Quality of Service Karrie Karahalios Spring 2007.

Quality of Service

Karrie KarahaliosSpring 2007

Announcements

• Midterm

• Final Projects

Distributed Multimedia System

Disk

Network CPUNetwork

Application

Server Client

Process

Device

Steps:

Resources:

Guarantees:

Consider End-to-End Behavior

NetworkMemoryDisk CPU

Apps

Operating System

Network MemoryDisk CPU

Operating System

MM apps Apps MM apps

Meta-scheduler attempts to reserve resourcesto guarantee end-to-end behavior

Basic Challenges

• Identification

• Negotiation

• Translation

• Specification

• Enforcement

Preliminaries

• Real-time systems

• Resources

• Management

• Reservation

• Allocation

Real Time (RT) Process

• Delivers results in predictable time– not necessarily fast– requests deterministic or stochastic

• Correctness means– errorless computation– meeting deadlines

Deadlines

• Hard deadlines– cannot be violated (if so, system fault)– cost money or human life

• Soft deadlines– misses can be tolerated if– not too many and not missed by much

• Examples

Characteristics of RT Systems

• Predictable response times to time-critical events

• Accurate system clocks and timing

• Able to schedule almost all resources

• Stability under overload

RT & Multimedia

• To meet demands of multimedia, use RT techniques along entire data path

• Relative to RT– relaxes deadline requirements and allows

some deadlines to be missed entirely– periodic requests ease scheduling– continuous data allows adaptive allocation

Resource

• Required by tasks for manipulating data– CPU, disk, memory, network, etc.

• A resource has a capacity– space, utilization, bandwidth

• A resource can be:– active or passive– exclusive or shared

Resource Management

• Maps multimedia requirements onto respective capacities of the system

• Specified through a QOS model– parameters + relaxation procedures

• Carried out by a resource manager– ensure adherence to QoS specification

Resource Management

Time

Audio

Mpeg-1

Mpeg-2

InteractiveVideo

Abundant

Sufficient, b

ut stre

ssed

Insuff

icien

t

Res

ou

rce

Nee

ds

Resource Reservation

• Test schedulability– determine if enough remaining capacity

• Negotiate QOS parameters– If not, determine how close it can come and when – application decides if this is acceptable

• Reserve resources– allocates resources to meet negotiated QoS

• Schedule resources– compute appropriate schedule for each resource– algorithm affects previous steps

Resource Allocation

• Pessimistic– reserve for the worst case– worse utilization, but more guarantees

• Optimistic– reserve for average or minimum needs– better utilization, but less guarantees

Quality of Service (QoS)

• Refers to how good provided services are– the more applications demand, the more

difficult it is to meet those demands

• Resource management realizes QoS– better management allow better QoS

• Examples

Layers of QoS

User

Application

System

MM devices Network

user QoS

application QoS

system QoS

network QoSdevice QoS

Layers of QoS

What are the right metrics?

How to specify them?

How to negotiate them?

User

Application

System

MM devices Network

user QoS

application QoS

system QoS

network QoSdevice QoS

How to enforce them?

How to translate among them?

User and Application QoS

• Startup time

• Sample rate

• Bits per sample

• Frame rate

• Resolution

• Skew relationships

• Response time for interaction

System QoS

Quantitative

• Bit rate

• Error rate

• Processing time

• Buffer sizes

• Throughput

Qualitative

• Ordered delivery

• Error recovery

• Scheduling options

Network QoS

• Network load – [min, avg, max] arrival times

• Packet/cell size

• Packet loss rate

• End-to-end delay (latency)

• Variability in delay (jitter)

Types of Services

• Guaranteed– threshold or range– deterministic or statistical

• Predictive– match current to historical performance

• Best Effort– none or only minimal guarantees

Linear Bounded Arrival Process

• Divides end-to-end system view into a pipeline of discrete sessions– one session corresponds to a single resource

• Defines parameterization of workload– arrival of messages at a particular interface

• A message is one unit of work– typically blocks of CM data (bytes or time)

Anderson, D. Metascheduling for Continuous Media, ACM TOCS, 11(3): 226-252

LBAP

• Models message arrival at a resource (I)– M - max message size (bytes)– R - max message rate (messages/second)– W – workload limit (max. messages that may

arrive ahead of schedule)

• Such that for all t0 < t1

NI(t0, t1) < R|t1 – t0| + W

Workload

• Workload W(t) of an LBAP at time t isw(t) = max{0, NI(t0, t) - R|t – t0| }

• Property 1:w(t) < W for all t

• Property 2:for all t1 < t2; NI(t1, t2) < w(t2) – w(t1) + R|t2 – t1|

Graph of w(t)

Logical Arrival Time (L)

• Let m0…mn denote sequence of messages and let a0…an be their arrival times

L(m0) = a0

L(mi+1)=max{ai+1, L(mi) + 1/R}

Logical Delay Between Interfaces

• Logical delay d(m) of message m between two interfaces is

d(m) = L2(m) - L1(m), where Li(m) isarrival time of m at interface i.

• Actual delay of message m may be> L(m), if m arrives ahead of schedule at I1

< L(m), if m completed ahead of schedule at I2

Resources and Sessions

• A resource handles incoming messages– arrive at input interface, delivered to output interface

• Clients must reserve resource– M (max message size)– R (max message rate)– Win = input max message burst (messages)– Wout = output max message burst (messages)– D = max logical delay (seconds),– A = min actual delay (seconds),– U = min unbuffered time (seconds)

• Arrival process at input interface specified by M, R, Win• Arrival process at output interface specified by M, R, Wout

Compound Session

• S is a sequence of sessions S1…SN

– input of S is that of S1; output is that of SN

– output of Si is input to Si+1

Delay and Buffer Size

• d(m) = LSN(m) – LS1

(m) < sum(di)

• Maximum shared buffer size for SW + R(D – U), D = sum(di); W = w(t) for S1

– realized when there is group of W messages followed by one message every 1/R and each resource uses its full delay for each message

QoS Specification

• Request S with given workload parameters and smallest delay bound

• If system can accept session– return minimum delay, reserves resources

• Otherwise– increase delay bound based on cost function– repeat request for S until allocated (or not)

Take Home Exercise

• Discuss tradeoffs of allowing people to pay for different qualities of service