cloud engineering

Cloud EngineeringSoftware in DatacentersGwendal SimonDepartment of Computer ScienceInstitut Telecom2009

Gartner Hype Cycle 2009

2 / 43 Gwendal Simon Cloud Engineering

Literature

A book:“The Datacenter as a Computer”, Luiz AndréBarroso and Urs Hölzle

and scientific publications including:ACM Sigops (SOSP) and Usenix OSDIHPDC, EuroPar, OOPSLA, etc.blogs (e.g. http://perspectives.mvdirona.com/)

Disclaimer

No discussion about the impact of cloud computing:net neutralityinteractions between CDNs and ISPsprivacy and electronic human rights. . .

Focus here on how cloud computing works

Disclaimer

No discussion about the impact of cloud computing:net neutralityinteractions between CDNs and ISPsprivacy and electronic human rights. . .

Focus here on how cloud computing works

Introduction

In a nutshell

Cloud computing is a model for enabling convenient,on-demand network access to a shared pool ofconfigurable computing resources (e.g., networks,servers, storage, applications, and services) that canbe rapidly provisioned and released with minimalmanagement effort or service provider interaction

NIST: http://csrc.nist.gov/groups/SNS/cloud-computing/index.html

Why Cloud Computing

The ∗aaS paradigm:SaaS: Software as a Service

applications for end-users (salesforce.com, Google, etc.)- email, office suite, photos sharing, video storage

PaaS: Platform as a Serviceservices for web app developers (Azure, Google, etc.)- workflow facilities and various basic services (http, database)

IaaS: Infrastructure as a Serviceresources for developers (Amazon, Joyent, etc.)- servers, network equipment, memory, CPU

Why Cloud Computing

An Engineer Vision

Benefits: Outsourcing Infrastructurereduce run time and response timeminimize infrastructure riskease deployment and upgrading

Challenge: Warehouse-Scale Computersthousands of individual computing nodescostly equipments (power, conditioning, cooling)large buildings with engineering teams

An Engineer Vision

Benefits: Outsourcing Infrastructurereduce run time and response timeminimize infrastructure riskease deployment and upgrading

Challenge: Warehouse-Scale Computersthousands of individual computing nodescostly equipments (power, conditioning, cooling)large buildings with engineering teams

Few Pictures

Datacenter is Different

Datacenter vs. Desktop Software:inherent parallelismsoftware controlplatform homogeneityfault-free requirement

Datacenter vs. High-Performance Computingunpredictable inputhigh volume of datanot only computing

Datacenter is Different

Datacenter vs. Desktop Software:inherent parallelismsoftware controlplatform homogeneityfault-free requirement

Datacenter vs. High-Performance Computingunpredictable inputhigh volume of datanot only computing

Basic Elements ofa Datacenter

Computing ArchitectureEnergyDealing with Failures

Architecture Basics

Main Elements

Storage: distributed file sys. (e.g. GFS) or NAS ?GFS is cheaper and faster for read operations

Network: 1-Gbps switch with 48 ports in a rack1 port per server, 8 ports for cluster rack

→ Oversubscription factor greater than 5scarce cluster-level bandwidthattractive rack-level networking

Main Elements

Storage: distributed file sys. (e.g. GFS) or NAS ?GFS is cheaper and faster for read operations

Network: 1-Gbps switch with 48 ports in a rack1 port per server, 8 ports for cluster rack

→ Oversubscription factor greater than 5scarce cluster-level bandwidthattractive rack-level networking

System Overview (in 2009)

Main Electric Components

Cooling Challenge

Evaluating Energy Efficiency

efficiency =computationtotal energy

Power Usage Effectiveness = total energyenergy in equipment

first generation datacenter PUE was poor (often ≥ 3.0)toward PUE around 1.2

Server PUE = critical component powertotal server power

basic SPUE is 1.7state-of-the-art servers reach 1.2

and computing efficiency

and computing efficiency19 / 43 Gwendal Simon Cloud Engineering

Computing Efficiency

Benchmarking cluster-level efficiency: on-going work

Benchmarking individual computer is easier

Benchmarking cluster-level efficiency: on-going workBenchmarking individual computer is easier

Toward Energy-Proportional Servers

Fault-Tolerant Application-Level

Fault-Tolerant software infrastructure layer:masks failures of lower-layer levelsreduces hardware costeases operational procedures (e.g., upgrade)

But application-level still experiences failuresservice is in degraded modeservice is unreachableservice is corrupted (loss of data)

Fault-Tolerant Application-Level

Fault-Tolerant software infrastructure layer:masks failures of lower-layer levelsreduces hardware costeases operational procedures (e.g., upgrade)

But application-level still experiences failuresservice is in degraded modeservice is unreachableservice is corrupted (loss of data)

Origin of Impacting Failures

Cause % of eventssoftware 33 %

configuration 28 %human 13 %network 12 %hardware 11 %

other 3 %

hardware faults are masked by fault-tolerant software

Origin of Impacting Failures

Cause % of eventssoftware 33 %

configuration 28 %human 13 %network 12 %hardware 11 %

other 3 %

hardware faults are masked by fault-tolerant software

Failures and Crash

Average machine availability is ' 99.9%95% of machines restart less than once a month80% of restart events last less than 10 minutes

Software most frequent faults (in one year):DRAM soft-errors: 1% experience uncorrectable errdisk soft-errors: 3% of drives see corrupted sectors

Failures and Crash

Average machine availability is ' 99.9%95% of machines restart less than once a month80% of restart events last less than 10 minutes

Software most frequent faults (in one year):DRAM soft-errors: 1% experience uncorrectable errdisk soft-errors: 3% of drives see corrupted sectors

SoftwareInfrastructure

FundamentalsCluster-Level: MapReduceApplication-Level: Web Search

Three Layers

Infrastructure-level software:kernel, operating systems, networking libraries

Cluster-level software (middleware):specific software operating a pool of servers

Application-level software:implementation of the Internet services

Three Layers

Main Software Components

replication

partitioningload-balancinghealth checkingintegrity check

compressionweak consistency

MapReduceDynamoBigTableHadoopSawzallChubbyDryad

replicationpartitioning

load-balancinghealth checkingintegrity check

load-balancing

health checkingintegrity check

load-balancinghealth checking

integrity checkcompression

weak consistency

compression

weak consistency

OS at a Cluster-Level Scale

Resource Management: mapping tasks to resourcesshould optimize energy usage

Hardware Abstraction: handling hardware elementsshould optimize performances

Deployment Maintenance: upgrading and monitoringshould reduce manual tasks

Programming Frameworks: easing implementationshould increase programmer productivity

Motivation

Map/Reduce is a software framework for easily writing applications whichprocess vast amounts of data (multi-terabyte data-sets) in-parallel on largeclusters (thousands of nodes) of commodity hardware in a reliable,fault-tolerant manner.

Functional Programming

Two fundamentals functions:map: apply a function to a list of elements

map f [] = []| map f [x::xs] = (f x) :: (map f xs)

map square [1,2,5] → [1,4,25]

reduce: build a value from a function and a listreduce f a [] = a| reduce f a [x::xs] = reduce f (f x a) xs

reduce add 0 [1,3,6] → 10

Functional Programming

Two fundamentals functions:map: apply a function to a list of elements

map f [] = []| map f [x::xs] = (f x) :: (map f xs)

map square [1,2,5] → [1,4,25]

reduce: build a value from a function and a listreduce f a [] = a| reduce f a [x::xs] = reduce f (f x a) xs

reduce add 0 [1,3,6] → 10

MapReduce

Implementing two functions w.r.t data (key,val)map: smaller sub-problems distributed to nodes

map (inKey, inVal) → list (outKey, v)produces intermediate values with an output key

reduce: combines results of sub-problemsreduce (outKey, list v) → outValproduces an output value from intermediate values

MapReduce

Implementing two functions w.r.t data (key,val)map: smaller sub-problems distributed to nodes

map (inKey, inVal) → list (outKey, v)produces intermediate values with an output key

reduce: combines results of sub-problemsreduce (outKey, list v) → outValproduces an output value from intermediate values

Example: Word Countmap(filename, content):

for each w in content:emitInt(w, 1)

reduce(word, partCount):int result = 0for pc in partCount:

result += pcemit(result)

map(file1, “hello me, goodbye me”)→<hello,1><me,1><goodbye,1><me,1>

map(file2, “hello you, bye you”)→<hello,1><you,1><bye,1><you,1>

a given key is allocated to a given server

reduce(hello,<1,1>) → 2. . .

<hello,2><me,2><you,2><goodbye,1><bye,1>

reduce(hello,<1,1>) → 2. . .

MapReduce Implemented

Basics

Input:the Web ' 100 billion file ' 400 terabytesthe pagerank algorithma query “w1 AND w2 AND · · · AND wn”

Output:a list of files containing all words wi , i ∈ [1, n]sorted by the pagerank algorithm

Basics

Input:the Web ' 100 billion file ' 400 terabytesthe pagerank algorithma query “w1 AND w2 AND · · · AND wn”

Output:a list of files containing all words wi , i ∈ [1, n]sorted by the pagerank algorithm

Implementation Overview

Offline task: index managementbased on keywords:

a word is associated with a tablea table contains all occurrences in the web

distributed on thousands of machinesmultiple copies and weak consistency

Online task: query managementfront-end Web servers to a subset of machines:

compute and rank their local resultsall best results are combined

intermediate servers to servers of file replicafrom file pointers to a set of metadata

Implementation Overview

Offline task: index managementbased on keywords:

a word is associated with a tablea table contains all occurrences in the web

distributed on thousands of machinesmultiple copies and weak consistency

Online task: query managementfront-end Web servers to a subset of machines:

compute and rank their local resultsall best results are combined

intermediate servers to servers of file replicafrom file pointers to a set of metadata

Discussion

The user-perceived latency is less than one second:read-only operationshigh parallelismmany thousands of queries per secondtraffic variations

Networking:tiny size of data exchangespossible packet loss around the front-end servers

Discussion

The user-perceived latency is less than one second:read-only operationshigh parallelismmany thousands of queries per secondtraffic variations

Networking:tiny size of data exchangespossible packet loss around the front-end servers

Conclusion

Key Challenges

Time-scale:datacenter are expected to last 10 yearsInternet apps gain popularity in weeks

Hardware components:processors are faster and more energy efficientmemory systems and networks are not

Server evolution:more cores mean more parallelism

Key Challenges

Personal Thoughts

A new era in the Internet:the industrial era of applicationscomputer science does really matter

About nano-datacenterusing your always-on devices

Personal Thoughts

A new era in the Internet:the industrial era of applicationscomputer science does really matter

About nano-datacenterusing your always-on devices

cloud engineering

Technology