High-Performance Architecture and Heterogeneous Computing

50 downloads 800 Views 419KB Size Report
Michael Gschwind. IBM Systems and Technology Group ... Cryptography, TOE, stream matching, data base query… – Programmable ... Text analytics and intrusion detection ... Protect software investment by ensuring program portability.
Michael Gschwind IBM Systems and Technology Group

High-Performance Architecture and Heterogeneous Computing

© 2011 IBM Corporation

Accelerator architecture

ƒ Heterogeneous Computing – Augment a general purpose CPU with accelerators ƒ Accelerators optimized for application domain

– Focus on power and performance optimization with respect to defined function ƒ Trade off generality and domain optimization – Fixed function • Cryptography, TOE, stream matching, data base query… – Programmable • Cell SPE, Azul Java accelerator, … ƒ Top down definition process

– Application needs drive system definition

2

ICCD 2011 M. Gschwind, High-Performance Architecture and Heterogeneous Computing

© 2011 IBM Corporation

Numeric acceleration

ƒ Numeric acceleration will be key to addressing the exascale challenge ƒ High-performance computing systems limited by power dissipation – Need to dramatically increase power efficiency ƒ Accelerators improve overall computational efficiency – Optimize for program controlled data menagement – Reduce core size and increase on-chip parallelism

Cell B.E.with 9 cores

Source: Horst Simon 3

ICCD 2011 M. Gschwind, High-Performance Architecture and Heterogeneous Computing

Source: Horst Simon © 2011 IBM Corporation

Middleware acceleration

ƒ Garbage collection for Java – Competitive performance and superior performance/resource tradeoff ƒ Text analytics and intrusion detection ƒ Query acceleration with FPGAs (Neteeza) (main memory only)

VEE 2008 4

ICCD 2011 M. Gschwind, High-Performance Architecture and Heterogeneous Computing

IPDPS 2008 © 2011 IBM Corporation

Heterogeneous System Challenges

ƒ Programming as the biggest challenge for accelerator exploitation –Programmer productivity –Realize accelerator performance benefits –Investment protection for applications

5

ICCD 2011 M. Gschwind, High-Performance Architecture and Heterogeneous Computing

© 2011 IBM Corporation

System Integration and Programmability

ƒ Programmability and Efficiency – True peer to peer model without callbacks to “master” CPU – Limit interventions by general purpose core

ƒ Integrate accelerators into system model – Common, coherent view of data – Shared view of virtual memory

6

ICCD 2011 M. Gschwind, High-Performance Architecture and Heterogeneous Computing

© 2011 IBM Corporation

Programming models

ƒ Focus on standardized program models in support of heterogeneous computing – Ensure ease of programming – Develop ecosystem and development skills – Protect software investment by ensuring program portability ƒ OpenMP-based autoparallelization – Ranked #1 in programmer choice in NA ƒ OpenCL-based computing kernels – Ranked #2 in programmer choice in NA ƒ Work on heterogeneous programming models driven by industry needs – Need for increased focus in academic research – Programming models for FPGAs • LIME effort at IBM Research

Programmer Choice Data as reported by Bergman, 2011 7

ICCD 2011 M. Gschwind, High-Performance Architecture and Heterogeneous Computing

© 2011 IBM Corporation

Leading the Way in Heterogeneous Computing

SPE 91% PPE 6%

Opteron 3%

RoadRunner: First to PetaFlop

zBX: System z BladeCenter Extension

PowerPC SoC •Ethernet Offload Engine •Cryptography Unit •Regular Expression Unit •XML Processing Unit •Compression Unit

IBM Power Edge of Network™ Processor 8

ICCD 2011 M. Gschwind, High-Performance Architecture and Heterogeneous Computing

Data Warehousing © 2011 IBM Corporation

Suggest Documents