Difference between revisions of "QMCChem"

From Qmcchem
Jump to navigation Jump to search
 
(9 intermediate revisions by the same user not shown)
Line 1: Line 1:
−
One of our main activities is the development of the massively parallel QMC code QMC=Chem.
+
During the last years we have been actively developing the QMC=Chem quantum Monte Carlo code. This code was initially designed for massively parallel simulations, and uses the manager/worker model.
 +
 
 +
[[File:manager_worker.png|center]]
 +
 
 +
When the program starts its execution, the manager runs on the master node and spawns two other processes: a worker process and a data server. The worker is an efficient Fortran executable with minimal memory and disk space requirements (typically a few megabytes for each), where the only MPI communication is the broadcast of the input data (wave function parameters, initial positions in the 3N-space and random seed). The outline of the task of a worker is the following:
 +
 
 +
  while ( Running )
 +
  {
 +
    compute_a_block_of_data();
 +
    Running = send_the_results_to_the_data_server();
 +
  }
 +
 
 +
The data server is a socket server implemented in Python. When it receives the computed data of a worker, it replies to the worker the order given by the manager to compute another block or to stop. The received data is then stored in a database using an asynchronous I/O mechanism. The manager is always aware of the results computed by all the workers and controls the running/stopping state of the workers and the interaction of the user during the simulation.
 +
 
 +
QMC=Chem is very well suited to massive parallelism and cloud computing:
 +
* All the implemented algorithms are CPU-bound
 +
* All workers are totally independent
 +
* The load balancing is optimal: the workers always work 100% of the time, independently of their respective CPU speeds
 +
* The code was written to be as portable as possible: the manager is written in standard Python and the worker is written in standard Fortran.
 +
* The network traffic is minimal and the amount of data transferred over the network can even be adjusted by the user
 +
* The number of simultaneous worker nodes can be variable during a calculation
 +
* Fault-tolerance is implemented
 +
* The input and output data are not presented as traditional input files and output files. All the input and output data are stored in a database and an API is provided to access the data. This allows different forms of interaction of the user: scripts, graphical user interfaces, command-line tools, web interfaces, etc.
  
 
== Current Features ==
 
== Current Features ==
 +
 +
[[File:screenshot.png|center|500px]]
  
 
=== Methods ===
 
=== Methods ===
 
* VMC
 
* VMC
−
* DMC
+
* DMC with stochastic reconfiguration
 
* Jastrow factor optimization
 
* Jastrow factor optimization
 
* CI coefficients optimization
 
* CI coefficients optimization
Line 12: Line 36:
 
* Single determinant
 
* Single determinant
 
* Multi-determinant
 
* Multi-determinant
−
* Multi-Jastrow
 
 
* Nuclear cusp correction
 
* Nuclear cusp correction
  
−
=== Properties ===
+
=== Links ===
−
* [[The Electron Pair Localization Function]]
 
−
 
 
−
=== Practical aspects ===
 
−
* Very low memory requirements
 
−
* Runs on a large number of processors : tested on 512 processors at the [http://www.calmip.cict.fr/spip/spip.php?article281 CALMIP cluster], and on 1000 processors on the [http://www.eu-egee.org/ EGEE European grid].
 
−
* Checkpointing
 
−
* Fault tolerant : if the code is aborted, the data is kept. This feature is very useful in grid environments.
 
 
* Easy development with the [http://irpf90.ups-tlse.fr/ IRPF90 tool].
 
* Easy development with the [http://irpf90.ups-tlse.fr/ IRPF90 tool].
−
* Variable number of MPI processes during the calculation : can be changed at any time
+
* The access to input/output files is provided via an API produced by the [http://ezfio.sf.net  Easy Fortran I/O library generator]
 
* [{{SERVER}}/qmcchem_doc/QMC_Chem_Documentation.html Documentation]
 
* [{{SERVER}}/qmcchem_doc/QMC_Chem_Documentation.html Documentation]
  
Line 30: Line 46:
  
 
[[File:qmcchem_speedup.png|400px|left]]
 
[[File:qmcchem_speedup.png|400px|left]]
 +
Number of computed blocks for the CuCl2 molecule. The simulation stops when the wall-time has reached 5 minutes. Each block is composed of 10 walkers realizing 2,000 VMC steps.
  
−
{| border="1" align="center"
 
−
! Number of processors
 
−
! Number of computed blocks
 
−
! Speed-up
 
−
|-
 
−
|  1 ||  5  || 1
 
−
|-
 
−
|  8 ||  41 || 8
 
−
|-
 
−
|  16 ||  81 || 16
 
−
|-
 
−
|  32 || 170 || 32
 
−
|-
 
−
|  64 || 321 || 64
 
−
|-
 
−
| 128 || 644 || 128
 
−
|-
 
−
| 256 || 1280 || 256
 
−
|-
 
−
| 512 || 2417 || 483
 
−
|-
 
−
|+ Number of computed blocks in 140 seconds, where each block
 
−
is composed of 50 walkers realizing 10,000 VMC steps (water molecule).
 
−
|}
 
 
<br style="clear: both" />
 
<br style="clear: both" />
  
Line 63: Line 56:
 
* Moments (dipole, quadrupole,...)
 
* Moments (dipole, quadrupole,...)
 
* Electron density
 
* Electron density
−
* ZV-ZB EPLF estimator
 
  
 
=== Practical aspects ===
 
=== Practical aspects ===
Line 74: Line 66:
 
== Papers related to the QMC=Chem code ==
 
== Papers related to the QMC=Chem code ==
  
−
* [[Large-scale quantum Monte Carlo electronic structure calculations on the EGEE grid]]
+
* [http://dx.doi.org/10.1002/jcc.23216 Quantum Monte Carlo for large chemical systems: Implementing efficient strategies for petascale platforms and beyond]
 +
* [http://link.springer.com/chapter/10.1007/978-1-4614-0508-5_13 QMC=Chem: a quantum Monte Carlo program for large-scale simulations in chemistry at the petascale level and beyond]
 +
* [http://link.springer.com/chapter/10.1007/978-1-4614-0508-5_13 Large-scale quantum Monte Carlo electronic structure calculations on the EGEE grid]

Latest revision as of 12:03, 21 November 2014

During the last years we have been actively developing the QMC=Chem quantum Monte Carlo code. This code was initially designed for massively parallel simulations, and uses the manager/worker model.

Manager worker.png

When the program starts its execution, the manager runs on the master node and spawns two other processes: a worker process and a data server. The worker is an efficient Fortran executable with minimal memory and disk space requirements (typically a few megabytes for each), where the only MPI communication is the broadcast of the input data (wave function parameters, initial positions in the 3N-space and random seed). The outline of the task of a worker is the following: 

 while ( Running )
 {
    compute_a_block_of_data();
    Running = send_the_results_to_the_data_server();
 }

The data server is a socket server implemented in Python. When it receives the computed data of a worker, it replies to the worker the order given by the manager to compute another block or to stop. The received data is then stored in a database using an asynchronous I/O mechanism. The manager is always aware of the results computed by all the workers and controls the running/stopping state of the workers and the interaction of the user during the simulation.

QMC=Chem is very well suited to massive parallelism and cloud computing:

  • All the implemented algorithms are CPU-bound
  • All workers are totally independent
  • The load balancing is optimal: the workers always work 100% of the time, independently of their respective CPU speeds
  • The code was written to be as portable as possible: the manager is written in standard Python and the worker is written in standard Fortran.
  • The network traffic is minimal and the amount of data transferred over the network can even be adjusted by the user
  • The number of simultaneous worker nodes can be variable during a calculation
  • Fault-tolerance is implemented
  • The input and output data are not presented as traditional input files and output files. All the input and output data are stored in a database and an API is provided to access the data. This allows different forms of interaction of the user: scripts, graphical user interfaces, command-line tools, web interfaces, etc.

Current Features

Screenshot.png

Methods

  • VMC
  • DMC with stochastic reconfiguration
  • Jastrow factor optimization
  • CI coefficients optimization

Wave functions

  • Single determinant
  • Multi-determinant
  • Nuclear cusp correction

Links

Parallel speed-up curve

Qmcchem speedup.png

Number of computed blocks for the CuCl2 molecule. The simulation stops when the wall-time has reached 5 minutes. Each block is composed of 10 walkers realizing 2,000 VMC steps.


Features under development

Properties

  • Molecular Forces
  • Moments (dipole, quadrupole,...)
  • Electron density

Practical aspects

  • Web interface for input and output

Input file creation

The QMC=Chem input file can be created using the web interface. Upload a Q5Cost file or an output file from GAMESS, Gaussian or Molpro, and you will download the QMC=Chem input directory.

Papers related to the QMC=Chem code

Retrieved from "http://qmcchem.ups-tlse.fr/index.php?title=QMCChem&oldid=1846"