Portable Heterogeneous High-Performance Computing via ...

ORNL is managed by UT-Battelle for the US Department of Energy

Portable HeterogeneousHigh-PerformanceComputing viaDomain-SpecificVirtualization

Dmitry I. [email protected]

This research used resources of the Oak Ridge Leadership Computing Facility at the Oak RidgeNational Laboratory, which is supported by the Office of Science of the U.S. Department of Energyunder contract No. DE-AC05-00OR22725.

2 ExaTensor

Application-Hardware Interface

HPC Application

Hardware/HPC-system

Interface Width

Interface Width ~ O(1): Manageable Performance Portability, regardless ofthe programming model(only O(1) parts of the codeinteract with raw hardware)

Direct Interface

Interface Width ~ O(N): Performance Portability canbecome an issue, even ifusing OpenMP/OpenACC

HPC Application

Runtime + Libraries

Interface Width

Performance Portability effort is shifted towards Runtime and Libraries

Uni-Indirect Interface

Hardware/HPC-system

Interface Width

HPC Application responsibility to be aware of Hardware

HPC Application

Domain-SpecificRuntime + Libraries

Interface Width

Performance Portability effort is shifted towardsdomain-specific Runtime and Libraries

Bi-Indirect Interface

DS Runtime is aware of boththe Application and Hardware

Hardware/HPC-system

Interface Width

Performance portability effort is shifted towards Runtime and Libraries

HPC Application is awareof Runtime, Runtime isresponsible to be awareof Hardware

3 ExaTensor

Domain-Specific Virtual Processor

HPC Application

Domain-SpecificVirtual Processor + Libraries

Interface Width

Hardware/HPC-system

Interface Width

A better structured runtime thatformally resembles a processorbut specialized to the domain-specific workloads

Understands the specificityof domain-specific data,operations, and algorithms

Understands the specificityof a class of HPC architecturesvia a parameterized templateof node architectures

➢ The domain HPC applications express their algorithms in a domain-native language (higher-level),either via a standalone or embedded DSL;

➢ The domain algorithms are compiled into instructions for a domain-aware virtual architecture;➢ The virtual architecture is built from virtual units that map reasonably well to the physical hardware;➢ Hardware encapsulation, yet potential for a seemless co-design.

4 ExaTensor

Domain-Specific Virtual Processor

DSInstruction

Queue

DSMetadata

Cache

Device 0Instruction

Queue

Device 0

Device 0Instruction

Queue

Device 1

Device 1Instruction

Queue

Device 2

Device 2Instruction

Queue

Device 3

Device 3Instruction

Queue

DSHierarchical

DataCache

DS-RDMA

DSCheckpoint

Parameterized domain-specific virtual architecture template

DSInstructionScheduler

Requires: Building blocks for heterogeneous computing: HiHat?

5 ExaTensor

Performance Portability Strategy• Abstract computing system template:

– Distributed (weakly-coupled) level: The computing systemis composed of compute nodes interconnected via network interfaces in some topology

– (Semi-)shared (strongly-coupled) level: Each node is composed of multiple compute devices of the same or different kinds, possibly sharing the same (hierarchical) memory

• Algorithms are formulated for this abstract computer

• The hardware specificity is masked by driver libraries that provide a device-unified API interface for a set of necessary domain-specific primitives (DS-ISA)

• New hardware = New driver library

6 ExaTensor

Use Case: Numerical Tensor Algebra

• Quantum many-body theory: Electronic and Nuclear structure, Quantum computing: Wave-function

• Loop quantum gravity: Fine structure of space-time

• Multivariate data analytics: Extracting correlations, features, data compression

• Classical physics

• Any data arranged as a multi-dimensional array is a tensor (in the most general sense)

• Any function of multiple variables discretized over a grid or some functional inner-product space is a tensor

7 ExaTensor

Coupled-Cluster Theory

Electron correlation = Correlationbetween hole-particle excitations

i a

j b

i aj b

kc

i aj b

i aj b

8 ExaTensor

Automated Development Tools and DSL• Automated equation/code generators are unavoidable

• DSLangs and DSLibs can provide a scalable development environment: DiaGen input/output:

Efficient flexible runtime?

9 ExaTensor

Past/Present Experience• 2002-2005: CLUSTER: Automated code generation:

CAS-CCSD method = Millions of generated SLOC

• 2006-2008: CLUSTER moved to a direct interpretation of many-body CC equations: High-level spec → Bytecode

• 2009-2013: ACES III and ACES IV: Domain-specific super-instruction language and runtime (SIAL/SIP):SIAL (med.level) → SIP bytecode → Interpretation by SIP

• 2014+: ExaTENSOR framework: Direct interpretation of generic tensor expressions on heterogeneous HPC architectures with cross-domain applications

• Other efforts: Code generation (TCE), embedded DSL + Tensor Algebra library (Aquarius/Cyclops, TiledArray)

10 ExaTensor

Math Framework: Basic Tensor Algebra• Formal tensor: : n-D Array

Full tensor:

Tensor slice:

• Tensor addition:

• Tensor product:

• Tensor contraction:

Compute intensive (potentially)!

Parallelism!

Parallelism!

Few primitive operations:

,...),,,(...... srqpTT pq

rs SsRrQqPpsrqpT ,,,:),,,(

SSsRRr

QQqPPpsrqpT

','

,',':),,,(

pqrs

pqrs

pqrs RLTsrqp :,,,

qs

pr

pqrs RLTsrqp :,,,

qbcdrsai

paibcd

pqrs RLTsrqp :,,,

11 ExaTensor

Math Framework: Tensor Decompositions

• Graphical (diagrammatic) representation:

Matrix: Matrix*Matrix:

• Linear algebra: SVD is optimal in the 2-norm:

• Tensor (multi-linear) algebra: Many choices:

Canonicalpolyadic

Tucker Tensor tree Tensor train

12 ExaTensor

Tensor Sparsity• Dense tensors:

• Block-sparse tensors (container of dense slices):

• Regular sparse tensors:– Some regular sparsity pattern other than block sparsity

• Irregular sparse tensors:– Sparsity pattern has no regularity

13 ExaTensor

Adaptive (+Hierarchical) Tensor Algebra

• Large tensor elements can become tensors themselves (higher resolution);

• Weak tensor slices can be compressed by lowering the resolution, up to a single (complex) number;

• Adapt to the calculated electronic state and available HPC resources;

• Should be better than just black-and-white discarding.

D.I.L. Int. J. Quantum Chem. 2014; D.I.L. ArXiv 2017

Extrapolation of H/H2-matrixalgebra to TENSORS

14 ExaTensor

Computational Challenges• Dense tensor algebra:

– Communication bandwidth: Communication avoiding and regularization (similar to the matrix multiplication, e.g.,CARMA by Demmel et al.)

– Memory size: Out-of-core algorithms

• Block-sparse tensor algebra adds:– Irregular data placement and computational workload: Load

balancing → Task-based programming model

• Adaptive hierarchical block-sparse tensor algebra adds:– Dynamic irregular data placement and computational workload:

Load balancing, dynamic data layout → Adaptive task-based programming model

ExaTENSOR: 1. Implementation of the tensor algebra DSVP for heterogeneous computing; 2. Implementation of scalable tensor algebra algorithms for TA-DSVP.

15 ExaTensor

Domain-Specific Virtual Processing • Domain-specific algorithm → Domain-specific virtual processor(s)

→ Physical processor(s):

• Why this way?– Less unneeded detail, more focus

on the domain-specific algorithms (TA);– Better portability;– Better debugability/profileability

of domain-specific algorithms;– Better opportunities for co-design.

Domain-SpecificVirtual Processor

(abstract class)

Domain-SpecificInstruction

(abstract class)

Domain-SpecificOperand

(abstract class)

TensorOperand

TensorInstruction

Tensor AlgebraProcessor

Logic Processor(control, meta-data)

Numeric Processor(numeric tensor ops)

0 1 2 3

0 1 2 3

7

11

654

8 9 10

Heterogeneous DSVPHeterogeneous DSVP(ExaTENSOR)(ExaTENSOR)

16 ExaTensor

Node-Level Virtualization: Hiding HardwareHardware Processors (CPU, GPU, etc.)

17 ExaTensor

Node-Level Virtualization: Hiding HardwareDomain-Specific Virtual Processor

(DSVP)

ExaTENSOR DSVPDomain-Specific

Instructions

Data Data

18 ExaTensor

Node-Level Virtualization: Hiding HardwareDomain-Specific Virtual Processor

(DSVP)

ExaTENSOR DSVPDomain-Specific

Instructions

Data Data

Driver Libraries (Tensor Algebra)Driver Libraries (Tensor Algebra)

19 ExaTensor

Global Virtualization: Hiding HPC Scale

20 ExaTensor

Global Virtualization: Hiding HPC Scale

21 ExaTensor

Global Virtualization: Hiding HPC Scale

22 ExaTensor

Global Virtualization: Hiding HPC Scale

23 ExaTensor

Global Virtualization: Hiding HPC Scale

24 ExaTensor

Recursive Dynamic Task Scheduling

O(1)O(1) O(1)

Data storage granularity is decoupled from the task granularity

25 ExaTensor

TAL-SH: On-Node Tensor Algebra Layer• Basic tensor algebra operations: Tensor contraction, tensor

product, tensor addition, tensor norm, etc.

• Supports multicore CPU and (multiple) NVIDIA GPU

• Asynchronous task-based execution on accelerators

• Automatic memory and data transfer management,data transfer pipelining

• User-controlled data presence management (images)

• Tensor contractions:

– Up to 1.1 TFlop/s double precision on NVIDIA Kepler K20x for arithmetic intensities > 1000

– Up to 4.2 TFlop/s double precision on NVIDIA Pascal P100 for arithmetic intensities > 1000

26 ExaTensor

TAL-SH: On-Node Tensor Algebra Layer

https://github.com/DmitryLyakh/TAL_SH.git

27 ExaTensor

cuTT: Fast Tensor Transpose for GPU

https://github.com/ap-hynninen/cutt.git A.P.Hynninen, D.I.L. ArXiv 2017

28 ExaTensor

Conclusions• Domain-Specific Virtual Processor pros:

– Domain algorithm friendly (higher-level):Application-Runtime codesign, customizable virtual architecture;

– Hardware agnostic by design:Only requires new driver libraries with domain-specific primitives;

– HPC scale agnostic by design (via a recursive HPC system view);

– Better fault-tolerance (autonomous virtual nodes can be turned on/off);

– Good model for Hardware-Runtime codesign;

– Better debugging (in terms of higher-level operations);

– Better profiling (in terms of higher-level operations).

• Domain-Specific Virtual Processor cons:– Level of effort;

– Runtime overhead (like any virtual machine).

29 ExaTensor

Potential Programming Models• Direct non-task static: Algorithm → Target code (MPI+X or

PGAS+X) → Hardware

• Direct task static: Algorithm → Parameterized static DAG → Task-based runtime → Hardware

• Direct task dynamic: Algorithm → Dynamic DAG → Task-based runtime → Hardware

• Indirect non-task static: Algorithm → DSL → Domain-specific bytecode → Runtime interpretation → Hardware

• Indirect task static: Algorithm → DSL → Parameterized static DAG → Task-based runtime → Hardware

• Indirect task dynamic: Algorithm → DSL → Dynamic DAG → Task-based runtime → Hardware