Farms, Fabrics and Clouds

Steve LoughranJulio GuijarroHP Laboratories, Bristol, UK

November 2008

Farms, Fabrics and Clouds

[email protected]@hpl.hp.com

Page 2 www.smartfrog.org

Researcher at HP LaboratoriesArea of interest: DeploymentAuthor of Ant in Action

Steve Loughran

Page 3Farms, Fabrics and Clouds www.smartfrog.org

Julio Guijarro

Researcher at HP Laboratories

Area of interest: Deployment

In charge of OSS release

http://smartfrog.org/

Page 4 www.smartfrog.org

• How to host big applications across distributed resources– Automatically– Repeatably– Dynamically– Correctly– Securely

• How to manage them from installation to removal

• How to make dynamically allocated servers useful

Our research - see smartfrog.org


Who had breakfast this morning?

Question


Who harvested wheat or corn, or killed an animal

for that breakfast?

Question


Farms provide food.

It is somebody else's problem


Who is wearing clothes they wove or knitted themselves?

Question


Provisioning of clothing -fabrics- is outsourced

It is somebody else's problem


Future applications are on the Web

• Web Browser, AJAX clients

• Richer: Flash, XUL, Silverlight

• "… as a Service "

⇒Lots of code running in the server

⇒Unpredictable demand

⇒Data mining/analysis problems


Old world installation: single server

Single web server,Single DBRAID filestore

-SPOF-limitations of scale

firewall/router

Application serverDatabase

Database server

Application server

firewall/router


yesterday: clustering

Multiple web servers,Replicated DBRAID Network filestoreLoad-balancing router

-Cost-Complexity-Limitations of scale

replication protocol

Application servers

firewall/router

DB Master DB Slave

Maintains the illusion of a single server


Now: server farms

500 web servers,Distributed filestoreRented storage & CPU

Scales upNo capital outlayAgile infrastructure


Tomorrow? grid fabric. 50000 servers


Application architectures and deployment problems change radically in this world

Application architectures

September 2008

IE

Your friends arehaving more funthan you

Mozilla

Your friendsare havingmore funthan you

Chrome

Your friendsare havingmore funthan you

iPhone

You are havingmore fun thanyourfriends

Scatter/gather?Message Queue?

Tuple Space?

Diskless frontend

MemcachedJSP?PHP?

Cloud Layer

HDFSHBase

MapReduceDLucene

Ops View

92% servers upNetwork OKCost: $200/hour

Management view

David Hasslehof isvery popular today

Affiliate View

David Hasslehofkeywords costmore today

Cloud Owner View

Customer #17 isusing lots of diskspace. NormalHDD failure rate

Developer View

MR job 17 completed.

REST APIs


Application architectures

• ROA/REST• Virtualized• MapReduce• Shards• Tuple-spaces• XMPP


Virtualization


Why?

• Save on hardware (and power, space)• Dynamically move running servers• Demand creation of new images• Testing complex system configurations• Redistributing entire machine image• 'virtual appliance'


Assumptions that are now invalid

• Systems have a long lifespan• It is slow/expensive to create a new system• It is expensive to duplicate one• Systems can/should be managed by hand• Clocks proceed at the same rate• Physical RAM doesn’t get swapped out• Running machines can't be moved/cloned• Virtualization is only for testing.


Server Farms

http://www.linuxjournal.com/



• System failure is an unusual event• 100% availability can be achieved• Data is always near the server• You need physical access to the servers• Databases are the best form of storage• You need millions of $/£/€ to play


Who has the servers?

• Yahoo!, Google, MSN, Amazon, eBay: services• MMORPG Game Vendors:

World of Warcraft, Second Life• EU Grid: Scientists• HP, IBM, Sun: rent to companies (some resold)

-focus on CPU performance for enterprise• Amazon: rent to anyone with an Amazon account

-focus on startups


Amazon EC2

• Pay as you go Virtual Machine Hosting• No persistent storage other than S3 filestore -

uses HTTP GET/PUT/DELETE operations• $0.10 per CPU/hour• Resold OS images for more

(RedHat, Windows)• Rent static IP addresses for failover/balancing• New: RAID-like storage


Host

Amazon EC2

S3 Storage

AMI

(Xen VM)

AMI

(Xen VM)

/mnt

Host

AMI

(Xen VM)

AMI

(Xen VM)

Public Internet

/mnt /mnt /mnt

Fast (free) network

free access; slow initial read time pay per GET; per megabyte

$

$$$$


Demo


EC2 Limitations

• Can't talk to peers using public IP addresses• Persistent file system is a premium extra• Most addresses are dynamic• No managed redundancy/restart• No multicast IP• No movement of VMs off high-traffic racks


Amazon S3

• Multiple geo-located data storage• No limits on size• Cost of write is high (guarantee of written remotely)• Read is cheap; may be out of date• Cost: Low

S3 is a global file system that any project can afford


Amazon S3 Charges

Storage $0.15/GB/month

Upload $0.10 per GB - all data transfer in

Download$0.18 per GB - first 10 TB / month data transfer out$0.16 per GB - next 40 TB / month data transfer out$0.13 per GB - data transfer out / month over 50 TB

Requests$0.01 per 1,000 PUT or LIST$0.01 per 10,000 GET or HEAD $0 DELETE

•S3 sets the limit on costs for reliable data storage over the network

•For Amazon, indexing and writes are the big costs…small files are the enemy


MapReduce

Commodity data processing for commodity data



• Terabyte datasets are hard to work with• Code runs on a single machine• Sequential code is better than parallel code• RAID hardware is the best way to store data• Databases are better than filesystems• Low-value data isn't worth collecting

even if you don't have a use for it now


Shards



• A single farm needs to scale to infinity• You need to provide 100% availability to 100%

of users• You have to roll out simultaneous updates to

the application, changes to the DB schema, globally


XMPP

post extends GoogleChatClientWorkflow {

to "[email protected]";

login "smartfrog.two";

password xmpp.password;

message "hello, world";

}



• You can't send message to a laptop that moves around behind a firewall.

• You need to build your own monitoring infrastructure.

• Blocking RPC is a good metaphor for long-haul communications.

• You can't send messages to your server farm from your phone

• IT doesn't have their eyes on your protocol


Problems for us farmers

• Power management• Predictive disk failure management• Load balancing for availability, power • File management• Billing• Routing• Security/Isolation• How will this change server hardware?• Managing/Configuring Machine Images• Diagnostics when things go wrong


“Agile” Routers

• Handle hundreds to thousands of (concurrent) change requests/second

• Integrate with billing• Managed throttling to specific hosts• Propagation of state to peer rooters• 'agile' DHCP -short leases; mobile• Monitored bandwidth may trigger VM migration


“Agile” Operating Systems

• Design for VM-only use• Limited functionality• Limited lifespan• Fully configurable before initial boot• Adapt to changes in surrounding environment• Viable licensing model

Farms, Fabrics and Clouds

Technology

gb data

reliable data storage

data databases

amazon s3 charges s3

tb month data transfer

s3 filestore

persistent storage

multiple web servers