Kumar CSE@UTA 1

6306 Advanced Operating Systems

Instructor : Dr. Mohan KumarRoom : 315 NH

kumar@cse.uta.edu

Class : TTh 7- 8:20PM

Office Hours : TTh1-3 PM

GTA : Byung Sungsung@cse.uta.edu

Kumar CSE@UTA 2

Distributed SystemsR Chow, Distributed Operating Systems

Provide a conceptually simple model Efficient, flexible, consistent and robust

Goals Efficiency

Communication delays?? Scheduling, load balancing

Flexibility Users decide how, where, and when to use the system Environment for building additional tools and services System’s view - scalability, portability, and interoperability

Consistency Lack of global information, potential replication and partitioning

of data, component failures and interaction among modules Robustness

Kumar CSE@UTA 3

Transparency

Logical view of a Physical System Reduce the effect and awareness of the physical system

System-dependent information Tradeoff between simplicity and effectiveness

Access transparency –local and remote objects Location transparency –objects refs by logical names Migration transparency –location independence Concurrency transparency Replication transparency Parallelism, Failure, Performance, Size (modularity

and scalability), Revision (software)

Kumar CSE@UTA 4

Transparency

System Goals Transparencies

Efficiency Concurrency

Parallelism

Performance

Flexibility Access, Location, Migration, size, revision

Consistency Access

Replication

Performance

Robustness Failure

Replication

Size

Revision

Kumar CSE@UTA 5

Transparency

Major Issues TransparenciesCommunicationSynchronizationDistributed algorithms

Interaction and control transparency

Process schedulingDeadlock handlingLoad balancing

Performance transparency

Resource schedulingFile sharingConcurrency control

Resource transparency

Failure handlingConfigurationRedundancy

Failure transparency

Kumar CSE@UTA 6

Services

Primitive Message passing primitives – send and receive Synchronization Process management

Process creation, deletion, tracking, resource allocation

System servers Name server/directory server

Users,processes, machines,files,ports Network server

Path selection and routing Broadcast/multicast server Clock Synchronization

Kumar CSE@UTA 7

Architecture Models

Workstation-server model Processor-pool model Communication network protocols

Kumar CSE@UTA 8

Distributed Coordination

Barrier Synchronization Condition coordination Mutual Exclusion

Interprocess communication

Distributed Resources

Distributed Shared memory

Kumar CSE@UTA 9

Two Processor Model

Consider execution of application program that contains M tasks on a Two processor (or PE) system

We make the following assumptions to simplify analysis1. Each task executes in R units of time

all tasks are of equal complexity2. Each task communicates with every other task with anoverhead cost of C units of time when the communicating tasks are on different PEs and with zero cost when the communicating tasks are co-resident

3. Non-overlapped computation/communication

Indurkhya and H Stone, IEEE Software, 1997

Kumar CSE@UTA 10

Two PE model (Contd.)

(k) (M-k)

Kumar CSE@UTA 11PA252/502 Week 8

Two PE model (Contd.)

* Assign all tasks to one PE or* Partition tasks among the two PEsTotal execution time (TET) is the sum of the total computation time and the communication overheads

Let us assign (M-k) tasks to one PE and k tasks to the otherComputation time = R Maximum (M-k,k)Communication time = C (M-k)*kTotal Execution time = R * (Max (M-k,k) + C *(M-k)*k

Linear Quadratic

What is the maximum TET as a function of k?

Kumar CSE@UTA 12PA252/502 Week 8

Two PE model (Contd.)

0

20

40

60

80

100

0 10 20 25 30 40 50

k

R

comput.comm.

TET

R/C = 10

Kumar CSE@UTA 13

Two PE model (Contd.)

0

10

20

30

40

50

60

comput.comm.

TET

comput. 50 40 30 25 30 40 50

comm. 0 10 15 15.63 15 10 0

TET 50 50 45 40.63 45 50 50

0 10 20 25 30 40 50

R/C = 40

Minimum execution time when R/C = M/2

Kumar CSE@UTA 14

Two PE model (Contd.)

0

10

20

30

40

50

60

comput.comm.

TET

comput. 50 40 30 25 30 40 50

comm. 0 16 24 25 24 16 0

TET 50 56 54 50 54 56 50

0 10 20 25 30 40 50

R/C = 25

Kumar CSE@UTA 15

Two PE model (Contd.)

Total time on one processor = R* M units … (1)

R/C = M/2 is the critical point for 2 processor system

if R/C M/2 assign all tasks to one Processor

if R/C > M/2 divide tasks equally among all processors

Divide equally among the two processors

Then TET = R(M/2) + C(M-M/2)*M/2 … (2)

To get break even point equate (1) and (2)

R(M/2) + C(M-M/2)*M/2 = R*M,

Simplifying, we get R/C = M/2

Kumar CSE@UTA 16

Two PE model (Contd.)If R/C high then C is low, communication overheads are low,

we are justified in parallelising

If R/C is low then C is high, communication overheads are high

better to execute on one processor

Speedup = Serial time/Parallel time > 1 if and only if R/C >M/2 Check with the times

Kumar CSE@UTA 17

N PE model All processor sharea common communicationchannel

Assign ki tasks to Pei

k1 - PE1

k2 - PE2 …kN - PEN

k1+ k2+ . . . + kN= M

Computation time = R Max (ki)Communication time = C/2[k1*(M-k1)+k2*(M-k2)+ . . . +kN*(M-kN)

Kumar CSE@UTA 18

N PE model

Communication time = C/2[k1*(M-k1)+k2*(M-k2)+ . . . +kN*(M-kN)

= C/2 ki(M-ki) = C/2[ (M ki) - (ki

2)]

= C/2 [M2 - ki2]

3 3

Kumar CSE@UTA 19

N PE model

Total Execution time = R* Max(ki) + C/2 [M2 - ki2]

assuming all tasks to be of equal complexity,we assign M/N tasks per PE

Total execution time = R(M/N)+CM2/2(1-1/N) … (3)ki

2 = N(M/N)2 = M2/N; because, ki = M/N

Equate equation 3 with RM to get the break even point

R(M/N)+CM2/2(1-1/N) = RMRM(1-1/N) = CM2/2(1-1/N) R/C = M/2

if R/C M/2 assign all tasks to one Processor

if R/C > M/2 divide tasks equally among N PEs

Kumar CSE@UTA 20

N PEs with overlapped communication

The PEs can compute and communicate at the same timeTET = Max {computation time, communication time} = Max {R* Max(ki), C/2 [M2 - ki

2} assuming all tasks to be of equal complexity,

= Max {R(M/N), (CM2/2)*(1-1/N)}optimum performance when computation time = communication overheadR(M/N) = (CM2/2) * (1-1/N), ignoring 1/N if N is large,R(M/N) = CM2/2

R/C = MN/2; optimum number of PEs, N = 2R/MC

Kumar CSE@UTA 21

N PEs with multiple communication links

N Communication channels in the network

Total Execution time = R* Max(ki) + [C/(2N)]* [M2 - ki2]

assuming all tasks to be of equal complexity,

Total execution time = R(M/N)+CM2/2N(1-1/N)

NM

CR

NN

NC

NNRM

NN

NC

NM

RRM

2

)1

(2

)1

(

)1

(2

Kumar CSE@UTA 22

N PEs with multiple communication links

if R/C M/2N assign all tasks to one Processor

if R/C > M/2N divide tasks equally among N PEs

Speedup = TET on ONE PE / (TET on N PEs)

N

NNM

CR

CRN

NNCM

NRM

RM

2)1(

)1

1(2

2

=

If R/C >> M/2 then speedup increases with Nif R/C << M/2 then speedup is given by,

2NR/MC

Kumar CSE@UTA 23

Comments on Performance

Multiple number of PEs produce overhead costs

- scheduling- contention for shared resources- message passing- synchronisation- lack of load balancing

Overhead costs increase with the number of PEsR/C is a measure of the amount of program execution per

unit of overheadIf R/C is large, Communication overheads are low and the

problem execution is efficientIf R/C is small, Communication overheads are high …..

Kumar CSE@UTA 24

Active Middleware Services in a Decision Support System for Managing Highly AvailableDistributed ResourcesS A Fakhouri, W F Jerome, V K Naik, A Raina and P Varma

The cluster based system is called Mounties Manages resources and applications using rule based

constraints Mounties has four service components

Resource proxy objects Manipulates cluster configuration

Event notification mechanism Monitors and controls interdependent and distributed resources

Rule Evaluation and Decision processing mechanism Global optimization service

Provides decision making capabilities

Kumar CSE@UTA 25

What’s in the paper?

Architecture and design of service components Interaction between the components and the

decision making component General programming paradigm

Kumar CSE@UTA 26

Cluster?Nodes,disks,

adapters,databases etc.Heterogeneous

Components of a Cluster

Transparent

Modular

Scalable

Achieved by providing coordination and mapping of physical distributed resources onto a set of virtual resources and their services – very complex?

Kumar CSE@UTA 27

Authors’ approach

Clusters and resources – 2 dimensions Semistatic nature of resource

Type and quality of supporting services needed to enable its services

Formalized as simple rules Dynamic state of the services provided by the

cluster – captured by events Coordination and mapping – centralized and

rule-based Events and rules are combined only when

necessary

Kumar CSE@UTA 28

Resources

Attributes Unique Name Type – functionality Capacity – number of dependent sources it

can serve Priority (L1 .. 10H) State

ONLINE OFFLINE FAILED

Kumar CSE@UTA 29

Mounties Approach

Constraint-based methodology cluster configuration, startup and recovery

Constraints are used build relationships among supporting and dependent resources/services

Nominal State – ONLINE, OFFLINE Inter-resource relationships

DependsOn CollacatedWith Anti-CollocatedWith Equivalency

Set of resources with similar functionality– Choose one of these

Kumar CSE@UTA 30

Mounties..

Distinct Decision making, Resource allocation Monitor, control, manipulate

Example from the paper E1: DA0,DA1; E2: NA0,NA1, NA2

Database 1

DO E1, E2

CW E1,E2

Database 1

E1: DA0

E2 : NA0

Node 0

Database 2

E1: DA1

E2 : NA1

Node 1

Kumar CSE@UTA 31

Mounties..

Webserver

D : E2,E3

CW : E2

E2: NA2

E3: DB1

Node 2

Kumar CSE@UTA 32

Events and Clusters

Resource related Response to a command Enduser interactions/directives Alerts/alarms

Resource Specific Constraints

capacity, dependency, location,

Kumar CSE@UTA 33

To subscribe to LISTSERV

Send a message to

To: listserv@listserv.uta.edu

Subject: blank (no subject heading)

--------------------------------------

subscribe cse6306-l <student’s name>

OR go to

http://listserv.uta.edu/cgi-bin/wa.exe?SUBED1=cse6306-l&A=1

and fill out the form

Kumar CSE@UTA 34