Mapping Biomedical Applications onto GPU Platforms Joseph JaJa - - PowerPoint PPT Presentation

Mapping Biomedical Applications

• nto GPU Platforms

Joseph JaJa University of Maryland

• Collaboration between GWU (Balaras), UMD

(Solares, Wu), and University of Chicago (Dubey).

• Goal: Development of high performance algorithms

applicable to fluid-structure interactions in viscous incompressible flows.

• Example application: interactions between the red

blood cells and plasma

• Critical Components: Poisson equation solver

combined with a multigrid algorithm. Multi- dimensional FFTs and several types of matrix computations

Fluid-Structure I nteractions

2

• Collaboration between the University of Maryland

(Varshney and JaJa) and the University of Maryland at Baltimore (Gullapalli, Herskovits, etc.)

• Understanding of brain connectivity differences

between subjects with brain disorders and normal subjects using diffusion MRI.

• Dynamics of functional brain connectivity using

resting state fMRI, for subjects with moderate TBI.

Data-Driven Understanding of Brain Disorders

3

Connectivity Matrix

4

• Diffusion MRI images with

64 diffusion frames with resolution 128× 128× 52.

• Probabilistic Tractography
• Number of entries in the

sparse connectivity matrix: 100,000,000- 200,000,000.

• Number of voxels in ROI:

100,000-200,000.

I nflammatory Responses and Wound Healing in Vocal Fold

• N. Seekhao

Collaborators: N. Li, C. Shung, L. Mongeau (McGill U.)

Bio me c ha nic a l Stre ss Muc o sa l Da ma g e Ce ll Re c ruitme nt Ce ll F unc tio n

I nflammatory Responses & Wound Healing in Vocal Fold

Biome c ha nic a l Stre ss

Muc o sa l Da ma g e Ce ll Re c ruitme nt Ce ll F unc tio n

F

• rc e a pplie d o n

tissue . T a lking , sho uting e tc .

Image from : http://2.bp.blogspot.com/- DI0yRAeRKjA/TrDdREMzr_I/AAAAAAAAH6o/QFgZ7xFFRjg/s320/s hout png

I nflammatory Responses & Wound Healing in Vocal Fold

Bio me c ha nic a l Stre ss

Muc osa l Da ma g e

Ce ll Re c ruitme nt Ce ll F unc tio n

Da ma g e in the tissue o f the vo c a l fo ld

Image from : https://wiki.uiowa.edu/download/attachments/39001206/nodules%2 0op%205 png?api=v2

I nflammatory Responses & Wound Healing in Vocal Fold

Bio me c ha nic a l Stre ss Muc o sa l Da ma g e

Ce ll Re c ruitme nt

Ce ll F unc tio n

Attra c ting c e lls suc h a s pla te le ts, ne utro phils, a nd ma c ro pha g e s to the wo und site

Image from : http://www.biospectrumasia.com/IMG/362/44362/atherosclerotic-lesions- generates-robust-t-cell-anti-inflammatory-response-262x174 jpg

I nflammatory Responses & Wound Healing in Vocal Fold

Bio me c ha nic a l Stre ss Muc o sa l Da ma g e Ce ll Re c ruitme nt

Ce ll F unc tion

E a c h c e ll pe rfo rm its duty. One o r mo re o f the fo llo wing :

• Se c re te c he mic a l (I

L

• 1,

MMP-8 e tc .) to a ttra c t, e xc ite o r inhib it o the r c e lls

• De po sit E

CM pro te in (c o lla g e n, e la stin e tc .) to he a l da ma g e d tissue

• Cle a n up c e ll de b ris

Agent-Based Modeling (ABM)

• 1. Bottom-up, rule-based, discrete-event and

discrete-time computational model

• 2. Initial “World” and a collection of “agents.”
• 3. Interactions between agents and the world.

– Agents migrate to area of injury – Remove dead cells and tissue debris – Remodel ECM to heal damaged tissue

• 4. Stochastic moves
• 5. Emergent behavior

ABMs of Vocal Fold Wound Healing Process

T issue a re a o f inte re st (ABMs te rm: Wo rld) Slic e s o f tissue (ABMs te rm: Pa tc he s)

T issue a re a o f inte re st (ABMs te rm: Wo rld) Slic e s o f tissue (ABMs te rm: Pa tc he s)

IL

• 1

MMP-8 …

ABMs of Vocal Fold Wound Healing Process

T issue a re a o f inte re st (ABMs te rm: Wo rld) Slic e s o f tissue (ABMs te rm: Pa tc he s) Co mpo ne nts o f tissue (E CM) suc h a s Co lla g e n, E la stin, Hya luro nic Ac id

IL

• 1

MMP-8 …

ABMs of Vocal Fold Wound Healing Process

T issue a re a o f inte re st (ABMs te rm: Wo rld) Slic e s o f tissue (ABMs te rm: Pa tc he s) Co mpo ne nts o f tissue (E CM) suc h a s Co lla g e n, E la stin, Hya luro nic Ac id

IL

• 1

MMP-8 …

Che mic a l L e ve ls (ABMs te rm: Pa tc he s Attrib ute s)

ABMs of Vocal Fold Wound Healing Process

F ib ro b la st (Ce ll) (ABMs te rm: Ag e nts) Ne utro phil (Ce ll) (ABMs te rm: Ag e nts) Ma c ro pha g e (Ce ll) (ABMs te rm: Ag e nts)

ABMs of Vocal Fold Wound Healing Process

F ib ro b la st (Ce ll) (ABMs te rm: Ag e nts) Ne utro phil (Ce ll) (ABMs te rm: Ag e nts) Ma c ro pha g e (Ce ll) (ABMs te rm: Ag e nts)

ABMs of Vocal Fold Wound Healing Process

F ib ro b la st (Ce ll) (ABMs te rm: Ag e nts) Ne utro phil (Ce ll) (ABMs te rm: Ag e nts) Ma c ro pha g e (Ce ll) (ABMs te rm: Ag e nts)

ABMs of Vocal Fold Wound Healing Process

Problem Scale

* Number of cells increase throughout the simulation due to proliferation. Current model shows a doubling of number of cells after the end of “5-day” simulation.

• Computationally demanding applications with

irregular memory access patterns and involving large data sizes that cannot fit on the GPU memory

• Need to use heterogeneous platforms involving

multicore CPU with one or more many-core GPUs.

• Performance Goal: try to achieve the same

performance rate or throughput as in the case when the data fits on the GPU.

Characteristic Features of Applications

19

Heterogeneous Platforms

CPU Mem (128GB) Dual-socket Multi-core CPU -16 cores GPU GDDR5 Mem (5GB) Massively Parallel GPU K20 I/O I/O

5.7GB/s 73GB/s 208GB/s

Dense Matrix Multiplication

• DGEMM:

where the matrices are of dimensions:

• Why?

– An important kernel for many problems – Optimization ideas can be used in other problems – Perhaps the most-studied algorithm in high performance computing

• Can we solve very large DGEMM with the same

performance throughput as small DGEMM?

n m C n k B k m A × × × : , : , :

Block Matrix Multiplication

• Decompose into blocks

A B C

Multiple CUDA Stream Scheduling

Performance Evaluation

Performance Evaluation

• Many biomedical applications can make

effective use of heterogeneous platforms.

• But a significant amount of work is required

to organize the computation into multi- stream of data transfers and kernel executions with no or very small stall time.

• Portability of high performance code

remains a problem.

Conclusion

26