Performance-responsive Scheduling for Grid Computing

Performance-responsive Scheduling for Grid Computing

Dr Stephen JarvisHigh Performance Systems Group

University of Warwick, UK

High Performance Systems Group

Context

• Funded by / collaborating with – UK e-Science Core Programme– IBM (Watson, Hursley)– NASA (Ames)– NEC Europe– Los Alamos National Laboratory

• Integrate established performance tools into emerging grid middleware

High Performance Systems Group

What do we mean by ‘scheduling’• Users view

– Jobs run somewhere on the Grid– Notion of deadline– Execution is single domain (includes pre-staging)

• Resource providers view– Don’t mind which jobs are run where– As long as resources are well/evenly used– Maintaining customers deadlines is important

• System view– Jobs can run anywhere– Resources are heterogeneous– Throughput is important, as are scheduling overheads

High Performance Systems Group

Managing through Middleware

High Performance Systems Group

•Determine what resources are required (predict)

•Determine what resources are available (discover)

•Map requirements to available resources (schedule)

•Maintain contract of performance (QoS)

Managing through Middleware

Performance Services

• Intra-domain– Lab- / department-based

– Shared resources under local administration

• Multi-domain– Campus- / country-based

– Wide-area resource and task management

– Cross domain

High Performance Systems Group

Performance Services

High Performance Systems Group

• Intra-domain– Lab- / department-based

– Shared resources under local administration

• Multi-domain– Campus- / country-based

– Wide-area resource and task management

– Cross domain

Performance Services

High Performance Systems Group

• Intra-domain– Lab- / department-based

– Shared resources under local administration

• Multi-domain– Campus- / country-based

– Wide-area resource and task management

– Cross domain

Performance Prediction

• Performance prediction tools• Aim to predict

– Execution time– Communication usage– Data and resource requirements

• Provides best guess as to how an application will execute on a given resource

High Performance Systems Group

High Performance Systems Group

PACE User

Application

Resource

High Performance Systems Group

PACE User

Application

Resource

ApplicationModel

Resource Model

Application

ApplicationModel

Resource

Resource Model

PACE User

Evaluation Engine

Model parameters

Resource config.

High Performance Systems Group

Application

ApplicationModel

Resource

Resource Model

PACE User

Evaluation Engine

Model parameters

Resource config.

High Performance Systems Group

Why is prediction useful?

• Scaling properties

• Compare runtime options with– deadline

– available resources

– priority / other jobs

– etc.

High Performance Systems Group

0

5

10

15

20

25

30

35

40

45

50

1 4 7 10 13 16

The Number of Processors

Run

ning

Tim

e on

SG

IOri

gin2

000

(sec

)

sweep3d

fft

improc

closure

jacobi

memsort

cpi

Allows runtime scenarios to be explored before deployment

Run

-tim

e

1. Intra-Domain Co-Scheduling

High Performance Systems Group

• Augment Condor scheduler with additional

performance information

• Scheduler driver, or co-scheduler (called

Titan)

• Use predictive data for system improvement

– Time to complete tasks / utilisation of resources

– QoS – ability to meet deadlines

• Handle predictive and non-predictive tasks

Intra-Domain Co-Scheduling

High Performance Systems Group

• Non-predictive tasks

PORTALPRE-

EXECUTIONENGINE MATCHMAKER

SCHEDULEQUEUE

PACE

GA CLUSTERCONNECTOR

CONDORCONDOR

REQUESTS FROM USERS OR OTHERDOMAIN SCHEDULERS

RESOURCES

CLASSADS

Titan

Intra-Domain Co-Scheduling

High Performance Systems Group

• Non-predictive tasks

PORTALPRE-

EXECUTIONENGINE MATCHMAKER

SCHEDULEQUEUE

PACE

GA CLUSTERCONNECTOR

CONDORCONDOR

REQUESTS FROM USERS OR OTHERDOMAIN SCHEDULERS

RESOURCES

CLASSADS

Titan

Intra-Domain Co-Scheduling

High Performance Systems Group

• Non-predictive tasks• Tasks with prediction

data

PORTALPRE-

EXECUTIONENGINE MATCHMAKER

SCHEDULEQUEUE

PACE

GA CLUSTERCONNECTOR

CONDORCONDOR

REQUESTS FROM USERS OR OTHERDOMAIN SCHEDULERS

RESOURCES

CLASSADS

Titan

Intra-Domain Co-Scheduling

High Performance Systems Group

• Non-predictive tasks• Tasks with prediction

data

PORTALPRE-

EXECUTIONENGINE MATCHMAKER

SCHEDULEQUEUE

PACE

GA CLUSTERCONNECTOR

CONDORCONDOR

REQUESTS FROM USERS OR OTHERDOMAIN SCHEDULERS

RESOURCES

CLASSADS

Titan

Intra-Domain Co-Scheduling

High Performance Systems Group

• Non-predictive tasks• Tasks with prediction

data

PORTALPRE-

EXECUTIONENGINE MATCHMAKER

SCHEDULEQUEUE

PACE

GA CLUSTERCONNECTOR

CONDORCONDOR

REQUESTS FROM USERS OR OTHERDOMAIN SCHEDULERS

RESOURCES

CLASSADS

Titan

Intra-Domain Co-Scheduling

High Performance Systems Group

• Non-predictive tasks• Tasks with prediction

data

PORTALPRE-

EXECUTIONENGINE MATCHMAKER

SCHEDULEQUEUE

PACE

GA CLUSTERCONNECTOR

CONDORCONDOR

REQUESTS FROM USERS OR OTHERDOMAIN SCHEDULERS

RESOURCES

CLASSADS

Titan

Intra-Domain Co-Scheduling

High Performance Systems Group

• Non-predictive tasks• Tasks with prediction

data

PORTALPRE-

EXECUTIONENGINE MATCHMAKER

SCHEDULEQUEUE

PACE

GA CLUSTERCONNECTOR

CONDORCONDOR

REQUESTS FROM USERS OR OTHERDOMAIN SCHEDULERS

RESOURCES

CLASSADS

Titan

Intra-Domain Deployment

Without co-scheduler With co-scheduler

Time to complete = 70.08m Time to complete = 35.19m

High Performance Systems Group

• Publish intra-domain perf. data through

global information services (MDS)

• Augment service with agent system

– One agent per domain / VO

• When a task is submitted

– Agents query IS, and negotiate to discover best

domain to run task

• Scheme is tested on a 256-node exp. Grid

– 16 resource domains; 6 arch. types

High Performance Systems Group

2. Multi-Domain Management

High Performance Systems Group

Multi-Domain Management

time

High Performance Systems Group

Multi-Domain Management

time

High Performance Systems Group

Multi-Domain Management

time

High Performance Systems Group

Multi-Domain Management

Time to complete = 2752s

Multi-Domain Management

High Performance Systems Group

Time to complete = 467s; an improvement of 83%

Multi-Domain Management

High Performance Systems Group

Time to complete = 467s; an improvement of 83%

QoS: Ability to Meet Deadline

High Performance Systems Group

active inactive

Resource usage

High Performance Systems Group

active inactive

Other work

• OGSA compatibility• Prediction

– Accuracy– Other prediction techniques

• Workflow (CCGrid 2003)• Reservation• V. 1.1, Condor/GT2-based

– www.dcs.warwick.ac.uk/~hpsg– Documented at HPDC-12/GGF-8, FGCS

High Performance Systems Group