Internet Server ClustersInternet Server Clusters

Jeff Chase

Duke University, Department of Computer Science

CPS 212: Distributed Information Systems

Using Clusters for Scalable ServicesUsing Clusters for Scalable Services

Clustersare a common vehicle for improving scalability andavailability at a single service site in the network.

Are network services the “Killer App” for clusters?

• incremental scalabilityjust wheel in another box...

• excellent price/performancehigh-end PCs are commodities: high-volume, low margins

• fault-tolerance“simply a matter of software”

• high-speed cluster interconnects are on the marketSANs + Gigabit Ethernet...

cluster nodes can coordinate to serve requests w/ low latency

• “shared nothing”

[Fox/Brewer]: SNS, TACC, and All That[Fox/Brewer]: SNS, TACC, and All That

[Fox/Brewer97] proposes a cluster-based reusable softwareinfrastructure for scalable network services (“SNS”), such as:

• TranSend: scalable, active proxy middleware for the Web

think of it as a dial-up ISP in a box, in use at Berkeley

distills/transforms pages based on user request profiles

• Inktomi/HotBotsearch engine

core technology for Inktomi Inc., today with $15B market cap.

“bringing parallel computing technology to the Internet”

Potential services are based onTransformation,Aggregation,Caching, andCustomization (TACC), built above SNS.

TACCTACC

Vision: deliver “the content you want” by viewing HTML content asa dynamic, mutable medium.

1. TransformInternet content according to:

• network and client needs/limitations

e.g., on-the-fly compression/distillation [ASPLOS96], packagingWeb pages for PalmPilots, encryption, etc.

• directed by user profile database

2. Aggregatecontent from different back-end services or resources.

3. Cachecontent to reduce cost/latency of delivery.

4. Customize(seeTransform)

TranSendTranSendStructureStructure

$

$

$

FrontEnds

ProfilesControlPanel

html gif jpg

To Internet

SAN (high speed)Utility (10baseT)Coordination bus

$ Cache partition

... Datatype-specific distiller

[adapted from Armando Fox (throughhttp://ninja.cs.berkeley.edu/pubs)]

SNS/TACC PhilosophySNS/TACC Philosophy

1. Specify services by plugging generic programs into the TACCframework, and compose them as needed.

sort of like CGI with pipes

run by long-livedworkerprocesses that serve request queues

allows multiple languages, etc.

2. Worker processes in the TACC framework are looselycoordinated, independent, and stateless.

ACID vs. BASE

serve independent requests from multiple users

narrow view of a “service”: one-shot readonly requests, and staledata is OK

3. Handle bursts with designatedoverflow poolof machines.

TACC ExamplesTACC Examples

HotBotsearch engine

• Query crawler’s DB

• Cache recent searches

• Customize UI/presentation

TranSendtransformation proxy

• On-the-fly lossy compression of inline images(GIF, JPG, etc.)

• Cache original & transformed

• User specifies aggressiveness, “refinement”UI, etc.

C TT

$$AA

TT

$$

C

DBDB

htmlhtml

[Fox]

(Worker) Ignorance Is Bliss(Worker) Ignorance Is Bliss

What workers don’t need to know

• Data sources/sinks

• User customization (key/value pairs)

• Access to cache

• Communication with other workers by name

Common case: stateless workers

C, Perl, Java supported

• Recompilation often unnecessary

• Useful tasks possible in <10 lines of (buggy) Perl

[Fox]

QuestionsQuestions

1. What are the research contributions of the paper?system architecture decouples SNS concerns from content

TACC programming model composes stateless worker modules

validation using two real services, with measurements

How is this different from clusters for parallel computing?

2. How is this different from clusters for parallel computing?

3. What are the barriers to scale in SNS/TACC?

4. How are requests distributed to caches, FEs, workers?

5. What can we learn from the quantitative results?

6. What about services that allow client requests to update shareddata?

e.g., message boards, calendars, mail,

SNS/TACC Functional IssuesSNS/TACC Functional Issues

1. What about fault-tolerance?

• Service restrictions allow simple, low-cost mechanisms.

Primary/backup process replication is not necessary with BASEmodel and stateless workers.

• Uses aprocess-peerapproach to restart failed processes.

Processes monitor each other’s health and restart if necessary.

Workers and manager find each other with “beacons” on well-known ports.

2. Load balancing?

• Manager gathers load info and distributes to front-ends.

• How are incoming requests distributed to front-ends?

Porcupine: A Highly Available ClusterPorcupine: A Highly Available Cluster--based Mail Servicebased Mail Service

Yasushi Saito

Brian Bershad

Hank Levy

University of WashingtonDepartment of Computer Science and Engineering,

Seattle, WA

http://porcupine.cs.washington.edu/

[Saito]

Why Email?Why Email?

Mail is importantReal demand

Mail is hardWrite intensive

Low locality

Mail is easyWell-defined API

Large parallelism

Weak consistency

[Saito]

How much of Porcupine isreusable to other services?

Can we use the SNS/TACCframework for this?

GoalsGoals

Use commodity hardware to build a large, scalable mail service

Three facets of scalability ...

Performance: Linear increase with cluster size

Manageability: React to changes automatically

Availability: Survive failures gracefully

[Saito]

Conventional Mail SolutionConventional Mail Solution

Static partitioning

Performance problems:No dynamic load balancing

Manageability problems:Manual data partition decision

Availability problems:

Limited fault tolerance

SMTP/IMAP/POP

Bob’smbox

Ann’smbox

Joe’smbox

Suzy’smbox

NFS servers

[Saito]

Key Techniques and RelationshipsKey Techniques and Relationships

Functional Homogeneity“any node can perform any task”

AutomaticReconfiguration

LoadBalancingReplication

Manageability PerformanceAvailability

Framework

Techniques

Goals

[Saito]

Porcupine ArchitecturePorcupine Architecture

Node A ...Node B Node Z...

SMTPserver

POPserver

IMAPserver

Mail mapMailboxstorage

Userprofile

Replication Manager

MembershipManager

RPC

Load Balancer

User map

[Saito]

Porcupine OperationsPorcupine Operations

ÿþýüûþüý

A B...

A

� � � �üþ����� ý�����

� � � � ���þ��ü����� ÿ �

� � � �üû������ �

�� ���� ý�ü �ü�ýþ��ü� ý� �ý�ûüþü� ��� ÿ �

� � � �� ��ü�ü�ý��þ

� � �� � ���� � ������ �þ� �þ� � � � �� ý�ûü

����B

C

Protocolhandling

Userlookup

LoadBalancing

Messagestore

...C

[Saito]

Basic Data StructuresBasic Data Structures“bob”

BCA CA BA C

bob : {A,C}ann : {B}

BCA CA BA C

suzy : {A,C} joe : {B}

BCA CA BA C

Apply hashfunction

User map

Mail map/user info

Mailboxstorage

A B C

Bob’sMSGs

Suzy’sMSGs

Bob’sMSGs

Joe’sMSGs

Ann’sMSGs

Suzy’sMSGs

[Saito]

fragment list

mailbox fragments

Porcupine AdvantagesPorcupine Advantages

Advantages:Optimal resource utilization

Automatic reconfiguration and task re-distribution upon nodefailure/recovery

Fine-grain load balancing

Results:Better Availability

Better Manageability

Better Performance

[Saito]

AvailabilityAvailability

Goals:Maintain function after failuresReact quickly to changes regardless of cluster sizeGraceful performance degradation / improvement

Strategy:Two complementary mechanisms

Hard state:email messages, user profileÿ Optimistic fine-grain replication

Soft state: user map, mail mapÿ Reconstruction after membership change

[Saito]

SoftSoft--state Reconstructionstate Reconstruction

B C A B A B A C

bob : {A,C}

joe : {C}B C A B A B A C

B A A B A B A B

bob : {A,C}

joe : {C}B A A B A B A B

A C A C A C A C

bob : {A,C}

joe : {C}A C A C A C A C

suzy : {A,B}

ann : {B}

1. Membership protocolUsermap recomputation

2. Distributeddisk scan

suzy :

ann :

Timeline

A

B

ann : {B}

B C A B A B A Csuzy : {A,B}

Cann : {B}

B C A B A B A Csuzy : {A,B}

ann : {B}

B C A B A B A Csuzy : {A,B}

[Saito]

suzy

ann

How does Porcupine React toHow does Porcupine React toConfiguration Changes?Configuration Changes?

300

400

500

600

700

0 100 200 300 400 500 600 700 800Time(seconds)

Messages/second

No failure

One nodefailureThree nodefailuresSix nodefailures

Nodesfail

Newmembershipdetermined

Nodesrecover

Newmembershipdetermined

[Saito]

HardHard--state Replicationstate Replication

Goals:Keep serving hard state after failures

Handle unusual failure modes

Strategy: Exploit Internet semanticsOptimistic, eventually consistent replication

Per-message, per-user-profile replication

Efficient during normal operation

Small window of inconsistency

[Saito]

How will Porcupine behave in a partition failure?

More on Porcupine ReplicationMore on Porcupine Replication

To add/delete/modify a message:• Find and update any replica of the mailbox fragment.

Do whatever it takes: make a new fragment if necessary...pick anew replica if chosen replica does not respond.

• Replica asynchronously transmits updates to other fragment replicas.continuous reconciling of replica states

• Log/force pending update state, and target nodes to receive update.on recovery, continue transmitting updates where you left off

• Order updates byloosely synchronizedphysical clocks.Clock skew should be less than the inter-arrival gap for a sequence

of order-dependent requests...usenodeIDto break ties.

• How many node failures can Porcupine survive? What happens ifnodes fail “forever”?

How Efficient is Replication?How Efficient is Replication?

0100200300400500600700800

0 5 10 15 20 25 30Cluster size

Mes

sage

s/se

cond

Porcupine no replication

Porcupine with replication=2

68m/day

24m/day

[Saito]

How Efficient is Replication?How Efficient is Replication?

0100200300400500600700800

0 5 10 15 20 25 30Cluster size

Mes

sage

s/se

cond

Porcupine no replication

Porcupine with replication=2

Porcupine with replication=2, NVRAM

68m/day

24m/day33m/day

[Saito]

Load balancing: Deciding where to store messagesLoad balancing: Deciding where to store messages

Goals:Handle skewed workload well

Support hardware heterogeneity

No voodoo parameter tuning

Strategy: Spread-based load balancingSpread:soft limit on # of nodes per mailbox

Large spreadÿ better load balance

Small spreadÿ better affinity

Load balanced within spread

Use # of pending I/O requests as the load measure

[Saito]

QuestionsQuestions• How to select the front-end node to handle the request? Does it

matter which one we choose?

• Don’t we already know how to build big mail servers? (e.g.,Earthlink, Christenson USITS97) Why do we need Porcupine?

• What properties of the mail “data model” allow this approach, withweaker consistency guarantees than a database?

• How does the system leverage/exploit the weaker semantics?

• Can the architecture accommodate new features, e.g., Pachyderm-like storage/indexing of large mail collections?

• Could I run Porcupine on the same cluster with other applications?

• Could this have been built on Microsoft’s MSCS? How muchapplication effort would have been saved?

Clusters: A Broader ViewClusters: A Broader View

MSCS (“Wolfpack”) is designed as basic infrastructure forcommercial applications on clusters.

• “A cluster service is a package of fault-tolerance primitives.”

• Service handles startup, resource migration, failover, restart.

• But: apps may need to be “cluster-aware”.

Apps must participate in recovery of their internal state.

Use facilities for logging, checkpointing, replication, etc.

• Service and node OS supports uniform naming and virtualenvironments.

Preserve continuity of access to migrated resources.

Preserve continuity of the environment for migrated resources.

Wolfpack: ResourcesWolfpack: Resources

• The components of a cluster arenodesandresources.

Shared nothing: each resource is owned by exactly one node.

• Resources may bephysicalor logical.

Disks, servers, databases, mailbox fragments, IP addresses,...

• Resources have types, attributes, and expected behavior.

• (Logical) resources are aggregated inresource groups.

Each resource is assigned to at most one group.

• Some resources/groups depend on other resources/groups.

Admin-installed registry lists resources and dependency tree.

• Resources can fail.

cluster service/resource managers detect failures.

FaultFault--Tolerant Systems: The Big PictureTolerant Systems: The Big Picture

messagingsystem

file/storagesystem

database mail service clusterservice

applicationservice

applicationservice

redundant hardwareparityECC

replicationRAID parity

checksumack/retransmission

replicationlogging

checkpointingvoting

replicationlogging

checkpointingvoting

Note:dependenciesredundancy at any/each/every levelwhat failure semantics to the level above?

Wolfpack: Resource Placement and MigrationWolfpack: Resource Placement and Migration

The cluster service detects component failures and responds byrestarting resources or migrating resource groups.

• Restart resource in place if possible...

• ...else find another appropriate node and migrate/restart.

Ideally, migration/restart/failover is transparent.

• Logical resources (processes) execute in virtual environments.

uniform name space for files, registry, OS objects (NT mods)

• Node physical clocks are loosely synchronized, with clock drift lessthan minimal time for recovery/migration/restart.

guarantees migrated resource sees monotonically increasing clocks

• Route resource requests to the node hosting the resource.

• Is the failure visible to other resources that depend on the resource?

Membership 101Membership 101

Cluster nodes must agree on the set of cluster members (theview).

• distribute resource ownership effectively

shift resources on node failures or additions

• eliminate dangerous/expensive interactions with faulty nodes

• “keep everyone in the loop” on updates and events

e.g., multicast groups and group communication

The literature on group membership is tangled up with the problemof ordered multicast (e.g., “CATOCS”).

• What are the ordering guarantees for message delivery, especiallywith respect to membership changes?

• Ordered group communication is controversial, but everyone needsa solution for the separate but related membership problem.

Failure DetectorsFailure Detectors

First problem: how to detect that a member has failed?

• pings, timeouts, beacons, heartbeats

• recovery notifications

“I was gone for awhile, but now I’m back.”

Is the failure detectoraccurate?

Is the failure detectorlive?

In an asynchronous system, it is possible for a failure detector to beaccurate or live, but not both.

• As it turns out, it is impossible for an asynchronous system to agreeon anythingwith accuracy and liveness!

• But this is academic...

Failure Detectors in Real SystemsFailure Detectors in Real Systems

Common solution:• Use a failure detector that is live but not accurate.

Assume bounded processing delays and delivery times.

Timeout with multiple retries detects failure accurately with highprobability.

If a “failed” site turns out to be alive, then kill it (fencing).

• Use a recovery detector that is accurate but not live.“I’m back....hey, did anyone hear me?”

What do we assume about communication failures?How much pinging is enough?

1-to-N, N-to-N, ring?

What about network partitions?

Membership ServiceMembership Service

Second problem: How to propagate knowledge of failure/recoveryevents to other nodes?

• Surviving nodes should agree on the new view (regrouping).

• Convergence should be rapid.

• The regrouping protocol should itself be tolerant of message drops,message reorderings, and failures.

liveness and accuracy again

• The regrouping protocol should be scalable.

• The protocol should handle network partitions.

• Behavior of the messaging system (e.g., group multicast) acrossmembership changes must be well-specified and understood.

Example: WombatExample: Wombat

• Wombat is a new membership protocol, an outgrowth of Porcupine.

Gretta Bartels, University of Washington, Duke ‘98

• Wombat is empirically more efficient/scalable than competingalgorithms such asThree Round.

• But: Wombat makes no guarantees about the relative ordering ofmembership events and messages.

Adherents of group communication would not accept it as a “real”membership protocol.

• Wombat’s assumptions have not been formally defined, and itsproperties have not been proven.

If you can’t prove that it works, you can’t believe that it works.

• Disclaimer: Wombat is a promising work in progress.

Wombat BasicsWombat Basics

ping

ping

leader

minions

Nodes are ranked by unique IDs.

Node IDs are permanent.

Nodei pingspredecessor(i).

The highest-ranked node is theleader.

All other nodes areminions.

The leader periodically broadcasts itsviewto all known minions.

physicalbroadcast

Minions adopt the leader’s view.

determine pred from leader’s view

Node Arrival/Recovery in WombatNode Arrival/Recovery in Wombat

If node i joins the cluster:

1. i waits for the leader’s next beacon.

2. i detects that the leader’s view doesnot includei.

3. i notifies the leader.

4. The leader updates its view.

5. The leader broadcasts its new view.

6. Minions adopt the leader’s view.

“I’m here too.”

i

Node Failure in WombatNode Failure in Wombat

If a node fails:

1. Its successor notifies the leader.

2. The leader updates its view.

3. The leader broadcasts its view.

4. Minions adopt the leader’s new view.

5. Life goes on.

X

“Node i has failed.”

i

Leader Failure in WombatLeader Failure in Wombat

If the leader fails:

1. Successor detects the failure.

2. Successor knows that the failed nodewas the leader.

3. Successor broadcasts as leader.

4. Minions adopt the new leader’s view.

5. Life goes on.

X “I am in control.”

Multiple Failures in WombatMultiple Failures in Wombat

If the leader and its successor(s) fail(s),the next ranking node must assumecommand on its own.

1. Each node has abroadcast timer; ifthe timer goes off, broadcast as leader.

2. Each node’s timer is set by its rank.

if i< j thentimer(i)<timer(j)

3. Reset timer on each beacon.

4. Leader’s timer value is adaptive.

Go faster if things are changing.

X

“I must be incontrol.”

X

Suppressing False LeadersSuppressing False Leaders

If a node falsely broadcasts as leader:

1. All nodes that know of a better leaderrecognize the usurper as such.

2. The real leader recognizes that it is abetter leader than the usurper.

3. The real leader broadcasts the union ofits view and the usurper’s view.

4. The usurper shuts up and adopts thereal leader’s view.

What if the “real leader” is dead?

X

“I must be incontrol.”

“I don’tthink so.”

Partitions in WombatPartitions in WombatpartitionleaderIf a network failure partitions the

cluster:

1. The old partition continues.

2. The leader of the new partitioneventually broadcasts its view.

3. Minions accept the new leader’sview.

partitionleader

notion

Healing a PartitionHealing a Partitiondominating

leaderWhen the partition heals, either:

1. The dominating partition leaderhears a false broadcast, and...

2. ...corrects it by broadcasting theunion of the views.

- or -

1. The dominating partition leaderbroadcasts first, and...

2. ...minions respond “I’m here”.

partitionleader

Wombat: WrinklesWombat: Wrinkles

1. What are the assumptions about:• network?

• clocks?

2. Are these reasonable/realistic assumptions?

3. How to ensure a single cluster view in the event of a partition?

4. How long does it take for the view to converge after a partition?

5. How do we start a cluster? What if a node starts or recovers butnever receives a beacon?

6. What about the ordering of messages and membership events?

7. How do minions come to accept a new leader?

8. What about “message storms”?