1 / 19

An Extensible Global Address Space Framework with Decoupled Task and Data Abstractions *

An Extensible Global Address Space Framework with Decoupled Task and Data Abstractions *. Sriram Krishnamoorthy Jarek Nieplocha Umit Catalyurek Atanas Rountev P. Sadayappan The Ohio State University Pacific Northwest Natl. Lab.

wlillian
Download Presentation

An Extensible Global Address Space Framework with Decoupled Task and Data Abstractions *

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. An Extensible Global Address Space Framework with Decoupled Task and Data Abstractions* Sriram Krishnamoorthy Jarek Nieplocha Umit Catalyurek Atanas Rountev P. Sadayappan The Ohio State UniversityPacific Northwest Natl. Lab. *Supported in part by the National Science Foundation

  2. Theory #Terms #F77Lines Year CCD 11 3209 1978 CCSD 48 13213 1982 CCSDT 102 33932 1988 CCSDTQ 183 79901 1992 Time Crunch in Quantum Chemistry Two major bottlenecks in computational chemistry: • Very computationally intensive models (performance) • Extremely time consuming to develop codes(productivity) The vicious cycle in many areas of computational science: • More powerful computers make more and more accurate models computationally feasible :-) • But efficient parallel implementation of complex models takes longer and longer • Hence computational scientists spend more and more time with MPI programming, and less time actually doing science :-( • Coupled Cluster family of models in electronic structure theory • Increasing number of terms => explosive increase in code complexity • Theory well known for decades but efficient implementation of higher-order models took many years

  3. CCSD Doubles Equation (Quantum Chemist’s Eye Test Chart :-) hbar[a,b,i,j] == sum[f[b,c]*t[i,j,a,c],{c}] -sum[f[k,c]*t[k,b]*t[i,j,a,c],{k,c}] +sum[f[a,c]*t[i,j,c,b],{c}] -sum[f[k,c]*t[k,a]*t[i,j,c,b],{k,c}] -sum[f[k,j]*t[i,k,a,b],{k}] -sum[f[k,c]*t[j,c]*t[i,k,a,b],{k,c}] -sum[f[k,i]*t[j,k,b,a],{k}] -sum[f[k,c]*t[i,c]*t[j,k,b,a],{k,c}] +sum[t[i,c]*t[j,d]*v[a,b,c,d],{c,d}] +sum[t[i,j,c,d]*v[a,b,c,d],{c,d}] +sum[t[j,c]*v[a,b,i,c],{c}] -sum[t[k,b]*v[a,k,i,j],{k}] +sum[t[i,c]*v[b,a,j,c],{c}] -sum[t[k,a]*v[b,k,j,i],{k}] -sum[t[k,d]*t[i,j,c,b]*v[k,a,c,d],{k,c,d}] -sum[t[i,c]*t[j,k,b,d]*v[k,a,c,d],{k,c,d}] -sum[t[j,c]*t[k,b]*v[k,a,c,i],{k,c}] +2*sum[t[j,k,b,c]*v[k,a,c,i],{k,c}] -sum[t[j,k,c,b]*v[k,a,c,i],{k,c}] -sum[t[i,c]*t[j,d]*t[k,b]*v[k,a,d,c],{k,c,d}] +2*sum[t[k,d]*t[i,j,c,b]*v[k,a,d,c],{k,c,d}] -sum[t[k,b]*t[i,j,c,d]*v[k,a,d,c],{k,c,d}] -sum[t[j,d]*t[i,k,c,b]*v[k,a,d,c],{k,c,d}] +2*sum[t[i,c]*t[j,k,b,d]*v[k,a,d,c],{k,c,d}] -sum[t[i,c]*t[j,k,d,b]*v[k,a,d,c],{k,c,d}] -sum[t[j,k,b,c]*v[k,a,i,c],{k,c}] -sum[t[i,c]*t[k,b]*v[k,a,j,c],{k,c}] -sum[t[i,k,c,b]*v[k,a,j,c],{k,c}] -sum[t[i,c]*t[j,d]*t[k,a]*v[k,b,c,d],{k,c,d}] -sum[t[k,d]*t[i,j,a,c]*v[k,b,c,d],{k,c,d}] -sum[t[k,a]*t[i,j,c,d]*v[k,b,c,d],{k,c,d}] +2*sum[t[j,d]*t[i,k,a,c]*v[k,b,c,d],{k,c,d}] -sum[t[j,d]*t[i,k,c,a]*v[k,b,c,d],{k,c,d}] -sum[t[i,c]*t[j,k,d,a]*v[k,b,c,d],{k,c,d}] -sum[t[i,c]*t[k,a]*v[k,b,c,j],{k,c}] +2*sum[t[i,k,a,c]*v[k,b,c,j],{k,c}] -sum[t[i,k,c,a]*v[k,b,c,j],{k,c}] +2*sum[t[k,d]*t[i,j,a,c]*v[k,b,d,c],{k,c,d}] -sum[t[j,d]*t[i,k,a,c]*v[k,b,d,c],{k,c,d}] -sum[t[j,c]*t[k,a]*v[k,b,i,c],{k,c}] -sum[t[j,k,c,a]*v[k,b,i,c],{k,c}] -sum[t[i,k,a,c]*v[k,b,j,c],{k,c}] +sum[t[i,c]*t[j,d]*t[k,a]*t[l,b]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[k,b]*t[l,d]*t[i,j,a,c]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[k,a]*t[l,d]*t[i,j,c,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[k,a]*t[l,b]*t[i,j,c,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[j,c]*t[l,d]*t[i,k,a,b]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[j,d]*t[l,b]*t[i,k,a,c]*v[k,l,c,d],{k,l,c,d}] +sum[t[j,d]*t[l,b]*t[i,k,c,a]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,c]*t[l,d]*t[j,k,b,a]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,a]*t[j,k,b,d]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,b]*t[j,k,d,a]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,k,c,d]*t[j,l,b,a]*v[k,l,c,d],{k,l,c,d}] +4*sum[t[i,k,a,c]*t[j,l,b,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,k,c,a]*t[j,l,b,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,k,a,b]*t[j,l,c,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,k,a,c]*t[j,l,d,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,k,c,a]*t[j,l,d,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,c]*t[j,d]*t[k,l,a,b]*v[k,l,c,d],{k,l,c,d}] +sum[t[i,j,c,d]*t[k,l,a,b]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,j,c,b]*t[k,l,a,d]*v[k,l,c,d],{k,l,c,d}] -2*sum[t[i,j,a,c]*t[k,l,b,d]*v[k,l,c,d],{k,l,c,d}] +sum[t[j,c]*t[k,b]*t[l,a]*v[k,l,c,i],{k,l,c}] +sum[t[l,c]*t[j,k,b,a]*v[k,l,c,i],{k,l,c}] -2*sum[t[l,a]*t[j,k,b,c]*v[k,l,c,i],{k,l,c}] +sum[t[l,a]*t[j,k,c,b]*v[k,l,c,i],{k,l,c}] -2*sum[t[k,c]*t[j,l,b,a]*v[k,l,c,i],{k,l,c}] +sum[t[k,a]*t[j,l,b,c]*v[k,l,c,i],{k,l,c}] +sum[t[k,b]*t[j,l,c,a]*v[k,l,c,i],{k,l,c}] +sum[t[j,c]*t[l,k,a,b]*v[k,l,c,i],{k,l,c}] +sum[t[i,c]*t[k,a]*t[l,b]*v[k,l,c,j],{k,l,c}] +sum[t[l,c]*t[i,k,a,b]*v[k,l,c,j],{k,l,c}] -2*sum[t[l,b]*t[i,k,a,c]*v[k,l,c,j],{k,l,c}] +sum[t[l,b]*t[i,k,c,a]*v[k,l,c,j],{k,l,c}] +sum[t[i,c]*t[k,l,a,b]*v[k,l,c,j],{k,l,c}] +sum[t[j,c]*t[l,d]*t[i,k,a,b]*v[k,l,d,c],{k,l,c,d}] +sum[t[j,d]*t[l,b]*t[i,k,a,c]*v[k,l,d,c],{k,l,c,d}] +sum[t[j,d]*t[l,a]*t[i,k,c,b]*v[k,l,d,c],{k,l,c,d}] -2*sum[t[i,k,c,d]*t[j,l,b,a]*v[k,l,d,c],{k,l,c,d}] -2*sum[t[i,k,a,c]*t[j,l,b,d]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,c,a]*t[j,l,b,d]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,a,b]*t[j,l,c,d]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,c,b]*t[j,l,d,a]*v[k,l,d,c],{k,l,c,d}] +sum[t[i,k,a,c]*t[j,l,d,b]*v[k,l,d,c],{k,l,c,d}] +sum[t[k,a]*t[l,b]*v[k,l,i,j],{k,l}] +sum[t[k,l,a,b]*v[k,l,i,j],{k,l}] +sum[t[k,b]*t[l,d]*t[i,j,a,c]*v[l,k,c,d],{k,l,c,d}] +sum[t[k,a]*t[l,d]*t[i,j,c,b]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,d]*t[j,k,b,a]*v[l,k,c,d],{k,l,c,d}] -2*sum[t[i,c]*t[l,a]*t[j,k,b,d]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,c]*t[l,a]*t[j,k,d,b]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,j,c,b]*t[k,l,a,d]*v[l,k,c,d],{k,l,c,d}] +sum[t[i,j,a,c]*t[k,l,b,d]*v[l,k,c,d],{k,l,c,d}] -2*sum[t[l,c]*t[i,k,a,b]*v[l,k,c,j],{k,l,c}] +sum[t[l,b]*t[i,k,a,c]*v[l,k,c,j],{k,l,c}] +sum[t[l,a]*t[i,k,c,b]*v[l,k,c,j],{k,l,c}] +v[a,b,i,j]

  4. Tensor Contraction Engine • Automatic transformation from high-level specification • Chemist specifies computation in high-level mathematical form • Synthesis system transforms it to efficient parallel program • Code is tailored to target machine • Code can be optimized for specific molecules being modeled • Multi-institutional collaboration (OSU, LSU, U. Waterloo, ORNL, PNNL, U. Florida) • Prototypes of TCE are operational • a) Full functionality but fewer optimizations (So Hirata, PNNL), b) Partial functionality but more sophisticated optimizations • Already used to implement new models, included in latest release of NWChem • First parallel implementation for many of the methods • Significant interest in quantum chemistry community • Built on top of the Global Arrays parallel library range V = 3000; range O = 100; index a,b,c,d,e,f : V; index i,j,k : O; mlimit = 10000000; function F1(V,V,V,O); function F2(V,V,V,O); procedure P(in T1[O,O,V,V], in T2[O,O,V,V], out X)= begin A3A == sum[ sum[F1(a,b,e,k) * F2(c,f,b,k), {b,k}] * sum[T1[i,j,c,e] * T2[i,j,a,f], {i,j}], {a,e,c,f}]*0.5 + ...; end

  5. Physically distributed data Global Address Space Global Arrays Library Distributed dense arrays that can be accessed through a shared view single, shared data structure with global indexing e.g.,accessA(4,3) rather than Alocal(1,3) on process 4

  6. Shared Object copy to shared object put compute/update local memory local memory local memory Global Array Model of Computations Shared Object copy to local memory get • Shared memory view for distributed dense arrays • MPI-Compatible; Currently usable with Fortran, C, C++, Python • Data locality and granularity control similar to message passing model • Used in large scale efforts, e.g. NWChem (million+ lines/code)

  7. Parochiality 3 P’s of Parallel Programming ParadigmsPerformance Productivity Parochiality “Holy Grail” TCE GA MPI OpenMP Performance Autoparallelized Productivity C/Fortran90

  8. Pros and Cons of the GA Model • Advantages • Provides convenient global-shared view, but the get-compute-put model ensures user focus on data-locality optimization => good performance • Inter-operates with MPI to enable general applicability • Limitations • Only supports dense multi-dimensional arrays • Data view more convenient than MPI, but computation specification is still “process-centric” • No support for load balancing of irregular computations

  9. Main Ideas in Approach Proposed • Decoupled task and data abstractions • Layered, multi-level data abstraction • Global-shared view highly desirable, but efficient word-level access to shared data infeasible (latency will always be much larger than 1/bandwidth) => chunked access • Lowest abstraction layer: Globally addressable “chunk” repository • Multi-dimensional block-sparse array is represented as a collection of small dense multi-dimensional bricks (chunks) • Bricks are distributed among processors; globally addressable and efficiently accessible • Computation Abstraction • Non-process-specific collection of independent tasks • Only non-local data task can access: parts of global arrays (patches of dense GA’s or bricks of block-sparse arrays) • Task-data affinity is modeled as a hypergraph; hypergraph partitioning used to achieve • Locality-aware load balancing • Transparent access to in-memory versus disk-resident data

  10. Hypergraph Formulation for Task Mapping • Vertices • One per task, weight = computation cost • Hyper-edges • One per data brick, weight = size of brick (Communication cost) • Connects tasks that access brick • Hypergraph partitioning to determine task mapping • Pre-determined distribution for data bricks (e.g. round-robin on hash) • Zero-weight vertex for each data brick, “pinned” to predetermined processor

  11. Example C = A * B Array B 3 3 3 3 3 2 3 3 3 Array C 3 3 2 Array A

  12. * = Tasks: [c-brick, a-brick, b-brick] [0x0, 0x0, 0x0] [0x0, 0x1, 1x0] [0x1, 0x0, 0x1] [0x1, 0x1, 1x1] [1x0, 1x0, 0x0] [1x0, 1x1, 1x0] [1x1, 1x0, 0x1] [1x1, 1x1, 1x1] [4x4, 4x2, 2x4] [4x4, 4x3, 3x4] [2x2, 2x4, 4x2] [2x3, 2x4, 4x3] [3x2, 3x4, 4x2] [3x3, 3x4, 4x3] [5x5, 5x5, 5x5]

  13. 27 27 [0x1, 0x0, 0x1] [0x0, 0x0, 0x0] 0 72 0 0x1 0x1 72 0x0 27 0x0 [0x1, 0x1, 1x1] 27 [0x0, 0x1, 1x0] 27 27 0 [4x4, 4x2, 2x4] [1x1, 1x0, 0x1] 4x4 0 72 72 1x1 1x1 4x4 27 [1x1, 1x1, 1x1] 27 27 [1x0, 1x0, 0x0] [4x4, 4x3, 3x4] 0 72 0 1x0 1x0 27 72 27 2x2 2x2 [2x2, 2x4, 4x2] [1x0, 1x1, 1x0] 0 3x3 0 0 72 27 5x5 2x3 2x3 [2x3, 2x4, 4x3] 32 72 5x5 3x3 0 27 72 8 27 3x2 3x2 [3x2, 3x4, 4x2] [5x5, 5x5, 5x5] [3x3, 3x4, 4x3]

  14. Transparent Memory Hierarchy Management • Problem: Schedule computation and disk I/O operations • Objective: Minimize disk I/O • Constraint: Memory limit constraint • Solution: Hypergraph partitioning formulation • Efficient solutions to the hypergraph problem exist • Typically used in the context of parallelization • Number of parts known • No constraints such as memory limit • Only balancing constraint

  15. Hypergraph Formulation for Memory Management • Formulation • Tasks -> vertices • Data -> Nets • No pre-assignment of nets to certain parts • Balance: Memory usage in the parts • Guarantees solution for some #parts, if it exists • Determine dynamically #parts • Modify the inherent recursive procedure of hypergraph partitioning.

  16. Read-Once Partitioning • Direct solution to above problem • Similar to approaches to parallelization • No refined reuse relationships • All tasks within a part have reuse, and none outside • Read-Once Partitioning • Group tasks into steps • Identify data common across steps and load into memory • For each step, read non-common (step-exclusive) data, process tasks, and write/discard step-exclusive data • Better utilization of memory available -> reduced disk I/O

  17. Read-Once Partitioning: Example Disk I/O: 8 data elements Disk I/O: 9 data elements

  18. Related Work • Sparse or block-sparse matrices in parallel libraries supporting sparse linear algebra • Aztec, PETSc etc. • Load-balancing • Work-stealing - Cilk • Object migration – Charm++ • Dynamic scheduling of loop iterations – OpenMP • Unaware of runtime systems support for: • Flexible global-shared abstractions for semi-structured data • Locality-aware, load-balanced scheduling of parallel tasks • Transparent access to in-memory and disk-resident data

  19. Summary and Future Work • Summary • Global-shared abstraction and distributed implementation on clusters, for multi-dimensional block-sparse arrays • Locality-aware load balancing of tasks operating on block-sparse arrays • Ongoing/Future Work • Improve performance through overlapping communication with computation • Transparent extension of global-shared data/computation abstraction for “out-of-core” case • Extension of the approach for oct-tree and other data structures – target application: multi-wavelet quantum chemistry code recently developed by Robert Harrison at ORNL

More Related