www.cs.man.ac.uk/cnc

Performance Control in RealityGrid

Ken Mayes, Mikel Luján, Graham Riley,Rupert Ford, Len Freeman, John Gurd

!st RealityGrid Workshop, June 2003

Overview

Preamble• The case for Performance Control

Context• Malleable, component-based Grid

applications

RealityGrid Application• The LB3D Use Case

RealityGrid Performance Control System• Design and status report

Generalisation . . .

“Performance Analysis and Distributed Computing” . . .

We want specific computational performance and resource levels

We want these in an environment that is• Distributed (physically)• Heterogeneous (inherently)• Dynamic (by necessity)

We want them with a minimum effort

. . . Implies Performanceand Resource Control

There is no viable alternative apart from automating the process of achieving them• monitoring and analysis (sensing)• change when unsatisfactory (actuation)

This is a classical control system scenario, and we need to use established control system techniques to deal with it

Control Systems Overview

Control systems involve appropriate change of actuated variables informed by (timely) feedback of monitoring information:

actuator

feedback function

error

Summary

Traditional performance “control” amounts to “open loop” control, mediated by human “performance engineers”

Distributed environments introduce enough extra complexity that “closed loop” control becomes essential

Successful closed loop control demands accurate and rapid feedback data, thus affecting achievable control limits

Context

Grid applications will be distributed and, in some sense, component-based.

To deliver the power grid model, adaptation is key!• Elevate users above Grid resource and performance

details.

Our work is considering adaptation and its impact on performance:• Adaptation in component deployment,

- Initial model configuration and deployment on resources.- Flexibility in deployment of compositions of coupled

models.

• Adaptation via malleable components.- At run time, re-allocation of resources in response to

changes.

RealityGrid - LB3D Use Case

Display simulation time equal

LB3D

LB3D

Output rate 1Resolution 1

Output rate 2Resolution 2

Tracking a parameter search

UserChanges

dynamically

T=100

T=100

Params 1

Params 2

LB3D Use Case

User Display rates ~equal

LB3D

LB3D

Output rate 1Resolution 1

Output rate 2Resolution 2

Tracking a parameter study

Params 1

Params 2

“Malleable” LB3D - mechanisms

LB3D will respond to requests to change resources• Use more (or less) pes (& mem) on current system• Move from one system to another

- Via machine independent (parallel) checkpoint/restart

LB3D will output science data (for remote visualisation) at higher or lower rates

LB3D will (one day) respond to requests to continue running at higher (or lower) lattice resolution

Each of the above affects performance (e.g. timesteps per second rate)

Each has an associated cost . . .

Use Case - continued

User might be tracking many parameter set developments (one per LB3D instance)

Some will be uninteresting (for a while)• Lower output rate / resolution / terminate

Some will become interesting• Increase output rate / resolution

One aim: Re-distribute resources across all LB3D instances to maintain highest possible timestep rate

A General Grid Application

GenerateData

ProcessData

VisualiseData

Component 1 Component 2 Component 3

Computational Grid Resources

Applications and components exhibit phased behaviour

Life is Hierarchical

Can we use hierarchy to “divide and conquer” complex system problems?

Introduce “Performance Steerers” at component- and application-levels . . .

Performance Steerers - Design

CPS CPS CPS

Component 1 Component 2 Component 3

Computational Grid Resources

APS

Component Framework

ExternalResourceScheduler

Initial deployment+

Run-time adaptation

Full System

CPS

Component 2

ExternalResourceScheduler

APS

Loader

Predictor

Predictor

ComponentPerformanceRepository

ApplicationPerformanceRepository

ResourceUsage

Monitor

Performance Prediction

Role of APS• To distribute available resources to the

components in such a way that the predicted performance of components gives a minimum predicted execution time.

Role of CPS• To utilise allocated resources and

component performance effectors (actuators) so that the predicted execution time of the steered component is minimised.

Life is Repetitive

Many programs iterate more-or-less the same thing over-and-over again

We can take advantage of this• e.g. for load balance in weather prediction• and, possibly, for performance control

Application Progress Points

Assume that the execution of the application proceeds through phases. Phase boundaries are marked by Progress Points.

NB. Can take decisions about performance and take actions at the progress points:• Must be safe points

0 1 2 3 4 5 6 7

Ph 1 Ph 2 Ph 3 Ph 4 Ph 5 Ph 6 Ph 7

Component Progress Points

Information about progress points will be contained in some repository.

Application progress points

Component progress points

APS

CPS

Component

Time

Implementation:

APS

CPS

Comms I/f

Comms I/f

LB3DComponent

Interface

APS as daemon, CPS as library

CPS Progress

Component control

Procedurecalls

Sockets(rpc)

Implementation:

APS

Machine 1

CPS LB3D

MPIRUN

Component Loader

DUROC + RSL + GLOBUS + GRIDFTP

Machine 3

Component Loader

Machine 2

shme

m

socket

socket

shme

m

Start-up Process

1. GlobusRun, RSL script for Component loaders (one per machine in Grid) plus APS daemon.

2. Loaders connect to each other.3. LB3D started by Loader (via MPIRUN), calls

CPS (a library) at start-up.4. CPS connects to APS.5. LB3D calls CPS at each subsequent progress

point and CPS communicates with APS.6. Continue until LB3D completed (e.g. no. of

timesteps complete).

Status We have a prototype implementation (basic

mechanisms)First Experiment:• Every N tsteps, move LB3D between machines 2 and

3 – determined by APS.• At tstep mod N progress point in LB3D, APS tells CPS

(which tells the component) to checkpoint, CPS writes certain status information to the shmem area and then LB3D (and CPS) dies.

• Loader on machine it ran on communicates to loader on machine it is to run on. The restart file is Gridftp’d along with restart info. (e.g. tstep to shmem area of new loader).

• New LB3D is started and CPS manages the restart.• Continue until no more tsteps.

Status Preliminary performance results

np 4, data 64x64x64checkpoint file (XDR) size 32.8 MBaverage resident tstep time for cronus 3.384 (s.)average migration tstep time to cronus 43.718 (s.)average resident tstep time for centaur3 6.675 (s.)average migration tstep time to centaur3 55.280

(s.)

np 4, data 8x8x8checkpoint file (XDR) size 64 KBaverage resident tstep time for cronus 0.017 (s.)average migration tstep time to cronus 0.504 (s.)average resident tstep time for centaur3 0.061 (s.)average migration tstep time to centaur3 3.038 (s.)

cronus = SGI Origin 3400 and centaur3 = Sun TCF

Status We have a prototype implementation

Second Experiment:• Three platforms (2 SGI, 1 Solaris – Linux

coming soon)• Move LB3D at random between machines,

as determined by APS.• Has exposed Gridftp problems – worked

around.• Have timings for 2 machines, but not yet 3

• Now looking at ICENI framework.

Status

Developing implementation of performance repository• Berkeley Database (linked as library)

Survey of prediction and machine learning algorithms • runtime vs. off line analysis• accuracy of predictions• amount of history data required

Learning control theory and understand how to apply it to Performance Control.

Generalisation

LB3D as a component is an “easy” case• how does the above scheme generalise?

Are components necessary? Is hierarchy necessary?

• component < application < scheduler

What different kinds of component and/or application might there be?• redefinition of the “process”• a topic for my forthcoming leave of absence

Summary Aim:

Develop an architecture that enables us to investigate different mechanisms for Performance Control of malleable component based applications on the Grid

Main characteristic: • dynamic adaptation

Design/implementation tensions:• general vs. specific purpose• APS <-> CPS communication• APS <-> CPS ratio• performance prediction algorithms - accuracy vs.

execution time• APS/CPS overhead vs. application execution time

Work in development (first year in one sentence): • A Grid Framework for Malleable Component-based

Application Migration

Related Projects at Manchester

Met Office – FLUME - design of next generation Unified Model software• Coupled models

Tyndall Centre – SoftIAM - Climate Impact, Integrated assessment modelling• Coupling climate and economic models

Computational Markets • 1 RA and 1 PhD - positions being filled at present• UK e-Science funded (Imperial College led)

For more information check http://www.cs.man.ac.uk/cnc

http://www.realitygrid.org