# Benchmarks

### From NWChem

(→Current developments for high accuracy: alternative task schedulers) |
(→Non-iterative part of the CCSD(T) approach: Comparing Xeon Phi and NVidia K20X performance) |
||

(22 intermediate revisions not shown) | |||

Line 7: | Line 7: | ||

=Hybrid density functional calculation on the C<sub>240</sub> Buckyball= | =Hybrid density functional calculation on the C<sub>240</sub> Buckyball= | ||

- | Performance of the Gaussian basis set DFT module in NWChem. This calculation involved performing a PBE0 calculation (in direct mode) on the on C<sub>240</sub> system with the 6-31G* basis set (3600 basis functions) without symmetry. These calculations were performed on the | + | Performance of the Gaussian basis set DFT module in NWChem. This calculation involved performing a PBE0 calculation (in direct mode) on the on C<sub>240</sub> system with the 6-31G* basis set (3600 basis functions) without symmetry. These calculations were performed on the Cascade supercomputer located at PNNL. The [[Media:input_c240_pbe0.nw|input file]] is available. |

- | [[File: | + | [[File:c240_web4.png|center|300px]] |

=Parallel performance of ''Ab initio'' Molecular Dynamics using plane waves= | =Parallel performance of ''Ab initio'' Molecular Dynamics using plane waves= | ||

Line 23: | Line 23: | ||

See the [[Media:input_gfpc.nw| input file]] for details. | See the [[Media:input_gfpc.nw| input file]] for details. | ||

- | [[File:creomccsd_t.png|center| | + | [[File:creomccsd_t.png|center|500px| ]] |

- | And more recent scalability test of the CR-EOMCCSD(T) formalism (Jaguar Cray XT5 at ORNL, see K. Kowalski, S. Krishnamoorthy, R.M. Olson, V. Tipparaju, E. | + | And more recent scalability test of the CR-EOMCCSD(T) formalism (Jaguar Cray XT5 at ORNL, see K. Kowalski, S. Krishnamoorthy, R.M. Olson, V. Tipparaju, E. Aprà , |

SC2011, for details). | SC2011, for details). | ||

- | [[File:curve_bg_lumo.png|center| | + | [[File:curve_bg_lumo.png|center|500px]] |

- | + | ||

=Parallel performance of the multireference coupled cluster (MRCC) methods = | =Parallel performance of the multireference coupled cluster (MRCC) methods = | ||

In collaboration with Dr. Jiri Pittner's group from Heyrovsky Institute of Physical Chemistry implementations of two variants of state-specific MRCC approaches | In collaboration with Dr. Jiri Pittner's group from Heyrovsky Institute of Physical Chemistry implementations of two variants of state-specific MRCC approaches | ||

- | have been developed. During his internship at PNNL Jirka Brabec, using novel processor-group-based algorithms, implemented Brillouin-Wigner and Mukherjee MRCC models with singles and doubles. The scalabililty tests for the Brillouin-Wigner MRCCSD approach have been performed on Jaguar XT5 system at ORNL for <math>\beta </math>-carotene in 6-31 basis set (472 orbitals, 216 correlated electrons, 20 reference functions; see J.Brabec, J. Pittner, H.J.J. van Dam, E. | + | have been developed. During his internship at PNNL Jirka Brabec, using novel processor-group-based algorithms, implemented Brillouin-Wigner and Mukherjee MRCC models with singles and doubles. The scalabililty tests for the Brillouin-Wigner MRCCSD approach have been performed on Jaguar XT5 system at ORNL for <math>\beta </math>-carotene in 6-31 basis set (472 orbitals, 216 correlated electrons, 20 reference functions; see J.Brabec, J. Pittner, H.J.J. van Dam, E. Aprà, K. Kowalski, JCTC |

2012, 8(2), pp 487–497). Currently, PNNL postdoctoral fellow Dr. Kiran Bhaskaran Nair is developing perturbative MRCCSD(T) approaches, which accounts for the effect of triple excitations. | 2012, 8(2), pp 487–497). Currently, PNNL postdoctoral fellow Dr. Kiran Bhaskaran Nair is developing perturbative MRCCSD(T) approaches, which accounts for the effect of triple excitations. | ||

- | [[File:mrccsd_scalability.png|center| | + | [[File:mrccsd_scalability.png|center|500px]] |

Line 44: | Line 43: | ||

BW-MRCCSD(T) implementation of NWChem have been performed on the Jaguar Cray-XK6 computer system of the National Center for Computational Sciences at Oak Ridge National Laboratory. | BW-MRCCSD(T) implementation of NWChem have been performed on the Jaguar Cray-XK6 computer system of the National Center for Computational Sciences at Oak Ridge National Laboratory. | ||

- | [[File:Mrccsd_t_titan.png|center| | + | [[File:Mrccsd_t_titan.png|center|500px]] |

=Timings of CCSD/EOMCCSD for the oligoporphyrin dimer = | =Timings of CCSD/EOMCCSD for the oligoporphyrin dimer = | ||

Line 76: | Line 75: | ||

=Performance tests of the GPU implementation of non-iterative part of the CCSD(T) approach= | =Performance tests of the GPU implementation of non-iterative part of the CCSD(T) approach= | ||

- | Recent tests of the GPU CCSD(T) implementation performed on Titan Cray XK7 system at ORNL | + | Recent tests of the GPU CCSD(T) implementation performed on Titan Cray XK7 [https://www.olcf.ornl.gov/computing-resources/titan-cray-xk7/] system at ORNL |

- | ( | + | (C<sub>22</sub>H<sub>14</sub>, 378 basis set functions, C1 symmetry; 98 nodes: 8 cores per node + 1GPU) |

+ | '''Using 8 CPU cores''' | ||

Using CUDA CCSD(T) code | Using CUDA CCSD(T) code | ||

Using 0 device(s) per node | Using 0 device(s) per node | ||

Line 89: | Line 89: | ||

Cpu & wall time / sec 9229.9 9240.3 | Cpu & wall time / sec 9229.9 9240.3 | ||

- | + | '''Using 7 CPU cores and one GPU''' | |

Using CUDA CCSD(T) code | Using CUDA CCSD(T) code | ||

Using 1 device(s) per node | Using 1 device(s) per node | ||

Line 100: | Line 100: | ||

Cpu & wall time / sec 1468.0 1630.7 | Cpu & wall time / sec 1468.0 1630.7 | ||

+ | '''Using 1 CPU core and one GPU''' | ||

+ | Using CUDA CCSD(T) code | ||

+ | Using 1 device(s) per node | ||

+ | CCSD[T] correction energy / hartree = -0.150973754993069 | ||

+ | CCSD[T] correlation energy / hartree = -3.067917061063028 | ||

+ | CCSD[T] total energy / hartree = -844.*************** | ||

+ | CCSD(T) correction energy / hartree = -0.147996460406749 | ||

+ | CCSD(T) correlation energy / hartree = -3.064939766476708 | ||

+ | CCSD(T) total energy / hartree = -844.400399501861216 | ||

+ | Cpu & wall time / sec 1410.9 1756.5 | ||

Without GPU 9240.3 sec. | Without GPU 9240.3 sec. | ||

Line 106: | Line 116: | ||

Next release: GPU implementation of non-iterative part of the MRCCSD(T) approach | Next release: GPU implementation of non-iterative part of the MRCCSD(T) approach | ||

(K. Bhaskarsan-Nair, W. Ma, S. Krishnamoorthy, O. Villa, H. van Dam, | (K. Bhaskarsan-Nair, W. Ma, S. Krishnamoorthy, O. Villa, H. van Dam, | ||

- | E. | + | E. Aprà, K. Kowalski, J. Chem. Theory Comput. 9, 1949 (2013)) |

- | [[File:GPU_MRCCSD_T.png|center| | + | [[File:GPU_MRCCSD_T.png|center|500px]] |

- | =Current developments for high accuracy: alternative task schedulers= | + | =Performance tests of the Xeon Phi implementation of non-iterative part of the CCSD(T) approach= |

+ | |||

+ | Recent tests (January 2015) of the Xeon Phi CCSD(T) implementation performed on EMSL cascade [http://www.emsl.pnl.gov/emslweb/instruments/computing-cascade-atipa-1440-intel-xeon-phi-node-fdr-infiniband-linux-cluster] system at PNNL | ||

+ | |||

+ | [[File:Triplet_updated_feb2015.png|center|728px]] | ||

+ | |||

+ | Aprà, E.; Klemm, M.; Kowalski, K., "Efficient Implementation of Many-Body Quantum Chemical Methods on the Intel® Xeon Phi Coprocessor," High Performance Computing, Networking, Storage and Analysis, SC14: International Conference for , vol., no., pp.674,684, 16-21 Nov. 2014 | ||

+ | [http://dx.doi.org/10.1109/SC.2014.60] | ||

+ | |||

+ | (Triplet state of Si<sub>4</sub>C<sub>3</sub>N<sub>2</sub>H<sub>12</sub>, 706 basis set functions, C1 symmetry) | ||

+ | |||

+ | = Non-iterative part of the CCSD(T) approach: Comparing Xeon Phi and NVidia K20X performance = | ||

+ | |||

+ | Wall time to solution (in seconds) of non-iterative triples part of the single-reference CCSD(T) approach for the pentacene molecule using Intel MIC and Nvidia GPU implementations. | ||

+ | Tests were performed using 96 compute nodes on the Cascade system at EMSL (Intel® Xeon™ Phi 5110P) and Titan system at ORNL (NVIDIA Tesla® K20X). | ||

+ | |||

+ | ([[Media:Pentacene_ccsdt.nw| input file]]) | ||

+ | |||

+ | {| border='1' cellspacing='0' | ||

+ | |- | ||

+ | | Tilesize | ||

+ | |Intel Xeon Phi 5110P | ||

+ | |Nvidia K20X | ||

+ | |- | ||

+ | | 18 | ||

+ | | 1806.4 | ||

+ | | 1824.9 | ||

+ | |- | ||

+ | | 21 | ||

+ | | 1652.2 | ||

+ | | 1699.3 | ||

+ | |- | ||

+ | | 24 | ||

+ | | 1453.3 | ||

+ | | 1554.4 | ||

+ | |- | ||

+ | |} | ||

+ | |||

+ | =Current developments for high accuracy: alternative task schedulers (ATS)= | ||

Currently various development efforts are underway for high accuracy methods that will be available in future releases of NWChem. The examples below shows the first results of the performance of the triples part of Reg-CCSD(T) on GPGPUs (left two examples) and of using alternative task schedules for the iterative CCSD and EOMCCSD. | Currently various development efforts are underway for high accuracy methods that will be available in future releases of NWChem. The examples below shows the first results of the performance of the triples part of Reg-CCSD(T) on GPGPUs (left two examples) and of using alternative task schedules for the iterative CCSD and EOMCCSD. | ||

Line 125: | Line 173: | ||

- | + | ||

+ | Other tests: | ||

+ | |||

+ | The impact of the tilesize on the CCSD(ATS) timings: All tests have been performed for uracil trimer (6-31G* basis set; all core electrons frozen) | ||

+ | on Hopper using 25 nodes (600 cores). One can observe almost 10-fold speedup of the CCSD(ATS) code for tilesize=40 compared to standard TCE CCSD implementation | ||

+ | using tilesize=12. | ||

+ | |||

+ | [[File:uracil_trimer.png|center|550px]] | ||

+ | |||

+ | |||

+ | Performance tests for water clusters | ||

+ | |||

+ | [[File:CTF.png|center|550px]] | ||

- | |||

Luciferin (aug-cc-pVDZ basis set; RHF reference; frozen core) - time per CCSD iteration ([[Media:luciferin.nw| input file]]) | Luciferin (aug-cc-pVDZ basis set; RHF reference; frozen core) - time per CCSD iteration ([[Media:luciferin.nw| input file]]) |

## Latest revision as of 18:29, 7 January 2016

# Benchmarks performed with NWChem

This page contains a suite of benchmarks performed with NWChem. The benchmarks include a variety of computational chemistry methods on a variety of high performance computing platforms. The list of benchmarks available will evolve continuously as new data becomes available. If you have benchmark information you would like to add for your computing system, please contact one of the developers.

# Hybrid density functional calculation on the C_{240} Buckyball

Performance of the Gaussian basis set DFT module in NWChem. This calculation involved performing a PBE0 calculation (in direct mode) on the on C_{240} system with the 6-31G* basis set (3600 basis functions) without symmetry. These calculations were performed on the Cascade supercomputer located at PNNL. The input file is available.

# Parallel performance of *Ab initio* Molecular Dynamics using plane waves

# Parallel performance of the CR-EOMCCSD(T) method (triples part)

An example of the scalability of the triples part of the CR-EOMCCSD(T) approach for Green Fluorescent Protein Chromophore (GFPC) described by cc-pVTZ basis set (648 basis functions) as obtained from NWChem. Timings were determined from calculations on the Franklin Cray-XT4 computer system at NERSC. See the input file for details.

And more recent scalability test of the CR-EOMCCSD(T) formalism (Jaguar Cray XT5 at ORNL, see K. Kowalski, S. Krishnamoorthy, R.M. Olson, V. Tipparaju, E. Aprà ,
SC2011, for details).

# Parallel performance of the multireference coupled cluster (MRCC) methods

In collaboration with Dr. Jiri Pittner's group from Heyrovsky Institute of Physical Chemistry implementations of two variants of state-specific MRCC approaches have been developed. During his internship at PNNL Jirka Brabec, using novel processor-group-based algorithms, implemented Brillouin-Wigner and Mukherjee MRCC models with singles and doubles. The scalabililty tests for the Brillouin-Wigner MRCCSD approach have been performed on Jaguar XT5 system at ORNL for β-carotene in 6-31 basis set (472 orbitals, 216 correlated electrons, 20 reference functions; see J.Brabec, J. Pittner, H.J.J. van Dam, E. Aprà, K. Kowalski, JCTC 2012, 8(2), pp 487–497). Currently, PNNL postdoctoral fellow Dr. Kiran Bhaskaran Nair is developing perturbative MRCCSD(T) approaches, which accounts for the effect of triple excitations.

Scaling of the triples part of the BW-MRCCSD(T) method for β-carotene in 6-31 basis set (JCP 137, 094112 (2012)). The scalability tests of the
BW-MRCCSD(T) implementation of NWChem have been performed on the Jaguar Cray-XK6 computer system of the National Center for Computational Sciences at Oak Ridge National Laboratory.

# Timings of CCSD/EOMCCSD for the oligoporphyrin dimer

CCSD/EOMCCSD timings for oligoporphyrin dimer (942 basis functions, 270 correlated electrons, D2h symmetry, excited-state calculations were performed for state of b1g symmetry, in all test calculation convergence threshold was relaxed, 1024 cores were used). See the input file for details.

-------------------------------------------------------- Iter Residuum Correlation Cpu Wall -------------------------------------------------------- 1 0.7187071521175 -7.9406033677717 640.9 807.7 ...... MICROCYCLE DIIS UPDATE: 10 5 11 0.0009737920958 -7.9953441809574 691.1 822.2 -------------------------------------------------------- Iterations converged CCSD correlation energy / hartree = -7.995344180957357 CCSD total energy / hartree = -2418.570838364838890 EOM-CCSD right-hand side iterations -------------------------------------------------------------- Residuum Omega / hartree Omega / eV Cpu Wall -------------------------------------------------------------- ...... Iteration 2 using 6 trial vectors 0.1584284659595 0.0882389635508 2.40111 865.3 1041.2 Iteration 3 using 7 trial vectors 0.0575982107592 0.0810948687618 2.20670 918.0 1042.2

# Performance tests of the GPU implementation of non-iterative part of the CCSD(T) approach

Recent tests of the GPU CCSD(T) implementation performed on Titan Cray XK7 [1] system at ORNL
(C_{22}H_{14}, 378 basis set functions, C1 symmetry; 98 nodes: 8 cores per node + 1GPU)

**Using 8 CPU cores**

Using CUDA CCSD(T) code Using 0 device(s) per node CCSD[T] correction energy / hartree = -0.150973754992986 CCSD[T] correlation energy / hartree = -3.067917061062492 CCSD[T] total energy / hartree = -844.403376796441080 CCSD(T) correction energy / hartree = -0.147996460406684 CCSD(T) correlation energy / hartree = -3.064939766476190 CCSD(T) total energy / hartree = -844.400399501854849 Cpu & wall time / sec 9229.9 9240.3

**Using 7 CPU cores and one GPU**

Using CUDA CCSD(T) code Using 1 device(s) per node CCSD[T] correction energy / hartree = -0.150973754993019 CCSD[T] correlation energy / hartree = -3.067917061062597 CCSD[T] total energy / hartree = -844.403376796441307 CCSD(T) correction energy / hartree = -0.147996460406693 CCSD(T) correlation energy / hartree = -3.064939766476270 CCSD(T) total energy / hartree = -844.400399501854963 Cpu & wall time / sec 1468.0 1630.7

**Using 1 CPU core and one GPU**

Using CUDA CCSD(T) code Using 1 device(s) per node CCSD[T] correction energy / hartree = -0.150973754993069 CCSD[T] correlation energy / hartree = -3.067917061063028 CCSD[T] total energy / hartree = -844.*************** CCSD(T) correction energy / hartree = -0.147996460406749 CCSD(T) correlation energy / hartree = -3.064939766476708 CCSD(T) total energy / hartree = -844.400399501861216 Cpu & wall time / sec 1410.9 1756.5

Without GPU 9240.3 sec. With GPU 1630.7 sec.

Next release: GPU implementation of non-iterative part of the MRCCSD(T) approach (K. Bhaskarsan-Nair, W. Ma, S. Krishnamoorthy, O. Villa, H. van Dam, E. Aprà, K. Kowalski, J. Chem. Theory Comput. 9, 1949 (2013))

# Performance tests of the Xeon Phi implementation of non-iterative part of the CCSD(T) approach

Recent tests (January 2015) of the Xeon Phi CCSD(T) implementation performed on EMSL cascade [2] system at PNNL

Aprà, E.; Klemm, M.; Kowalski, K., "Efficient Implementation of Many-Body Quantum Chemical Methods on the Intel® Xeon Phi Coprocessor," High Performance Computing, Networking, Storage and Analysis, SC14: International Conference for , vol., no., pp.674,684, 16-21 Nov. 2014 [3]

(Triplet state of Si_{4}C_{3}N_{2}H_{12}, 706 basis set functions, C1 symmetry)

# Non-iterative part of the CCSD(T) approach: Comparing Xeon Phi and NVidia K20X performance

Wall time to solution (in seconds) of non-iterative triples part of the single-reference CCSD(T) approach for the pentacene molecule using Intel MIC and Nvidia GPU implementations. Tests were performed using 96 compute nodes on the Cascade system at EMSL (Intel® Xeon™ Phi 5110P) and Titan system at ORNL (NVIDIA Tesla® K20X).

( input file)

Tilesize | Intel Xeon Phi 5110P | Nvidia K20X |

18 | 1806.4 | 1824.9 |

21 | 1652.2 | 1699.3 |

24 | 1453.3 | 1554.4 |

# Current developments for high accuracy: alternative task schedulers (ATS)

Currently various development efforts are underway for high accuracy methods that will be available in future releases of NWChem. The examples below shows the first results of the performance of the triples part of Reg-CCSD(T) on GPGPUs (left two examples) and of using alternative task schedules for the iterative CCSD and EOMCCSD.

Other tests:

The impact of the tilesize on the CCSD(ATS) timings: All tests have been performed for uracil trimer (6-31G* basis set; all core electrons frozen) on Hopper using 25 nodes (600 cores). One can observe almost 10-fold speedup of the CCSD(ATS) code for tilesize=40 compared to standard TCE CCSD implementation using tilesize=12.

Performance tests for water clusters

Luciferin (aug-cc-pVDZ basis set; RHF reference; frozen core) - time per CCSD iteration ( input file)

tilesize = 30 256 cores 644 sec. 512 378 sec. 664 314 sec. 1020 278 sec. 1300 237 sec.

tilesize = 40 128 998 sec. 256 575 sec.

Sucrose (6-311G** basis set; RHF reference; frozen core) - time per CCSD iteration ( input file)

tilesize = 40 256 cores 1486 sec. 512 910 sec. 1024 608 sec.

Cytosine-OH (POL1; UHF reference; frozen core) - time per EOMCCSD iteration ( input file)

tilesize = 30 256 cores 44.5 sec.

tilesize = 40 128 cores 55.6 sec.