January 29, 20032003 Tsunami Research, Inc.

Presented to COOUG

February 11, 2003

Hive ComputingA New Vision for

Mission Critical Applications

Agenda

• Definition• Executive Summary• Our History• The Problem• The Five Pillars• Benefits• Demo• The Technology• eXtreme Programming

Definition

HIVE COMPUTING

A revolutionary approach to thedevelopment, deployment, and management

of applicationsthat involves enabling

large numbers of dedicated, commodity computersto join together and form

a mission critical computing environmentcalled a Hive

Executive Summary

The goal of Hive Computing isto make reliability affordable

In Essence

CommodityComputers

High PerformanceLow Price

In Essence

CommodityComputers

High PerformanceLow Price

Super-computing

Performance- Scalability

BeowulfUnlimited Scale

QlustersPlatform Computing

United DevicesEntropia

Data Synapse

Linux ClustersDistributed Supercomputers

In Essence

MissionCritical

Computing

Tsunami Research

Hive Computing

Affordability- Adaptability- Maintainability- Usability

Reliability- Scalability- Availability- Predictability

+

CommodityComputers

High PerformanceLow Price

Super-computing

Performance- Scalability

BeowulfUnlimited Scale

QlustersPlatform Computing

United DevicesEntropia

Data Synapse

Linux ClustersDistributed Supercomputers

• Significant transaction processing experience• Core team previously founded PaylinX Corporation• Leading real-time enterprise payment processing solution• PaylinX sold to CyberSource in 2000 for $130+ million• Unfinished business…

– Scalability– Reliability– Complexity– Cost

Our History

The Problem

MainframesFault

TolerantComputers

ApplicationServers

LinuxClusters

DistributedSupercomputers Hive

Reliability

Scalability

Availability

Predictability

Affordability

Adaptability

Maintainability

Usability HP/Tandem

StratusBEAIBM

BeowulfQlusters

Unlimited Scale

United DevicesData Synapse

Entropia

Existing solutions force organizations to make trade-offs

IBMUnisysFujitsu

Why Can’t Solutions…

• Scale without limits?• Be immune to failure?• Guarantee the job will get done?• Keep up with a changing world?• Maintain themselves?• Be easy to use and develop for?

Old Problems Result From Old Assumptions

• TCP/IP• Not a better centralized or decentralized network• The core assumption…

– Networks are unreliable

• Design implications– The network could not have any centralized authority– Messages would be divided up into equally-sized, independently-routed packets– The system should be easy to expand by adding additional, inexpensive resources

New Assumptions Create New Possibilities

• The application is the only thing that matters• Self-organizing• Hardware and operating system layers relatively unimportant

• Computers will fail• Self-healing• Failure happens, deal with it

• Computers are disposable• Self-maintaining• Inexpensive• Make it very easy to add and remove hardware• Dedicate hardware to simplify solution & improve results

Application Fabric

Application

Operating SystemCommodity

Hardware

Allows you to leverage commodity hardwareand inexpensive operating systems

(e.g. Windows 2000 and XP, FreeBSD, Linux)

Hive

Hive Computing

Application

Hive

CommodityComputer

CommodityComputer

CommodityComputer

CommodityComputer

CommodityComputer

CommodityComputer

Delivers scalability and reliability without the cost

The Five Pillars of Hive Computing• Collective Intelligence

- The intelligence of a Hive resides in every member- Peer-to-peer- Bottom-up, not top-down- Role-based

• Commodity Components– PC-grade computers and blade servers

• Assimilation– The Hive automatically brings new Workers on line

• Self-Monitoring– The Hive monitors and recovers all Tasks

• Self-Maintaining– The Hive maintains own configuration

Benefits

• Reliability :: A Hive has no reason to ever go down– Scalability

• A Hive can scale in small increments and in a near-linear manner– Availability

• No single point of failure– Predictability

• The job gets done in a specified period of time

• Affordability :: Low acquisition, maintenance, & development costs– Adaptability

• A Hive can easily adapt to changing conditions– Maintainability

• Dramatically reduced administration and maintenance costs– Usability

• Developers can focus on the application

Hive Architecture

Logical, not Physical

Sample ArchitectureClient Client Client ClientClient Client Client Client Client

Presentation Layer

Business Logic Layer

Hive

Application Server

Access ControlAuthenticationPage Rendering

DatabaseHigh Value OperationsHigh Volume Operations

Sample Architecture

Hive

HiveSystem

System

Client

Client

Client

Enterprise ApplicationsWeb Services

Sample Architecture

Hive Hive

ResourcePool

Client

Grid

ResourcePool

Demo

• Collective Intelligence• Load balancing and fail-over• Fault tolerance• Scalability

Request/Response Loop

Client

Hive Mind

requestresponse

task taskresult

taskresult

task

WorkerMachine A

WorkerMachine B

WorkerMachine C

WorkerMachine

WorkerMachine

Security

• SSL encryption between Workers and the Hive• Authentication required to access Operations, Process Flows, and

Tasks• Additional layers of security planned or in the works

Terminology

• Hive• Worker• Package• Application• Operation• Process Flow• Task

Hive     Worker     Worker     Worker          Package               Application                    Operation                    Operation               Application                    Operation                    Operation                    Operation                         Process Flow                         Task

Operation• A brokerage firm might perform the following Operations in a Hive... 

– GetRealtimeQuoteForCUSIP – GetDelayedQuoteForCUSIP – GetChartForCUSIP – CalculateCurrentPortfolioValue

• A credit card company might perform the following Operations in a Hive...– PerformAuthorization – PerformFraudCheck – CaptureTransaction

• A mortgage company might perform the following Operations in a Hive...– GetScoreForApplication – CalculateMonthlyPayment – CalculateEscrow

• A security application might perform the following Operations in a Hive...– AnalyzeFingerprint – AnalyzeFace– AuthenticateUser

Worker

Process Flow

Task• An atomic function to be performed by the Hive (e.g. a syntax check

or database lookup)• A Task can be written in C/C++ or Java• A Task can optionally implement OnStartup and OnShutdown

functions to initialize memory and or the environment

Task Examplevoid makeStringUppercase(TASK_HANDLE request){

string data = task_getAttribute( request, "argument" );data = ToUpper(data);task_setAttribute( request, "argument", data );task_setResult( request, "COMPLETE" );

}

• All data passes into the task through the TASK_HANDLE data structure• Hive Developer’s task code reads arguments from the TASK_HANDLE object• The task performs operations as necessary • Any data updates are made to the TASK_HANDLE object

Limited Tasks

• Some tasks should not run on every machine such as connections to limited resources

• These tasks can be ‘limited’ by the Hive to run on either X total machines or Y percentage of machines

Rollbacks• Application level rollbacks are handled via compensating tasks

within the process flow• Only the application developer is aware of how to “undo” any of the

atomic actions that have been completed• The Hive’s process flow allows the application developer to catch

errors generated by the task and send the transaction to a task to undo any actions within the process flow

Databases and a Hive• Hive Developers should leverage existing database solutions for

data storage• Database connection code can be written as tasks for the Hive to

manage• If one machine with a database connection goes down, the Hive can

bring another connection up

Hive API

• The Hive is accessed through thin client side API’s written in C/C++ or Java in Windows or UNIX

• Client API addresses the Hive as a whole, not a single machine• Client API can use SSL connections when communicating with the

Hive

API ExampleHAPI_PROCESS process;process = hapi_createProcess();

hapi_setAttribute( process, "aString", "ABCDEF");long rc = hapi_executeProcess( process,

"integration_test_app", "CaseProcess", "224.0.0.1", 31000, HAPI_NO_ENCRYPTION );

string resultString = hapi_getAttribute( process, "aString" );string processResult =hapi_getProcessResult( process );

hapi_destroyProcess( process );

Deploying Hive Applications

• Hivepkg bundles process flows, tasks, and support files and provides syntax checking before deployment

• Hivectl deploys applications to a Hive• Applications can be deployed to a running Hive• The Hive gracefully updates a subset of workers while the majority of the

workers continue processing transactions

eXtreme Programming

• Test-Driven Development– Automated build process– Each build is subjected to 3,000+ unit tests– Extremely tight code

• Paired Programming• The Planning Game

Suitable Applications

• Transaction-Oriented applications– Request/Response loop

• Financial Services & Insurance– Credit Card Processors– Trading Systems– Financial and process modeling

• Homeland Security, Government, & Defense– Applications where scalability, reliability, & survivability are important

• Biotechnology & Life Sciences– Bioinformatics– In-silico drug discovery

For More Information

Chris O’Leary314.336.2841 – Directchris@tsunamiresearch.com

Don Mang314.336.2840 – Directdon@tsunamiresearch.com

www.tsunamiresearch.com