COMMUNICATION COMMUNICATE COMMUNITY Henri Bal A PUBLIC-PRIVATE RESEARCH COMMUNITY.

COMMUNICATION

COMMUNICATE

COMMUNITY

Henri Bal

A PUBLIC-PRIVATE RESEARCH COMMUNITY

Agenda

• General introduction to COMMIT/– (slides by Arnold Smeulders)

• Example collaboration NLeSC & COMMIT/: global climate modelling– NLeSC eSalsa project– COMMIT/ IV-e project

COMMIT/

The public-private research community for ICT-research of 60+ (non)profit & 20 science partners.

Our mission is to deliver world-class ICT-research to create opportunities with more & more partners.

DIMENSIONS

ICT-science use-inspired science

Synergy between projects

Dissemination tell the story outside science

Valorization towards use in industry

Internationalization NL is not alone in this field

SUCCESSESICT-science

Synergy

SWEET

Euro-Par 2014 Achievement award

SUCCESSESDissemination

Valorization

InternationalizationInfoSys and Figshare become partners, EU projects, EIT-ICT, …….

THE BIG FUTURE OF DATAWe have presented

60 demonstrators: ICT Science Application Alternative application

for four audiences:

The Big Future of DataBig Data are not always Big.

Often they are: private interactive uncertain unordered new media

In short they are

overwhelming data.

The Big Future of Data

Data-Science will surpass ICT-Science,

that is our future.

IV-e: e-Infrastructure Virtualization fore-Science Applications (P20)

• e-Infrastructures become very complicated– Distributed & heterogeneous

• P20 tries to simplify the e-Infrastructure, ease programming and integration

• WP5 studies how to effectively use GPUs for data processing (imaging) and computing

vrije Universiteit

Global Climate Modeling

COMMIT/

GPU Computing

• Offload expensive kernels for Parallel Ocean Program (POP) from CPU to GPU– Many kernels, fairly easy to port to

GPUs– Execution time becomes virtually 0

• New bottleneck: I/O between CPU & GPUCPU

hostmemory

GPU devicememory

Host

Device

PCI Express link

Different methods for CPU-GPU communication

• Memory copies (explicit)– No overlap with GPU computation

• Device-mapped host memory (implicit)– Allows fine-grained overlap between computation

and communication in either direction• CUDA Streams or OpenCL command-queues

– Allows overlap between computation and communication in different streams

• Any combination of the above

Problem

• ICT problem:– Which method will be most efficient for a given

GPU kernel? Implementing all can be a large effort• Solution:

– Create a performance model that identifies the best implementation:

• Which strategy for overlapping computation and communication is best for the given program?

Ben van Werkhoven, Jason Maassen, Frank Seinstra & Henri Bal: Performance models for CPU-GPU data transfers, CCGrid2014(nominated for best-paper-award, top 1%)

MOVIE

Example result

Measured Model

Discussion

• Good example of use-inspired Computer Science research

• Many other application domains that use accelerators face the same problem– Digital forensics, water management, astronomy …

• All details:

27 October 2014 1 October 2014

Different GPUs (state kernel)

Different GPUs (buoydiff)

Global Climate Modeling

• Understand future local sea level changes• Needs high-resolution simulations• Combine two approaches:

– Distributed computing (multiple resources)• Ibis couples models for land, ice, ocean, atmosphere

– GPUs(Graphics processing units)

COMMIT/

Comes with spreadsheet