications

• Old way: come up with idea, code, get VC money, lease space, buy AC, buy boxes, buy network, upgrade plan for boxes …

• New way: Come up with idea, deploy demo on cloud with your credit card. If it works, get VC/angel money. Done.

• Dramatically lower cost of entry

22

Private Clouds

• For the cases were you can’t outsource because of concerns about data, you can set up “private clouds”

• Use the same software as the commercial providers, but with a data center your organization controls

• For example, maybe a DoD private cloud, or an NPS private cloud

• Amazon EC does some FIPS (http://aws.amazon.com/federal/)

23

Examples

• “I want to run a publicly visible web server”– Call Amazon EC, small web sites free, roughly 5

cents per compute hour after that (depending on lots of options), plus data bandwidth charges

• “I need 500 nodes to do my social modeling compute sim”– Large compute VM at 10 cents per compute

hour

• Compare this to the cost of setting up your own data center

24

Implications for M&S?

• If you can do web-based M&S, you can put the back end on the cloud

• Implications for dynamic scaling at short notice

• From a program management standpoint, fewer big projects depending on up-front planning, lower risk