Parallel Execution and Global Variables
Hi, does anyone know how parallel execution works with global variables? In random_walk_metropolis_hastings.m we define at the beginning global objective_function_penalty_base objective_function_penalty_base = Inf; How is this variablem treated in parallel execution. In particular, what happens if one of the instances/nodes changes this value for objective_function_penalty_base? Does this affect all other nodes? Or does each node keep its own instance of the global variable? I would be grateful for any insights. Best, Johannes -- ------ Johannes Pfeifer Department of Economics University of Mannheim L7, 3-5, Room 242 68131 Mannheim Germany +49 (0)621 181-3430
Hi Johannes, My understanding is that each node keeps its own instance of the globals (we run different matlab sessions that do not communicate). Also, if I remember correctly, the global variable storing the penalty, is not updated in the metropolis (it should always be equal to Inf). This variable is used for the optimization (csminwel) algorithm which is not parallelized. Best, Stéphane. -- Stéphane Adjemian Université du Maine, Gains & Dynare Team
Le 27 avr. 2015 à 15:55, Johannes Pfeifer <pfeifer@uni-mannheim.de> a écrit :
Hi, does anyone know how parallel execution works with global variables? In random_walk_metropolis_hastings.m we define at the beginning
global objective_function_penalty_base objective_function_penalty_base = Inf;
How is this variablem treated in parallel execution. In particular, what happens if one of the instances/nodes changes this value for objective_function_penalty_base? Does this affect all other nodes? Or does each node keep its own instance of the global variable?
I would be grateful for any insights.
Best,
Johannes
-- ------ Johannes Pfeifer Department of Economics University of Mannheim L7, 3-5, Room 242 68131 Mannheim Germany +49 (0)621 181-3430
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
That makes sense, but I can not see this anywhere from the code. In the parallel branch of random_walk_metropolis_hastings.m , we have globalVars = struct(); But nothing else is passed. In masterParallel.m we seem to call different independent Matlab sessions that do not communicate. But I have no clue where they get a value for objective_function_penalty_base from. Am 27.04.2015 um 16:19 schrieb Stéphane Adjemian:
Hi Johannes,
My understanding is that each node keeps its own instance of the globals (we run different matlab sessions that do not communicate). Also, if I remember correctly, the global variable storing the penalty, is not updated in the metropolis (it should always be equal to Inf). This variable is used for the optimization (csminwel) algorithm which is not parallelized.
Best, Stéphane.
-- Stéphane Adjemian Université du Maine, Gains & Dynare Team
Le 27 avr. 2015 à 15:55, Johannes Pfeifer <pfeifer@uni-mannheim.de> a écrit :
Hi, does anyone know how parallel execution works with global variables? In random_walk_metropolis_hastings.m we define at the beginning
global objective_function_penalty_base objective_function_penalty_base = Inf;
How is this variablem treated in parallel execution. In particular, what happens if one of the instances/nodes changes this value for objective_function_penalty_base? Does this affect all other nodes? Or does each node keep its own instance of the global variable?
I would be grateful for any insights.
Best,
Johannes
-- ------ Johannes Pfeifer Department of Economics University of Mannheim L7, 3-5, Room 242 68131 Mannheim Germany +49 (0)621 181-3430
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
Hi Johannes. All global and local variables are wrapped and sent to workers by masterParallel.m. then only the variables that are needed co complete random_walk_metropolis_hastings.m are updated after parallel execution. best m -- Marco Ratto, Financial and Economic Analysis Joint Research Centre The European Commission, TP 361, 21027 ISPRA(VA), ITALY Tel: +39 0332 78 3794 Fax: +39 0332 78 5752, marco.ratto@jrc.ec.europa.eu http://www.macfinrobods.eu/ On 4/27/2015 4:29 PM, Johannes Pfeifer wrote:
That makes sense, but I can not see this anywhere from the code. In the parallel branch of random_walk_metropolis_hastings.m , we have
globalVars = struct();
But nothing else is passed. In masterParallel.m we seem to call different independent Matlab sessions that do not communicate. But I have no clue where they get a value for objective_function_penalty_base from.
Am 27.04.2015 um 16:19 schrieb Stéphane Adjemian:
Hi Johannes,
My understanding is that each node keeps its own instance of the globals (we run different matlab sessions that do not communicate). Also, if I remember correctly, the global variable storing the penalty, is not updated in the metropolis (it should always be equal to Inf). This variable is used for the optimization (csminwel) algorithm which is not parallelized.
Best, Stéphane.
-- Stéphane Adjemian Université du Maine, Gains & Dynare Team
Le 27 avr. 2015 à 15:55, Johannes Pfeifer<pfeifer@uni-mannheim.de> a écrit :
Hi, does anyone know how parallel execution works with global variables? In random_walk_metropolis_hastings.m we define at the beginning
global objective_function_penalty_base objective_function_penalty_base = Inf;
How is this variablem treated in parallel execution. In particular, what happens if one of the instances/nodes changes this value for objective_function_penalty_base? Does this affect all other nodes? Or does each node keep its own instance of the global variable?
I would be grateful for any insights.
Best,
Johannes
-- ------ Johannes Pfeifer Department of Economics University of Mannheim L7, 3-5, Room 242 68131 Mannheim Germany +49 (0)621 181-3430
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
_______________________________________________ Dev mailing list Dev@dynare.org https://www.dynare.org/cgi-bin/mailman/listinfo/dev
participants (4)
-
Johannes Pfeifer -
Johannes Pfeifer -
Marco Ratto -
Stéphane Adjemian