17 - university of california, berkeley

Copyright 1999, University of California 4-Aug-99

1

Understanding Networked Applications:A First Course

Chapter 17

by

David G. Messerschmitt

Understanding Networked Applications A First Course2

Summary

• Performance and quality

• Concurrency

• Scalability

• Operating systems


2


What is performance

• Characterizing non-functional applicationattributes

• Quantitative metrics– Latency, speed, etc

– Parameterized by load, users, etc

• Performance of application as a wholeusually not meaningful– Subdivide into tasks


Tasks

• Collection of actions that are not meaningfullyseparated– Only completed task is of interest; partial results not

useful

– How does this relate to atomicity in transactions?

• For repetitive tasks, users or operators often careabout– Task completion time (sec)

– Task throughput (tasks/sec)


3


Important questions

• For any application

• What are the tasks?– Which is important: completion time,

throughput, or both, or something else?

– What are the quantitative objectives?


Example: Web surfing

• Task:– User: Click on URL, retrieve page, display

– Server: HTTP request received, HTML pageformed as a message, message transmitted

• What performance metric(s) are importantto:– User?

– Server?


4


Examples

• What are the tasks and their appropriateperformance metrics for the followingapplications:– Schwab stock trading system?

– Stock ticker?

– ATM system?

– Travel reservation system?

– Remote conferencing?


Typical components of taskcompletion time

Client Server

Formulaterequest

Message latency

Message latency

Processing time toformulate response

Interpretresponse


5


Time

Large completion timeand high throughput

Small completion timeand low throughput

Concurrency

Completion time and throughputare unrelated


Concurrency

• Two tasks are concurrent if they overlap intime

• Four cases:

Time


6


Concurrency may be….• …a way to satisfy intrinsic application

requrements– e.g. a number of users may generate concurrent tasks

• …a way to improve fairness

• …a way to increase task throughput– e.g. assign tasks to different hosts to get more done

• …a way to reduce task completion time– e.g. divide a task into subtasks and assign those

subtasks to different hosts


Time

Tasks not concurrentScheduling delay

Tasks concurrent

Short task completes sooner

Concurrency for fairness


7


Time

Tasks not concurrentScheduling delay

Tasks concurrent

Even though each task takes longer,short task still completes sooner

Concurrency may contribute tofairness on a single host


User view Reality

Time slice

Context switch

Concurrency on a single host

Why do this?

What is effect oncompletion time ofeach task?


8


Messages arrive at a communication linksimultaneously

Transmitting one message at a time

Transmitting messages concurrently

Message arrivesearlier

Divide each message into packets

Why networks use packets


Transaction_1

Transaction_2

Shared resources

Transaction_1Transaction_2

Transaction_1 Transaction_2

Two concurrenttransactions

Give the sameresult as:

Or:

Isolation of transactions


9


Queue

Concurrency

Messages among active objects

An active objectdelegates a task toanother active object,which retrieves fromqueue when ready

Task

Task


Scalability

by



10


Definition

• An application is scalable if its performanceparameters can be improved as necessary– Adding but not replacing equipment is

acceptable

– Re-configuration but not re-programming isacceptable

– Normally equipment cost should increase nomore than linearly with performance


Basic technique

• Concurrent tasks are a key to scalability

• Some obstacles:– Concurrent tasks aren’t there

– Dependency among tasks

– Communication overhead

– Variations in task loads result in congestion


11


Some problems to try to avoid

• Blocking– Processing blocked waiting for external events

– Make sure processor can work on somethingelse

• Duplication– In creating concurrent tasks, duplication of

effort is created


Problems to avoid (con’t)

• Load balancing– Workload not equal among processors

– Some processors not fully utilized

– More difficult if concurrent tasks are dependent

• Congestion– Workload fluctuates

– Makes load balancing more difficult becauseutilization can’t be predicted


12


Mux&

queue

Irregulararrival oftasks

Irregulartaskcompletiontimes

Server

Source of congestion


Congestion model

Queue

1. Tasks arrive atrandom times 2. Tasks wait in

queue waitingfor processing

3. Taskservice timeis random


13


Statistical waiting timeWaiting time

0

5

10

15

20

25

30

0.1 0.3 0.5 0.7 0.9

Utilization

Series1

A specific statisticalmodel for arrivaltimes and servicetimes is assumed

Tradeoff between taskcompletion time (waitingtime) and task throughput(utilization)


Example: Web server

How do users know the server is reachingfull utilization?

When a single Web server/processor isexhausted, what do we do?


14


Operating systems


Some OS goals

• Abstraction: hide hardware details

• Manage concurrency– Multitasking (time-slicing) concurrency

– Process and thread

– Inter-process communicaton

• Manage resources– Memory, storage, processing, network access


15


Processor cache

Main memory

Online (non-volatile) external storage

Archive storage

Off-line

Size

Speed

Memory/storage hierarchy


Processor

Sharedmemory

Processor Processor

Networkinterface

Processor

Privatememory

Processor ProcessorNetworkinterface

Privatememory

Privatememory

High-speed communication bus


16


OS kernel

Network Storage Memory

Process

Mem

ory

Process

Mem

ory

Userlevel

Kernellevel

Resources


Process

Kernel performsservice, then returns

KernelKernelmodemode

UserUsermodemode

TimeTime

write_file(),send_message(),etc.


17


Thread n Thread n+1Request toHTTP serverto returnhomepage

Page returned

Another request

Sharedrepositoryof WWWlinks

Another return

Threads


Supplements

By



18


Execution models for objects

• Active object– Independent center of activity

– Works without methods being called

• Passive object– Works only in response to method invocations


Two observations

• For anything to get done, there has to be at leastone active object!

• Natural for each active object to work on one taskat a time– If so there must be at least two active objects for

concurrency

– (although in principle an active object could time sliceamong tasks, this is much more complicated and not agood idea)


19


Active object Passive objects

Executing

Time

One object is executingat a time

Natural for each active object towork on one task

Task


Activeobject

ActiveobjectPassive objects

Concurrency with two activeobjects


20


Concurrency

Activeobject

ActiveobjectPassive objects

Potential resourceconflict

A possible problem


Stage1 Stage2 Stage3

Partially completedpiece of work

Task

Without pipelining


21


Stage1 Stage2 Stage3 Task_completeTask_start

Repetitivetasks

With pipelining

17 - university of california, berkeley

Documents