Distributed-Systems

CS60002 Distributed Systems

Text Book:• “Advanced Concepts in Operating

Systems” by Mukesh Singhal and Niranjan G. Shivaratri

will cover about half the course, supplemented by copies of papers

Xerox, notes, copies of papers etc. will cover the rest.

What is a distributed system?

A very broad definition:

A set of autonomous processes communicating among themselves to perform a task

Autonomous: able to act independently

Communication: shared memory or message passing

“Concurrent system” : a better term probably

A more restricted definition:• A network of autonomous

computers that communicate by message passing to perform some task

A practical “distributed system” will probably have both– Computers that communicate by

messages– Processes/threads on a

computer that communicate by messages or shared memory

Advantages

• Resource Sharing• Higher Performance• Fault Tolerance

• Scalability

Why is it hard to design them?

The usual problem of concurrent systems:– Arbitrary interleaving of actions

makes the system hard to verify

Plus• No globally shared memory

(therefore hard to collect global state)

• No global clock• Unpredictable communication

delays

Models for Distributed Algorithms

– Topology : completely connected, ring, tree etc.

– Communication : shared memory/message passing (reliable? Delay? FIFO/Causal? Broadcast/multicast?)

– Synchronous/asynchronous– Failure models (fail stop, crash,

omission, Byzantine…)

An algorithm need to specify the model on which it is supposed to work

Complexity Measures

• Message complexity : no. of messages

• Communication complexity/Bit Complexity : no. of bits

• Time complexity : For synchronous systems, no. of rounds. For asynchronous systems, different definitions are there.

Some Fundamental Problems

• Ordering events in the absence of a global clock

• Capturing the global state• Mutual exclusion• Leader election• Clock synchronization

• Termination detection• Constructing spanning trees• Agreement protocols

Ordering of Events and

Logical Clocks

Ordering of Events

Lamport’s Happened Before relationship:

For two events a and b, a → b if a and b are events in the same

process and a occurred before b a is a send event of a message

m and b is the corresponding receive event at the destination process

a → c and c → b for some event c

a → b implies a is a potential cause of b

Causal ordering : potential dependencies

“Happened Before” relationship causally orders events• If a → b, then a causally affects

b• If a → b and b → a, then a and b

are concurrent ( a || b)

Logical Clock

Each process i keeps a clock Ci.

• Each event a in i is timestamped C(a), the value of Ci when a

occured

• Ci is incremented by 1 for each

event in i• In addition, if a is a send of

message m from process i to j, then on receive of m,

Cj = max(Cj, C(a)+1)

Points to note:

• if a → b, then C(a) < C(b)

• → is an irreflexive partial order

• Total ordering possible by arbitrarily ordering concurrent events by process numbers

Limitation of Lamport’s Clock

a → b implies C(a) < C(b)

BUT

C(a) < C(b) doesn’t imply a → b !!

So not a true clock !!

Solution: Vector Clocks

Ci is a vector of size n (no. of processes)

C(a) is similarly a vector of size n

Update rules:

• Ci[i]++ for every event at process

i• if a is send of message m from i

to j with vector timestamp tm, on receive of m:

Cj[k] = max(Cj[k], tm[k]) for all k

For events a and b with vector timestamps ta and tb,

• ta = tb iff for all i, ta[i] = tb[i]

• ta ≠ tb iff for some i, ta[i] ≠ tb[i]

• ta ≤ tb iff for all i, ta[i] ≤ tb[i]

• ta < tb iff (ta ≤ tb and ta ≠ tb)

• ta || tb iff (ta < tb and tb < ta)

• a → b iff ta < tb

• Events a and b are causally related iff ta < tb or tb < ta, else they are concurrent

• Note that this is still not a total order

Causal ordering of messages: application of vector clocks

• If send(m1)→ send(m2), then every recipient of both message m1 and m2 must “deliver” m1 before m2.

“deliver” – when the message is actually given to the application for processing

Birman-Schiper-Stephenson Protocol

• To broadcast m from process i, increment Ci(i), and timestamp m with VTm = Ci[i]

• When j ≠ i receives m, j delays delivery of m until – Cj[i] = VTm[i] –1 and

– Cj[k] ≥ VTm[k] for all k ≠ i

– Delayed messaged are queued in j sorted by vector time. Concurrent messages are sorted by receive time.

• When m is delivered at j, Cj is updated according to vector clock rule.

Problem of Vector Clock

• message size increases since each message needs to be tagged with the vector

• size can be reduced in some cases by only sending values that have changed

Capturing Global State

Global State Collection

Applications: – Checking “stable” properties,

checkpoint & recovery

Issues:– Need to capture both node and

channel states

– system cannot be stopped– no global clock

Some notations:– LSi : local state of process i– send(mij) : send event of

message mij from process i to process j

– rec(mij) : similar, receive instead of send

– time(x) : time at which state x was recorded

– time (send(m)) : time at which send(m) occured

send(mij) є LSi iff

time(send(mij)) < time(LSi)

rec(mij) є LSj iff

time(rec(mij)) < time(LSj)

transit(LSi,LSj) = { mij | send(mij) є LSi and rec(mij) є LSj}

inconsistent(LSi, LSj) = {mij | send(mij) є LSi and rec(mij) є LSj}

Global state: collection of local states

GS = {LS1, LS2,…, LSn}

GS is consistent iff for all i, j, 1 ≤ i, j ≤ n,

inconsistent(LSi, LSj) = Ф

GS is transitless iff for all i, j, 1 ≤ i, j ≤ n,

transit(LSi, LSj) = Ф

GS is strongly consistent if it is consistent and transitless.

Chandy-Lamport’s Algorithm

• Uses special marker messages.

• One process acts as initiator, starts the state collection by following the marker sending rule below.

• Marker sending rule for process P:– P records its state; then for each

outgoing channel C from P on which a marker has not been sent already, P sends a marker along C before any further message is sent on C

• When Q receives a marker along a channel C:

– If Q has not recorded its state then Q records the state of C as empty; Q then follows the marker sending rule

– If Q has already recorded its state, it records the state of C as the sequence of messages received along C after Q’s state was recorded and before Q received the marker along C

Points to Note:

• Markers sent on a channel distinguish messages sent on the channel before the sender recorded its states and the messages sent after the sender recorded its state

• The state collected may not be any state that actually happened in reality, rather a state that “could have” happened

• Requires FIFO channels• Network should be strongly

connected (works obviously for connected, undirected also)

• Message complexity O(|E|), where E = no. of links

Lai and Young’s Algorithm

• Similar to Chandy-Lamport’s, but does not require FIFO

• Boolean value X at each node, False indicates state is not recorded yet, True indicates recorded

• Value of X piggybacked with every application message

• Value of X distinguishes pre-snapshot and post-snapshot messages, similar to the Marker

Mutual Exclusion

Mutual Exclusion

• very well-understood in shared memory systems

• Requirements:– at most one process in critical

section (safety)– if more than one requesting

process, someone enters (liveness)

– a requesting process enters within a finite time (no starvation)

– requests are granted in order (fairness)

Classification of Distributed Mutual Exclusion Algorithms

• Non-token based/Permission based– Permission from all processes:

e.g. Lamport, Ricart-Agarwala, Raicourol-Carvalho etc.

– Permission from a subset: ex. Maekawa

• Token based

– ex. Suzuki-Kasami

Some Complexity Measures

• No. of messages/critical section entry

• Synchronization delay• Response time

• Throughput

Lamport’s Algorithm

• Every node i has a request queue qi, keeps requests sorted by logical timestamps (total ordering enforced by including process id in the timestamps)

To request critical section:– send timestamped REQUEST

(tsi, i) to all other nodes– put (tsi, i) in its own queue

On receiving a request (tsi, i):– send timestamped REPLY to the

requesting node i – put request (tsi, i) in the queue

To enter critical section:– i enters critical section if (tsi, i) is

at the top if its own queue, and i has received a message (any message) with timestamp larger than (tsi, i) from ALL other nodes.

To release critical section:– i removes it request from its own

queue and sends a timestamped RELEASE message to all other nodes

– On receiving a RELEASE message from i, i’s request is removed from the local request queue

Some points to note:

• Purpose of REPLY messages from node i to j is to ensure that j knows of all requests of i prior to sending the REPLY (and therefore, possibly any request of i with timestamp lower than j’s request)

• Requires FIFO channels. • 3(n – 1 ) messages per critical

section invocation• Synchronization delay = max.

message transmission time• requests are granted in order of

increasing timestamps

Ricart-Agarwala Algorithm

• Improvement over Lamport’s• Main Idea:

– node j need not send a REPLY to node i if j has a request with timestamp lower than the request of i (since i cannot enter before j anyway in this case)

• Does not require FIFO• 2(n – 1) messages per critical

section invocation• Synchronization delay = max.

message transmission time• requests granted in order of

increasing timestamps

To request critical section:– send timestamped REQUEST

message (tsi, i)

On receiving request (tsi, i) at j:– send REPLY to i if j is neither

requesting nor executing critical section or if j is requesting and i’s request timestamp is smaller than j’s request timestamp. Otherwise, defer the request.

To enter critical section:– i enters critical section on

receiving REPLY from all nodes

To release critical section:– send REPLY to all deferred

requests

Roucairol-Carvalho Algorithm

• Improvement over Ricart-Agarwala• Main idea

– once i has received a REPLY from j, it does not need to send a REQUEST to j again unless it sends a REPLY to j (in response to a REQUEST from j)

– no. of messages required varies between 0 and 2(n – 1) depending on request pattern

– worst case message complexity still the same

Maekawa’s Algorithm

• Permission obtained from only a subset of other processes, called the Request Set (or Quorum)

• Separate Request Set Ri for each process i

• Requirements:

– for all i, j: Ri ∩ Rj ≠ Φ

– for all i: i Є Ri

– for all i: |Ri| = K, for some K

– any node i is contained in exactly D Request Sets, for some D

• K = D = sqrt(N) for Maekawa’s

A simple version

To request critical section:– i sends REQUEST message to

all process in Ri

On receiving a REQUEST message:– send a REPLY message if no

REPLY message has been sent since the last RELEASE message is received. Update status to indicate that a REPLY has been sent. Otherwise, queue up the REQUEST

To enter critical section:– i enters critical section after

receiving REPLY from all nodes in Ri

To release critical section:

– send RELEASE message to all nodes in Ri

– On receiving a RELEASE message, send REPLY to next node in queue and delete the node from the queue. If queue is empty, update status to indicate no REPLY message has been sent.

• Message Complexity: 3*sqrt(N)• Synchronization delay =

2 *(max message transmission time)

• Major problem: DEADLOCK possible

• Need three more types of messages (FAILED, INQUIRE, YIELD) to handle deadlock. Message complexity can be 5*sqrt(N)

• Building the request sets?

Token based Algorithms

• Single token circulates, enter CS when token is present

• No FIFO required• Mutual exclusion obvious• Algorithms differ in how to find and

get the token• Uses sequence numbers rather

than timestamps to differentiate between old and current requests

Suzuki Kasami Algorithm

• Broadcast a request for the token• Process with the token sends it to

the requestor if it does not need it

Issues:

– Current vs. outdated requests– determining sites with pending

requests– deciding which site to give the

token to

• The token:– Queue (FIFO) Q of requesting

processes– LN[1..n] : sequence number of

request that j executed most recently

• The request message:– REQUEST(i, k): request

message from node i for its kth critical section execution

• Other data structures– RNi[1..n] for each node i, where

RNi[j] is the largest sequence number received so far by i in a REQUEST message from j.

To request critical section:– If i does not have token,

increment RNi[i] and send REQUEST(i, RNi[i]) to all nodes

– if i has token already, enter critical section if the token is idle (no pending requests), else follow rule to release critical section

On receiving REQUEST(i, sn) fat j:– set RNj[i] = max(RNj[i], sn)– if j has the token and the token is

idle, send it to i if RNj[i] = LN[i] + 1. If token is not idle, follow rule to release critical section

To enter critical section:– enter CS if token is present

To release critical section:

– set LN[i] = RNi[i]

– For every node j which is not in Q (in token), add node j to Q if RNi[ j ] = LN[ j ] + 1

– If Q is non empty after the above, delete first node from Q and send the token to that node

Points to note:

– No. of messages: 0 if node holds the token already, n otherwise

– Synchronization delay: 0 (node has the token) or max. message delay (token is elsewhere)

– No starvation

Raymond’s Algorithm

• Forms a directed tree (logical) with the token-holder as root

• Each node has variable “Holder” that points to its parent on the path to the root. Root’s Holder variable points to itself

• Each node i has a FIFO request queue Qi

• To request critical section:

– Send REQUEST to parent on the tree, provided i does not hold the token currently and Qi is empty. Then place request in Qi

• When a non-root node j receives a request from i

– place request in Qj

– send REQUEST to parent if no previous REQUEST sent

• When the root receives a REQUEST:– send the token to the requesting

node– set Holder variable to point to

that node• When a node receives the token:

– delete first entry from the queue– send token to that node– set Holder variable to point to

that node– if queue is non-empty, send a

REQUEST message to the parent (node pointed at by Holder variable)

• To execute critical section:– enter if token is received and

own entry is at the top of the queue; delete the entry from the queue

• To release critical section– if queue is non-empty, delete first

entry from the queue, send token to that node and make Holder variable point to that node

– If queue is still non-empty, send a REQUEST message to the parent (node pointed at by Holder variable)

Points to note:

• Avg. message complexity O(log n)

• Sync. delay (T log n)/2, where T = max. message delay

Leader Election

Leader Election in Rings

• Models– Synchronous or Asynchronous– Anonymous (no unique id) or

Non-anonymous (unique ids)– Uniform (no knowledge of ‘n’,

the number of processes) or non-uniform (knows ‘n’)

• Known Impossibility Result:– There is no Synchronous, non-

uniform leader election protocol for anonymous rings

– Implications ??

Election in Asynchronous Rings

Lelann-Chang-Robert’s Algorithm– send own id to node on left

– if an id received from right, forward id to left node only if received id greater than own id, else ignore

– if own id received, declares itself “leader”

• works on unidirectional rings

• message complexity = θ(n^2)

• Hirschberg-Sinclair Algorithm– operates in phases, requires

bidirectional ring– In kth phase, send own id to 2^k

processes on both sides of yourself (directly send only to next processes with id and k in it)

– if id received, forward if received id greater than own id, else ignore

– last process in the chain sends a reply to originator if its id less than received id

– replies are always forwarded– A process goes to (k+1)th phase

only if it receives a reply from both sides in kth phase

– process receiving its own id – declare itself “leader”

• Message Complexity: O(nlgn)

• Lots of other algorithms exist for rings

• Lower Bound Result:– Any comparison-based leader

election algorithm in a ring requires Ώ(nlgn) messages

– What if not comparison-based?

Leader Election in Arbitrary Networks

• FloodMax – synchronous, round-based– at each round, each process sends

the max. id seen so far (not necessarily its own) to all its neighbors

– after diameter no. of rounds, if max. id seen = own id, declares itself leader

– Complexity = O(d.m), where d = diameter of the network, m = no. of edges

– does not extend to asynchronous model trivially

• Variations of building different types of spanning trees with no pre-specified roots. Chosen root at the end is the leader (Ex., the DFS spanning tree algorithm we covered earlier)

Clock Synchronization

Clock Synchronization

• Multiple machines with physical clocks. How can we keep them more or less synchronized?

• Internal vs. External synchronization

• Perfect synchronization not possible because of communication delays

• Even synchronization within a bound can not be guaranteed with certainty because of unpredictability of communication delays.

• But still useful !! Ex. – Kerberos, GPS

How clocks work

• Computer clocks are crystals that oscillate at a certain frequency

• Every H oscillations, the timer chip interrupts once (clock tick). No. of interrupts per second is typically 18.2, 50, 60, 100; can be higher, settable in some cases

• The interrupt handler increments a counter that keeps track of no. of ticks from a reference in the past (epoch)

• Knowing no. of ticks per second, we can calculate year, month, day, time of day etc.

Clock Drift

• Unfortunately, period of crystal oscillation varies slightly

• If it oscillates faster, more ticks per real second, so clock runs faster; similar for slower clocks

• For machine p, when correct reference time is t, let machine clock show time as C = Cp(t)

• Ideally, Cp(t) = t for all p, t• In practice,

1 – ρ ≤ dC/dt ≤ 1 + ρ• ρ = max. clock drift rate, usually

around 10-5 for cheap oscillators• Drift => Skew between clocks

(difference in clock values of two machines)

Resynchronization

• Periodic resynchronization needed to offset skew

• If two clocks are drifting in opposite directions, max. skew after time t is 2 ρ t

• If application requires that clock skew < δ, then resynchronization period

r < δ /(2 ρ)

• Usually ρ and δ are known

Cristian’s Algorithm

• One m/c acts as the time server• Each m/c sends a message

periodically (within resync. period r) asking for current time

• Time server replies with its time• Sender sets its clock to the reply• Problems:

– message delay– time server time is less than

sender’s current time

• Handling message delay: try to estimate the time the message with the timer server’s time took to each the sender– measure round trip time and

halve it– make multiple measurements of

round trip time, discard too high values, take average of rest

– make multiple measurements and take minimum

– use knowledge of processing time at server if known

• Handling fast clocks– do not set clock backwards; slow

it down over a period of time to bring in tune with server’s clock

Berkeley Algorithm

• Centralized as in Cristian’s, but the time server is active

• time server asks for time of other m/cs at periodic intervals

• time server averages the times and sends the new time to m/cs

• M/cs sets their time (advances immediately or slows down slowly) to the new time

• Estimation of transmission delay as before

External Synchronization

• Clocks must be synchronized with real time

• Cristian’s algorithm can be used if the time server is synchronized with real time somehow

• Berkeley algorithm cannot be used

• But what is “real time” anyway?

Measurement of time

• Astronomical– traditionally used– based on earth’s rotation around

its axis and around the sun– solar day : interval between two

consecutive transits of the sun– solar second : 1/86,400 of a

solar day– period of earth’s rotation varies,

so solar second is not stable– mean solar second : average

length of large no of solar days, then divide by 86,400

• Atomic– based on the transitions of

Cesium 133 atom– 1 sec. = time for 9,192,631,770

transitions– about 50+ labs maintain Cesium

clock– International Atomic Time (TAI) :

mean no. of ticks of the clocks since Jan 1, 1958

– highly stable– But slightly off-sync with mean

solar day (since solar day is getting longer)

– A leap second inserted approx. occasionally to bring it in sync. (so far 32, all positive)

– Resulting clock is called UTC – Universal Coordinated Time

• UTC time is broadcast from different sources around the world, ex.– National Institute of Standards &

Technology (NIST) – runs radio stations, most famous being WWV, anyone with a proper receiver can tune in

– United States Naval Observatory (USNO) – supplies time to all defense sources, among others

– National Physical Laboratory in UK

– GPS satellites– Many others

NTP : Network Time Protocol

• Protocol for time sync. in the internet

• Hierarchical architecture– primary time servers (stratum 1)

synchronize to national time standards via radio, satelite etc.

– secondary servers and clients (stratum 2, 3,…) synchronize to primary servers in a hierrachical manner (stratum 2 servers sync. with stratum 1, startum 3 with stratum 2 etc.).

• Reliability ensured by redundant servers

• Communication by multicast (usually within LAN servers), symmetric (usually within multiple geographically close servers), or client server (to higher stratum servers)

• Complex algorithms to combine and filter times

• Sync. possible to within tens of milliseconds for most machines

• But, just a best-effort service, no guarantees

• RFC 1305 and www.eecis.udel.edu/~ntp/ for more details

Termination Detection

Termination Detection

Model– processes can be active or idle– only active processes send

messages– idle process can become active

on receiving an computation message

– active process can become idle at any time

– termination: all processes are idle and no computation message are in transit

– Can use global snapshot to detect termination also

Huang’s Algorithm

• One controlling agent, has weight 1 initially

• All other processes are idle initially and has weight 0

• Computation starts when controlling agent sends a computation message to a process

• An idle process becomes active on receiving a computation message

• B(DW) – computation message with weight DW. Can be sent only by the controlling agent or an active process

• C(DW) – control message with weight DW, sent by active processes to controlling agent when they are about to become idle

Let current weight at process = W

3. Send of B(DW):• Find W1, W2 such that W1 > 0, W2

> 0, W1 + W2 = W• Set W = W1 and send B(W2)

4. Receive of B(DW):• W += DW; • if idle, become active

5. Send of C(DW):• send C(W) to controlling agent• Become idle

6. Receive of C(DW):• W += DW• if W = 1, declare “termination”

Building Spanning Trees

Building Spanning Trees

Applications:• Broadcast• Convergecast

• Leader election

Two variations:• from a given root r• root is not given a-priori

Flooding Algorithm– starts from a given root r– r initiates by sending message M

to all neighbours, sets its own parent to nil

– For all other nodes, on receiving M from i for the first time, set parent to i and send M to all neighbors except i. Ignore any M received after that

– Tree built is an arbitrary spanning tree

– Message complexity = 2m – (n -1) where m = no of

edges– Time complexity ??

Constructing a DFS tree with given root

• plain parallelization of the sequential algorithm by introducing synchronization

• each node i has a set unexplored, initially contains all neighbors of i

• A node i (initiated by the root) considers nodes in unexplored one by one, sending a neighbor j a message M and then waiting for a response (parent or reject) before considering the next node in unexplored

• if j has already received M from some other node, j sends a reject to i

• else, j sets i as its parent, and considers nodes in its unexplored set one by one

• j will send a parent message to i only when it has considered all nodes in its unexplored set

• i then considers the next node in its unexplored set

• Algorithm terminates when root has received parent or reject message from all its neighbours

• Worst case no. of messages = 4m• Time complexity O(m)

What if no root given?

Main idea:

– Nodes have unique ids– A node starts building the DFS

tree with itself as root (a single node fragment) spontaneously as in the previous case

– Fragments of the spanning tree gets built in parallel, all nodes in each fragment is identified by the id of its root

– M carries the fragment id of the sender

– when M sent from node in lower id fragment to node in higher id fragment, lower id fragment is stalled by higher id fragment by not giving a response

– When M sent from higher to lower id fragment, node in lower id fragment switches parent to node in higher id tree, resets unexplored, and starts DFS again

– Eventually, the highest id node becomes the root (leader election!!)

– Message complexity O(mn) !!– Time complexity O(m)

What about MSTs??

Gallager-Humblet-Spira Algorithm

– much more complex! but similar to Kruskal’s

– no root given, edge weights assumed to be distinct

– MST built up in fragments (subtree of MST)

– initially each node in its own fragment

– fragments merge, finally just one fragment

– outgoing edge – edge that goes between two fragments

– known result – min. wt. outgoing edge of a fragment always in MST

Issues:

1. How does a node find its min. wt. outgoing edge?

2. How does a fragment finds its min. wt. outgoing edge?

3. When does two fragments merge?

4. How does two fragments merge?

Some definitions

• Each node has three states – Sleeping – initial state – Find – currently finding the

fragment’s min. wt. outgoing edge

– Found – found the min. wt. outgoing edge

• Each fragment has a level– initially, each node is in a

fragment of level 0

Merging rule for fragments

• Suppose F is a fragment with id X, level L, and min. wt. outgoing edge e. Let fragment at other end of e be F1, with id X1 and level L1. Then

– if L < L1, F merges into F1, new fragment has id X1, level L1

– if L=L1, and e is also the min. wt. outgoing edge for F1, then F and F1 merges; new fragment has id X2 = weight of e, and level L + 1; e is called the core edge

– otherwise, F waits until one of the above becomes true

How to find min. wt. outgoing edge of a fragment

• nodes on core edge broadcasts initiate message to all fragment nodes along fragment edges; contains level and id

• on receiving initiate, a node find its min. wt. outgoing edge (in Find state) – how?

• nodes send Report message with min. wt. edge up towards the core edge along fragment edges (and enters Found state)

• leafs send their min. wt. outgoing edge, intermediate nodes send the min. of their min. wt. outgoing edge and min. edge sent by children in fragment; path info to best edge kept

• when Report reaches the nodes on the core edge, min. wt. outgoing edge of the fragment is known.

What then???

• nodes on core edges send Change_core message to node i with min. wt. outgoing edge

• node i then sends a Connect message to node j at other end with its level

• If j’s fragment level is greater than i’s, initiate message sent from j to i. This updates level and id of all nodes in j’s old fragment; if j has not sent a Report message yet, nodes in i’s old fragment starts finding its min. wt. outgoing edge, else not.

• if j’s fragment level is less, no response sent and i just waits till j’s fragment id becomes higher

• if fragment id’s match and j sends Connect to i also, merge into a level L+1 fragment with new core edge and id, and send initiate message

• some more details skipped, read paper

• Algo. terminates when no outgoing edge found for a fragment

• Worst case message complexity = O(n log n + m)

Fault Tolerance

and

Recovery

Fault Tolerance & RecoveryClassification of faults:

– based on component that failed• program/process• processor/machine• link• storage• clock

– based on behavior of faulty component• Crash – just halts • Failstop – crash with

additional conditions • Omission – fails to perform

some steps• Byzantine – behaves

arbitrarily• Timing – violates timing

constraints

Types of tolerance:– Masking – system always

behaves as per specifications even in presence of faults

– Non-masking – system may violate specifications in presence of faults. Should at least behave in a well-defined manner

Fault tolerant system should specify:– Class of faults tolerated– what tolerance is given from

each class

Some building blocks (assumptions/primitives that help in designing fault tolerant systems):– Agreement (multiple processes

agree on some value)

– Clock synchronization– Stable storage (data accessible

after crash)– Reliable communication (point-

to-point, broadcast, multicast)– Atomic actions

Agreement Problems

Model:– total n processes, at most m of

which can be faulty– reliable communication medium– fully connected– receiver always knows the

identity of the sender of a message

– byzantine faults– synchronous system. In each

round, a process receives messages, performs computation, and sends messages.

Different problem variations

• Byzantine agreement (or Byzantine Generals problem)– one process x broadcasts a

value v• all nonfaulty processes must

agree on a common value (Agreement condition).

• The agreed upon value must be v if x is nonfaulty (Validity condition)

• Consensus– Each process broadcasts its

initial value• satisfy agreement condition• If initial value of all nonfaulty

processes is v, then the agreed upon value must be v

• Interactive Consistency– Each process i broadcasts its

own value vi

• all nonfaulty processes agree on a common vector (v1, v2,…,vn)

• If the ith process is nonfaulty, then the ith value in the vector agreed upon by nonfaulty processes must be vi

Solution to Byzantine agreement problem implies solution to other two

Byzantine Agreement Problem

• no solution possible if– asynchronous system, or – n < (3m + 1)

• needs at least (m+1) rounds of message exchange (lower bound result)

• “Oral” messages – messages can be forged/changed in any manner, but the receiver always knows the sender

Lamport-Shostak-Pease Algorithm

Recursively defined;

OM(m), m > 0• Source x broadcasts value to all

processes• Let vi = value received by process

i from source (0 if no value received). Process i acts as a new source and initiates OM(m -1), sending vi to remaining (n - 2) processes

• For each i, j, i ≠ j, let vj = value received by process i from process j in step 2 using O(m-1). Process i uses the value majority(v1, v2, …, vn -1)

OM(0)

2. Source x broadcasts value to all processes

3. Each process uses the value; if no value received, 0 is used

Time complexity = m+1 rounds

Message Complexity = O(nm)

You can reduce message complexity to polynomial by increasing time

Atomic Actions and Commit Protocols

An action may have multiple subactions executed by different processes at different nodes of a distributed system

Atomic action : either all subactions are done or none are done (all-or-nothing property/ global atomicity property) as far as system state is concerned

Commit protocols – protocols for enforcing global atomicity property

Two-Phase Commit• Assumes the presence of write-

ahead log at each process to recover from local crashes

• One process acts as coordinator

Phase 1:• coordinator sends

COMMIT_REQUEST to all processes

• waits for replies from all processes• on receiving a

COMMIT_REQUEST, a process, if the local transaction is successful, writes Undo/redo logs in stable storage, and sends an AGREED message to the coordinator. Otherwise, sends an ABORT

Phase 2:• If all processes reply AGREED,

coordinator writes COMMIT record into the log, then sends COMMIT to all processes. If at least one process has replied ABORT, coordinator sends ABORT to all. Coordinator then waits for ACK from all processes. If ACK is not received within timeout period, resend. If all ACKs are received, coordinator writes COMPLETE to log

• On receiving a COMMIT, a process releases all resources/locks, and sends an ACK to coordinator

• On receiving an ABORT, a process undoes the transaction using Undo log, releases all resources/locks, and sends an ACK

• Ensures global atomicity; either all processes commit or all of them aborts

• Resilient to crash failures (see text for different scenarios of failure)

• Blocking protocol – crash of coordinator can block all processes

• Non-blocking protocols possible; ex., Three-Phase Commit protocol; we will not discuss in this class

Checkpointing & Rollback Recovery

Error recovery:• Forward error recovery – assess

damage due to faults exactly and repair the erroneous part of the system state– less overhead but hard to assess

effect of faults exactly in general

• Backward error recovery – on a fault, restore system state to a previous error-free state and restart from there – costlier, but more general,

application-independent technique

Checkpoint and Rollback Recovery – a form of backward error recovery

Checkpoint : – local checkpoint – local state of a

process saved in stable storage for possible rollback on a fault

– global checkpoint – collection of local checkpoints, one from each process

Consistent and Strongly Consistent Global Checkpoint – similar to consistent and strongly consistent global state respectively (Also called “recovery line”)

Orphan message – a message whose receive is recorded in some local checkpoint of a global checkpoint but send is not recorded in any local checkpoint in that global checkpoint ( Note : A consistent global checkpoint cannot have an orphan message)

Lost message – a message whose send is recorded but receive is not in a global checkpoint

Is lost messages a problem??– not if unreliable channels

assumed (since messages can be lost anyway)

– if reliable channels assumed, need to handle this properly! Cannot lose messages !

– We will assume unreliable channels for simplicity

Performance measures for a checkpointing and recovery

algorithm

• during fault-free operation– checkpointing time– space for storing checkpoints

and messages (if needed)• in case of a fault

– recovery time (time to establish recovery line)

– extent of rollback (how far in the past did we roll back? how much computation is lost?)

– is output commit problem handled? (if an output was sent out before the fault, say cash dispensed at a teller m/c, it should not be resent after restarting after the fault)

Some parameters that affect performance

• Checkpoint interval (time between two successive checkpoints)

• Number of processes• Communication pattern of the

application• Fault frequency• Nature of stable storage

Classification of Checkpoint & Recovery Algorithms

• Asynchronous/Uncoordinated

– every process takes local checkpoint independently

– to recover from a fault in one process, all processes coordinate to find a consistent global checkpoint from their local checkpoints

– very low fault-free overhead, recovery overhead is high

– Domino effect possible (no consistent global checkpoint exist, so all processes have to restart from scratch)

– higher space requirements, as all local checkpoints need to be kept

– Good for systems where fault is rare and inter-process communication is not too high (less chance of domino effect)

• Synchronous/Coordinated

– all processes coordinate to take a consistent global checkpoint

– during recovery, every process just rolls back to its last local checkpoint independently

– low recovery overhead, but high checkpointing overhead

– no domino effect possible– low space requirement, since

only last checkpoint needs to be stored at each process

• Communication Induced

– Synchronize checkpointing with communication, since message send/receive is the fundamental cause of inconsistency in global checkpoint

– Ex. : take local checkpoint right after every send! Last local checkpoint at each process is always consistent. But too costly

– Many variations are there, more efficient than the above, we will not discuss them in this class.

• Message logging

– Take coordinated or uncoordinated checkpoint, and then log (in stable storage) all messages received since the last checkpoint

– On recovery, only the recovering process goes back to its last checkpoint, and then replays messages from the log appropriately until it reaches the state right before the fault

– Only class that can handle output commit problem!

– Details too complex to discuss in this class

Some Checkpointing Algorithms

Asynchronous/Uncoordinated– See Juang-Venkatesan’s

algorithm in text, quite well-explained

Synchronous/Coordinated– Chandy-Lamport’s global state

collection algorithm can be modified to handle recovery from faults

– See Koo-Toueg’s algorithm in text, quite well-explained