HPC2012 Workshop Cetraro, Italy

Supercomputing and Big Data: Where are the Real Boundaries and Opportunities for Synergy?

Bill BlakeCTO

Cray, Inc.

The “Big Data” Challenge

6/30/20122

Supercomputing minimizes data movement –

The focus is loading the “mesh”

in distributed memory, computing the answer with fast interconnects, and visualizing the answer in memory –

the high performance

“data movement”

is for loading, check pointing or archiving.

Data-intensive computing is all about data movement –

The focus is scanning, sorting, streaming and aggregating "all the data all the time" to get the answer or discover new knowledge from unstructured/structured data sources.

Big Data is “Data in Motion”

6/30/20123

•Set the stage for the fusion of numerically intensive and data intensive computing in future Cray systems

•Build on Cray's success of delivering the most scalable systems through heterogeneous and specialized nodes•

•

Nodes not only optimized for compute, but also storage and network I/O, all connected with the highest level of interconnect performance. Add system capability to the edge of the fabric

•We see this effort as increasing key application performance with an "appliance style" approach using Cray's primary supercomputing products with extensions configured as optimized HW/SW stacks –

adding value

around the edge of the high performance system network

Maximum Scalability: System Node Specialization

Net I/O

System Support

Service

Sys Admin

Users

File I/O

Compute

/home

Pric

e

Performance

PCs and Workstations• 10,000,000s units

Enterprise servers• 100,000s units

High-Performance Computing• 1,000s units

Courtesy of Dr. Bill Camp, Sandia National Laboratories, circa 2000

Key to Cray’s MPP scalability is system node o/s specialization combined with very high bandwidth, low latency interconnects

A very effective approach to “appliance”

design: Netezza and Cray examples

Back in Time, at the Beginnings of the Web, there was Dr. Codd and his Relational Database …

Algebraic Set Theory

Analyzing Big Data Analytics: A “Swim Lane” View of Supporting Business and Technical Decisions

KeyFunction

Language DataApproach

“Airline”Example

OLTP Declarative(SQL)

Structured(relational)

ATM transactionsBuying a seat on an airplane

OLAPAd Hoc

Declarative(SQL+UDF)

Structured(relational)

BI aggregate and analyze bookings for new ad placements

SemanticAd hoc

Declarative(SPARQL)

Linked, Open(graph-based)

Analyze social graphs and infer who might travel where

OLAPAd Hoc

Procedural(MapReduce)

Unstructured(Hadoop files)

Application Framework for weblog analysis

OptimizeModels

Procedural(Solver Libs)

Optimization<-> Simulation

Complex Scheduling Estimating empty seats

SimulateModels

Procedural(Fortran, C++)

Matrix Math(Systems of Eq’s)

Mathematical Modeling and simulation (design airplane)

For Perspective (1980’s) …

●

The relational database was invented on a system that merged server, storage and database

●

It was called a mainframe

●

Change focus to vector processing and memory performance and the mainframe becomes a supercomputer

KeyFunction

Language DataApproach

SMP Server

ClusterAnd MPP

Cloud And Grid

WebScale

OLTP Declarative(SQL)

Structured(relational)

CPU

Memory

IOPIOP

OLTP: Processing Tables with Queries

Processing millions of transactions without a hitch

●

Do entire transaction or nothing (don’t debit account without dispensing cash)

●

Disciplined approach to data forms data models/schema tables on disk

●

Tables at first glance look like the rows and columns of a spreadsheet

●

Transaction processing uses highly repetitive queries for entering/updating

KeyFunction

Language DataApproach

SMP Server

ClusterAnd MPP

Cloud And Grid

WebScale

OLTP Declarative(SQL)

Structured(relational)

OLTP: Processing Tables with Queries●

A transaction “touches” only the data it needs to accomplish its purpose●Since the workload involves many, many small

data payloads, speedups can be gained through caching and the use of indices that prevent “full table scans”.

●

The query states what it needs, and the database uses a (cost-based) planner/optimizer to create the program that actually manipulates the tables

●

Mostly writing to the database

Then The Rules Changed

●Mainframes attacked by killer micros!

●Memory grew large● I/O became weak●System costs dropped●Storage moved off to

the network

CPU CPU CPU CPU

CPU CPU CPU CPU

Very LargeMemory

I/O

Storage AreaNetwork

Capacity Was Added By Clustering

CPU CPU CPU CPU

Memory

I/O

CPU CPU CPU CPU

Memory

I/O

CPU

CPU

CPU

CPU

Mem

ory

I/O

CPU CPU CPU CPU

Memory

I/O

CPU CPU CPU CPU

Memory

I/O Storage AreaNetwork

SAN limits Moving Data to the Processors

The Online Analytical Processing Problem

●

Business Intelligence Applications generate reports of aggregations●

Need to read at all the data all the time (telecom, retail, finance, advertising, etc)

●

BI Analytics require ad hoc queries since you don’t know the next question to ask until you understand the answer to the last question

●

Standard SQL is limited by the Algebraic Set Theory basis of RDBMS, if you need Calculus then insert User Defined Functions into the SQL

●

Programming models in conflict as Modeling and Simulation combine with Business Intelligence in Predictive Analytics

KeyFunction

Language DataApproach

SMP Server

ClusterAnd MPP

Cloud And Grid

WebScale

OLAPAd Hoc

Declarative(SQL+UDF

)

Structured(relational)

OLAP DatabasesOLAP Databasesaka Data Warehouseaka Data Warehouse

Processing Hundreds Processing Hundreds of Terabytes/hourof Terabytes/hour

Extraction /Cleansing

Trans-formation Loading

ReportsReports

OLTP DatabasesOLTP Databases

Processing Millions Processing Millions Of Transactions/secOf Transactions/sec

Operational vs. Analytical RDBMS

ClientApplications

Local Applications

Hours or Days

$$$ Processing

$$$ Storage

$$$ Fiber Channels

SMP HOST 1

SMP HOST 2

SMP HOST N

C6F7

C12F13

C38 F39G13 G22

$$$

$$$

Data Flow – The SAN Bottleneck

The Netezza Idea: Moving Processing to the Data

●

Active Disk architectures●

Integrated processing power and memory into disk units

●

Scaled processing power as the dataset grew

●

Decision support algorithms offloaded to Active Disks to support key decision support tasks●

Active Disk architectures use stream-based model ideal for the software architecture of relational databases

●

Influenced by the success of Cisco and NetApp appliances, the approach combined software, processing, networking and storage leading to the first database warehouse appliance!

Netezza is an IBM Company

The HW Architecture Responded to the Needs of the Database Application

SQL Compiler

Query Plan

Optimize

Admin

Front End

DBOS

Execution Engine

Linux SMP Host Massively Parallel

Intelligent Storage

1

2

3

1000+

Gigabit Ethernet

Processor+streaming DB logic

High-Performance

Database Engine

Streaming joins, aggregations, sorts, etc.

Processor+streaming DB logic

Processor+streaming DB logic

Snippet Processing Unit (SPU)Netezza Performance ServerClientBI

Applications

Fast Loader/Unloader

Local Applications

ODBC 3.XJDBC Type 4SQL/92

Processor+streaming DB logic

Move processing to the data (maximum I/O to a single table)

The Database Application

Shifting from Analyst to Programmer

●Google, Yahoo and their friends are using a data-intensive application framework to analyze very large datasets (e.g.,weblogs) without transactions or structured data●What will this look like in 10 years?

●

MapReduce/Hadoop: an programming model/application framework performing “group-by (map), sort and aggregation (reduce)”

●

Not queries, but programs willing to forgo the need for transactional integrity or the performance of structured data (4X-5X disadvantage on equal hardware, but with excellent scaling on cheap hardware)

●

An increasingly popular approach with organizations that have the programming talent to use it, especially research organizations

●

Another frontal assault on the $40K per socket RDBMS licensing

KeyFunction

Language DataApproach

SMP Server

ClusterAnd MPP

Cloud And Grid

WebScale

OLAPAd Hoc

Procedural(MapReduce)

Unstructured(Hadoop Files)

●

Multi-dimensional predictive queries●

Connection networks, social network, Time, Space, Reasoning Find why the customers who live on the 25th street in Zurich did not switch their phone company between March 23 to August 19 while their friends and family switched their phone company and predict the trend moving forward?

●

Need to pick weak signals from the noise●

Data size will continue to grow and be more heterogeneous

●

Finding needle in a hay stack will become more important

●

Will require real time or shorter turnaround time●

Differentiation will be based on the speed of analysis of changing data

●

INFERENCE – “maybe” as a legitimate answer

Analysis is Getting Complex

18

Advancing from Generating Reports to Inference and Discovery

●The International W3C standards body has approved the key standards, called RDF and OWL to support the Semantic Web aka Web 3.0 with “machine readable”

open linked data

●Future databases will use triples (subject-predicate-object) vs tables and with RDF/OWL federate heterogeneous data●Future databases will support reasoning not just reporting●This work started as a combined European Defense and DARPA effort●Major RDBMS vendors are admitting Relational and XML are ill-suited to the needs of the semantic web of the future

KeyFunction

Language DataApproach

SMP Server

ClusterAnd MPP

Cloud And Grid

WebScale

SemanticAd hoc

Declarative(SPARQL)

Linked, Open(graph-based)

The Idea: Address Memory and Network Latency

●

Decision support algorithms offloaded to multi- threaded processors and in-memory database with new complex key decision support tasks●

Supports new SPARQL query processing for RDF triples database offering the speedup of XMT processing without low level API

●

Shared memory, multi-threaded Cray technology●

Integrated with semantic database (open source std compliant)

●

Fastest complex query response on open linked data in the industry●

Influenced by the success of database warehouse appliances, our combined (database) software, processing, networking and memory led to the first graph appliance!●

Deliver easy to deploy solutions requiring knowledge discovery for Intelligence, Bioinformatics, Finance, etc.

?

Performs like a supercomputer Uses open web 3.0 standards

Scales like a web engineOperates like a data warehouse

21

Now a new tool for data analysis….

Adapt the system to the application not the application to the systemVision: ExascaleCray’s Adaptive Supercomputing combines

multiple processing architectures into a single scalable system—CPU, GPU, or Multi-

threaded

The focus in on the user’s application where the adaptive software, the compiler or query processor, knows what types of processors are available on the heterogeneous system and targets code to the most appropriate processor

The next step is to evolve Adaptive Supercomputing to Big Data workloads

Enabling Simulation and Data Science

Adaptive Supercomputing