The Ibis Project:Simplifying Grid Programming &

Deployment

Henri Balbal@cs.vu.nl

Vrije Universiteit Amsterdam

The ‘Promise of the Grid’

Efficient and transparent (i.e. easy-to-use) wall-socket computing over a distributed set of resources [Sunderam ICCS’2004, based on Foster/Kesselman]

Parallel computing on grids

● Mostly limited to● trivially parallel applications

● parameter sweeps, master/worker● applications that run on one cluster at a

time● use grid to schedule application on a suitable

cluster

● Our goal: run real parallel applications on a large-scale grid, using co-allocated resources

Efficient wide-area algorithms

● Latency-tolerant algorithms with asynchronous communication● Search algorithms (Awari-solver [CCGrid’08])● Model checkers (DiVinE [PDMC’08])

● Algorithms with hierarchical communication● Divide-and-conquer● Broadcast trees

● …..

● Performance & scalability

● Heterogeneous● Low-level & changing

programming interfaces

● writing & deploying grid applications is hard

Reality: ‘Problems of the Grid’

● Connectivity issues

● Fault tolerance

● Malleability

Wide-Area Grid Systems

UseUserr

!

The Ibis Project

● Goal:● drastically simplify grid

programming/deployment

● write and go!

Approach (1)

● Write & go: minimal assumptions about execution environment● Virtual Machines (Java) deal with

heterogeneity

● Use middleware-independent APIs● Mapped automatically onto middleware

● Different programming abstractions● Low-level message passing

● High-level divide-and-conquer

Approach (2)

● Designed to run in dynamic/hostile grid environment● Handle fault-tolerance and malleability● Solve connectivity problems automatically

(SmartSockets)

● Modular and flexible: can replace Ibis components by external ones● Scheduling: Zorilla P2P system or external

broker

Global picture

Rest of talk

Satin: divide & conquer

Communication layer (IPL)

SmartSockets

Applications

JavaGATZorilla P2P

Outline

● Grid programming● IPL● Satin● SmartSockets

● Grid deployment● JavaGAT● Zorilla

● Applications and experiments

Ibis Design

Ibis Portability Layer

(IPL)● Java-centric “run-anywhere” library

● Sent along with the application (jar-files)● Point-to-point, multicast, streaming, ….

● Efficient communication● Configured at startup, based on

capabilities (multicast, ordering, reliability, callbacks)

● Bytecode rewriter avoids serialization overhead

Serialization

● Based on bytecode-rewriting ● Adds (de)serialization code to serializable

types● Prevents reflection overhead

during runtimeJava

compilerbytecoderewriter JVM

JVM

JVM

source bytecodebytecode

Membership Model

● JEL (Join-Elect-Leave) model● Simple model for tracking resources,

supports malleability & fault-tolerance● Notifications of nodes joining or leaving● Elections

● Supports all common programming models● Centralized and distributed

implementations● Broadcast trees, gossiping

Programming models

● Remote Method Invocation (RMI)● Group Method Invocation (GMI)● MPJ (MPI Java 'standard')● Satin (Divide & Conquer)

Satin: divide-and-conquer

● Divide-and-conquer isinherently hierarchical

● More general thanmaster/worker

● Cilk-like primitives (spawn/sync) in Java● Supports malleability and fault-tolerance● Supports data-sharing between different

branches through Shared Objects

fib(1) fib(0) fib(0)

fib(0)

fib(4)

fib(1)

fib(2)

fib(3)

fib(3)

fib(5)

fib(1) fib(1)

fib(1)

fib(2) fib(2)cpu 2

cpu 1cpu 3

cpu 1

Satin implementation

● Load-balancing is done automatically● Cluster-aware Random Stealing (CRS)● Combines Cilk’s Random Stealing with

asynchronous wide-area steals

● Self-adaptive malleability and fault-tolerance● Add/remove machines on the fly● Survive crashes by efficient

recomputations/checkpointing

Self-adaptation with Satin

● Adapt #CPUs to level of parallelism● Migrate work from overloaded to idle CPUs● Remove CPUs with poor network connectivity● Add CPUs dynamically when

● Level of parallelism increases● CPUs were removed or crashed

● Can also remove/add entire clusters● E.g., for network problems

[Wrzesinska et al., PPoPP’07 ]

Approach

● Weighted Average Efficiency (WAE):

1/#CPUs * Σ speedi * (1 – overheadi )

overhead is fraction idle+communication time

speedi= relative speed of CPUi (measured periodically)

● General idea:

Keep WAE between Emin (30%) and Emax(50%)

Overloaded network link

● Uplink of 1 cluster reduced to 100 KB/s● Remove badly connected cluster, get new one

Iter

atio

n d

ura

tion

Iteration

Connectivity Problems

● Firewalls & Network Address Translation (NAT) restrict incoming traffic

● Addressing problems● Machines with >1 network interface (IP

address)● Machine on a private network (e.g., NAT)

● No direct communication allowed● E.g., between compute nodes and external

world

SmartSockets library

● Detects connectivity problems● Tries to solve them automatically

● With as little help from the user as possible● Integrates existing and several new

solutions● Reverse connection setup, STUN, TCP splicing,

SSH tunneling, smart addressing, etc.● Uses network of hubs as a side channel

Example

Example

[Maassen et al., HPDC’07 ]

Overview

JavaGATZorilla P2P

JavaGAT

● GAT: Grid Application Toolkit● Makes grid applications independent of the

underlying grid infrastructure

● Used by applications to access grid services● File copying, resource discovery, job

submission & monitoring, user authentication

● API is currently standardized (SAGA)● SAGA implemented on JavaGAT

Grid Applications with GAT

GAT Engine

Remote

FilesMonitoring

Info

service

Resource

Management

GridLab Globus Unicore SSH P2P LocalGAT

Grid ApplicationFile.copy(...)

submitJob(...)

gridftp globus

Intelligentdispatching

[van Nieuwpoort et al., SC’07 ]

Zorilla: Java P2P supercomputing

middleware

Zorilla components● Job management

● Handling malleability and crashes

● Robust Random Gossiping● Periodic information exchange between nodes● Robust against Firewalls, NATs, failing nodes

● Clustering: nearest neighbor● Flood scheduling

● Incrementally search for resources at more and more distant nodes

[Drost et al., HPDC’07 ]

Overview

Ibis applications● e-Science (VL-e)

● Brain MEG-imaging● Mass spectroscopy

● Multimedia content analysis● Various parallel applications

● SAT-solver, N-body, grammar learning, …

● Other programming systems● Workflow engine for astronomy (D-grid),

grid file system, ProActive, Jylab, …

Overview experiments

● DAS-3: Dutch Computer Science grid

● Satin applications on DAS-3● Zorilla desktop grid experiment● Multimedia content analysis● High resolution video processing

DAS-3

272 nodes(AMD Opterons)

792 cores1TB memoryLAN: Myrinet 10G Gigabit

EthernetWAN

(StarPlane): 20-40 Gb/s

OPN

Heterogeneous: 2.2-2.6 GHz Single/dual-

core Delft no

Myrinet

Gene sequence comparison in Satin (on

DAS-3)

Speedup on 1 cluster Run times on 5 clusters

• Divide&conquer scales much better than master-worker

• 78% efficiency on 5 clusters (with 1462 WAN-msgs/sec)

Barnes-Hut (Satin) on DAS-3

• Shared object extension to D&C model improves scalability

• 57% efficiency on 5 clusters (with 1371 WAN-msgs/sec)

Speedup on 1 cluster Run times on 5 clusters

Zorilla Desktop Grid Experiment

● Small experimental desktop grid setup● Student PCs running Zorilla overnight ● PCs with 1 CPU, 1GB memory, 1Gb/s

Ethernet

● Experiment: gene sequence application● 16 cores of DAS-3 with Globus● 16 core desktop grid with Zorilla ● Combination, using Ibis-Deploy

Ibis-Deploy deployment tool

3574 sec 1099 sec

877 sec

• Easy deployment with Zorilla, JavaGAT & Ibis-Deploy

Multimedia content analysis

● Analyzes video streams to recognize objects

● Extract feature vectors from images● Describe properties (color, shape)● Data-parallel task implemented with C+

+/MPI

● Compute on consecutive images● Task-parallelism on a grid

MMCA application

Client

(Java)

Broker

(Java)

Parallel

Horus

Server

Parallel

Horus

Servers

Servers

(C++)

(any machine world-wide)

(local desk-top machine)IbisIbis

(Java)

(grid)

MMCA with Ibis● Initial implementation with TCP was unstable● Ibis simplifies communication, fault tolerance● SmartSockets solves connectivity problems● Clickable deployment interface● Demonstrated at many

conferences (SC’07)● 20 clusters on 3 continents, 500-800 cores

● Frame rate increased from 1/30 to 15 frames/sec

[Seinstra et al., IEEE Multimedia’07 ]

MMCA

‘Most Visionary Research’ award at AAAI 2007, (Frank Seinstra et al.)

High Resolution Video Processing

● Realtime processing of CineGrid movie data● 3840x2160 (4xHD) @ 30 fps = 1424 MB/sec

● Multi-cluster processing pipeline ● Using DAS-3, StarPlane and Ibis

CineGrid with Ibis● Use of StarPlane requires no configuration

● StarPlane is connected to local Myrinet network● Detected & used automatically by SmartSockets

● Easy setup of application pipeline● Connection administration of application is simplified by the IPL

election mechanism

● Simple multi-cluster deployment (Ibis-Deploy)● Uses Ibis serialization for high throughput

Summary● Goal: Simplify grid

programming/deployment

● Key ideas in Ibis● Virtual machines (JVM) deal with

heterogeneity● High-level programming abstractions

(Satin)● Handle fault-tolerance, malleability,

connectivity problems automatically (Satin, SmartSockets)

● Middleware-independent APIs (JavaGAT)● Modular

Acknowledgements Current members Rob van Nieuwpoort Jason Maassen Thilo Kielmann Frank Seinstra Niels Drost Ceriel Jacobs Kees Verstoep

Roelof Kemp Kees van Reeuwijk

Past members John Romein Gosia Wrzesinska Rutger Hofman Maik Nijhuis Olivier Aumage Fabrice Huet Alexandre Denis

More information● Ibis can be downloaded from

● http://www.cs.vu.nl/ibis

● Papers:● Satin [PPoPP’07], SmartSockets [HPDC’07],

Gossiping [HPDC’07], JavaGAT [SC’07],MMCA [IEEE Multimedia’07]

• Ibis tutorials• Next one at CCGrid 2008 (19 May, Lyon)