Parallel DYNARE Toolbox FP7 Funded Project MONFISPOL · PDF fileParallel DYNARE Toolbox FP7 Funded Project MONFISPOL Grant no.: 225149 ... Parallelize portions of code that require

Parallel DYNARE

Parallel DYNARE ToolboxFP7 Funded

Project MONFISPOL Grant no.: 225149

Marco RattoEuropean Commission, Joint Research Centre, Ispra, ITALY

with Ivano Azzini, Houtan Bastani, Sebastien Villemot

November 8, 2010

Parallel DYNARE

Outline

1 Introduction

2 The DYNARE environment

3 Installation and utilizationRequirementsThe user perspectiveThe Developers perspective

4 Parallel DYNARE: testing

5 Conclusions

Parallel DYNARE

Intro

Introduction

Parallelize portions of code that require a minimal (i.e. start-endcommunication) or no communications between differentprocesses, denoted in the literature as “embarrassingly parallel”(Goffe and Creel, 2008; Barney, 2009), with the following criteria:

provide the necessary input data to any sequence;

distribute the workload, automatically balancing between thecomputational resources;

collect the output data;

ensure the coherence of the results with the original serialexecution.

Parallel DYNARE

Intro

...

n=2;

m=10^6;

Matrix= zeros(n,m);

for i=1:n,

Matrix(i,:)=rand(1,m);

end,

Mse= Matrix;

...

Example 1

Parallel DYNARE

Intro

...

n=2;

m=10^6;

<provide to CPU1 and CPU2 input data m>

-

<Execute on CPU1> <Execute on CPU2>

Matrix1 = zeros(1,m); Matrix2 = zeros(1,m);

Matrix1(1,:)=rand(1,m); Matrix2(1,:)=rand(1,m);

save Matrix1 save Matrix2

-

retrieve Matrix1 and Matrix2

Mpe(1,:) = Matrix1;

Mpe(2,:) = Matrix2;

Example 2

Parallel DYNARE

Intro

The for cycle has disappeared and it has been split into twoseparated sequences that can be executed in parallel on two CPUs.We have the same result (Mpa=Mse) but the computational timecan be reduced up to 50%.

Parallel DYNARE

The DYNARE environment

The DYNARE environment IRoutines suitable for parallelization:

1 the RW Metropolis with multiple chains(random walk metropolis hastings.m)(independent metropolis hastings.m);

2 the diagnostic tests(McMCDiagnostics.m);

3 posterior IRF’s (posteriorIRF.m).

4 posterior statistics for filtered and smoothed variables,forecasts, smoothed shocks, etc..(prior posterior statistics.m).

5 plotting utility (pm3.m).

Parallel DYNARE

The DYNARE environment

The DYNARE environment II

MATLAB does not support multi-threads, without the use (andpurchase) of MATLAB Distributed Computing Toolbox.

When the execution should start in parallel:

1 pass the control to the operating system(Windows/Linux);

2 launched concurrent threads (i.e. MATLABinstances) on different processors/cores/machines;

3 when parallel is concluded the control is given backto the original MATLAB session .

Parallel DYNARE

Installation and utilization

Requirements

Requirements I

For a Windows grid:

1 a standard Windows network (SMB) must be in place;

2 PsTools (Russinovich, 2009) must be installed in the path ofthe master Windows machine;

3 the Windows user on the master machine has to be user ofany other slave machine in the cluster, and that user will beused for the remote computations.

Parallel DYNARE


Requirements

Requirements II

For a UNIX grid

1 SSH must be installed on the master and on the slavemachines;

2 the UNIX user on the master machine has to be user of anyother slave machine in the cluster, and that user will be usedfor the remote computations;

3 SSH keys must be installed so that the SSH connection fromthe master to the slaves can be done without passwords, orusing an SSH agent.

Parallel DYNARE


The user perspective

The interface I

Put the configuration of the cluster in a config file different fromthe MOD file, and to trigger the parallel computation withoption(s) on the dynare command line. The configuration file isdesigned as follows:

be in a standard location

$HOME/.dynare under Unix;c:\Documents and Setting\<username>\ApplicationData\dynare.ini on Windows;

have provisions for other Dynare configuration parametersunrelated to parallel computation

allow to specify several clusters, each one associated with anickname;

Parallel DYNARE



The interface IINode Options type default Win Unix

Local Remote Local Remote

Name string (stop) * * * *CPUnbr integer (stop) * * * *

or arrayComputerName string (stop) * *UserName string empty * *Password string empty *RemoteDrive string empty *RemoteDirectory string empty * *DynarePath string emptyMatlabOctavePath string emptySingleCompThread boolean true

Cluster Options type default Meaning Required

Name string empty name of the node *Members string empty list of members in this cluster *

Parallel DYNARE



Configuration file syntax (Unix)

[cluster]

Name = c1

Members = n1 n2 n3

[cluster]

Name = c2

Members = n2 n3

[node]

Name = n1

ComputerName = localhost

CPUnbr = 1

[node]

Name = n2

ComputerName = karaba.cepremap.org

CPUnbr = 5

UserName = houtanb

RemoteDirectory = /home/houtanb/Remote

DynarePath = /home/houtanb/dynare/matlab

MatlabOctavePath = matlab

[node]

Name = n3

ComputerName = hal.cepremap.ens.fr

CPUnbr = 3

UserName = houtanb

RemoteDirectory = /home/houtanb/Remote

DynarePath = /home/houtanb/dynare/matlab


Parallel DYNARE



Configuration file syntax (Win)

[cluster]

Name = local

Members = n1

[cluster]

Name = vonNeumann

Members = n2

[cluster]

Name = c2

Members = n1 n2

[node]

Name = n1

ComputerName = localhost

CPUnbr = 4

[node]

Name = n2

ComputerName = vonNeumann

CPUnbr = [4:6]

UserName = COMPUTOWN\John

Password = *****

RemoteDrive = C

RemoteDirectory = dynare_calcs\Remote

DynarePath = c:\dynare\matlab


Parallel DYNARE



DYNARE command line options

conffile=<path>: specify the location of the configurationfile if it is not standard

parallel: trigger the parallel computation using the firstcluster specified in config file

parallel=<clustername>: trigger the parallel computation,using the given cluster

parallel slave open mode: use the leaveSlaveOpen modein the cluster

parallel test: just test the cluster, dont actually run theMOD file

Parallel DYNARE


The Developers perspective

Core functions I

masterParallel is the entry point to the parallelization system:

It is called from the master computer, at thepoint where the parallelization system should beactivated;

all file exchange through the filesystem isconcentrated in this masterParallel routine;

there are two modes of parallel execution,triggered by optionparallel slave open mode;

slaveParallel.m/fParallel.m: are the top-level functions tobe run on every slave;

Parallel DYNARE



Core functions II

fMessageStatus.m: provides the core for simple message passingduring slave execution and typically replaces calls towaitbar.m;

closeSlave.m is the utility that sends a signal to remote slavesto close themselves;

Parallel DYNARE



Utilities

AnalyseComputationalEnviroment.m;

InitializeComputationalEnviroment.m;

distributeJobs.m;

generalized routines:

dynareParallelDelete.m : generalized delete;dynareParallelDir.m : generalized dir;dynareParallelGetFiles.m : generalized copy FROM slaves TO

master machine;dynareParallelMkDir.m : generalized mkdir on remote

machines;dynareParallelRmDir.m : generalized rmdir on remote

machined;dynareParallelSendFiles.m : generalized copy TO slaves

FROM master machine;

Parallel DYNARE

Parallel DYNARE: testing


We present here all tests performed with Windows Xp/Matlab.The model used for testing is a modification of Lubik (2003): asmall scale open economy DSGE model with 6 observed variables,6 endogenous variables and 19 parameters to be estimated.

Parallel DYNARE


Test 1.

Test 1

Bi-processor machine (Fujitsu Siemens, Celsius R630) powered withan Intel(R) Xeon(TM) CPU 2.80GHz Hyper Treading Technology.

Parallel DYNARE


Test 1.

3

Figure 1. X=Runs / Y=Computational time in minutes

Figure 2. X=Runs Y=Time Gain

0

0.1

0.2

0.3

0.4

0.5

0.6

0 200000 400000 600000 800000 1000000 1200000

Figure 2 plots the computation time gain against the chain length. The time gain

0

100

200

300

400

500

600

700

800

0 200000 400000 600000 800000 1000000 1200000

one processortwo processors

Figure: Computational time (in minutes) versus chain length for theserial and parallel implementation (Metropolis with two chains).

Parallel DYNARE


Test 1.

3

Figure 1. X=Runs / Y=Computational time in minutes

Figure 2. X=Runs Y=Time Gain

0

0.1

0.2

0.3

0.4

0.5

0.6

0 200000 400000 600000 800000 1000000 1200000

Figure 2 plots the computation time gain against the chain length. The time gain

0

100

200

300

400

500

600

700

800

0 200000 400000 600000 800000 1000000 1200000

one processortwo processors

Figure: Reduction of computational time (i.e. the ‘time gain’) using theparallel coding versus chain length. The time gain is computed as(Ts − Tp)/Tp, where Ts and Tp denote the computing time of the serialand parallel implementations respectively.

Parallel DYNARE


Test 2.

Test 2 I

Test different hardware platforms, with chain lengths of 1,000,000runs on the following hardware platforms:

Single processor machine: Intel(R) Pentium4(R) CPU3.40GHz with Hyper Treading Technology (Fujitsu-SiemensScenic Esprimo);

Bi-processor machine: two CPU’s Intel(R) Xeon(TM)2.80GHz Hyper Treading Technology (Fujitsu-Siemens,Celsius R630);

Dual core machine: Intel Centrino T2500 2.00GHz Dual Core(Fujitsu-Siemens, LifeBook S Series).

Parallel DYNARE


Test 2.

Test 2 II

Machine Single-processor Bi-processor Dual core

Parallel 8:01:21 7:02:19 5:39:38Serial 10:12:22 13:38:30 11:02:14Speed-Up rate 1.2722 1.9381 1.9498Ideal Speed-UP rate ∼1.5 2 2

Table: Trail results with normal PC operation. Computing time expressedin h:m:s. Speed-up rate is computed as Ts/Tp, where Ts and Tp are thecomputing times for the serial and parallel implementations.

Parallel DYNARE


Test 2.

Test 2 III

Environment Computing time Speed-up ratew.r.t. Table 1

Parallel Waitbar Not Visible 5:06:00 1.06

Parallel waitbar Not Visible,Real-time Process priority, Un-plugged network cable.

4:40:49 1.22

Table: Trail results with different software configurations (optimizedoperating environment for computational requirements).

Parallel DYNARE


Test 3

Test 3

We used a very simple computer class, quite diffuse nowadays:Netbook personal Computer. In particular we used the Dell Mini10 with Processor Intel Atom Z520 (1,33 GHz, 533 MHz), 1 GB diRAM (with Hyper-trading).

Parallel DYNARE


Test 3

8

Chains Length Time Serial Time Parallel 105 85 151 1005 246 287 5005 755 599 10005 1246 948 15005 1647 1250 20005 2068 1502 25005 2366 1675 Table3. Computational Time using all the parallel functions in DYNARE and the Open/Close strategy. We can also plot the results in table 3. We call this situation Complete Parallel …

Complete Parallel

0

500

1000

1500

2000

2500

105 1005 5005 10005 15005 20005 25005

MH Runs

Com

puta

tiona

l Tim

e (s

ec)

SerialParallel

Figure 3. The plot of data in Table 3. Figure 3. show as …

Figure: Computational Time (s) versus Metropolis length, running all theparallelized functions in DYNARE and the basic parallel implementation(the ‘Open/Close’ strategy). (Lubik, 2003).

Parallel DYNARE


Test 3

10

Figure 4. show as … Comments … As reported in … we also introduced a new computational Matlab instances are always open … Chain Lenght

Computational Times: Complete Parallel

Computational Times: Partial Parallel

105 103 951005 209 1225005 504 205

10005 915 30615005 1203 32020005 1506 33425005 1611 322

Table5. Computational Time with Always Open strategy. We can also plot and compare the results in table 5 with results in table 3 and 4.:

Open/Close vs AlwaysOpenComplete Parallel

0

200

400

600

800

1000

1200

1400

1600

1800

105 1005 5005 10005 15005 20005 25005

MH Runs

Com

puta

tiona

l Tim

e (s

ec)

Open/CloseAlways Open

Figure 5. The compared computational time for “complete” parallel . Figure: Comparison of the ‘Open/Close’ strategy and the ‘Always-open’strategy. Computational Time (s) versus Metropolis length, running allthe parallelized functions in DYNARE (Lubik, 2003).

Parallel DYNARE


Test 4

Tets 4 I

Here we increase the dimension of the test model, using theQUEST III model (Ratto et al., 2009), using a more powerfulNotebook Samsung Q 45 with an Dual core Processor IntelCentrino.

Parallel DYNARE


Test 4

12

Chains Length

Time Serial

Time Parallel

105 98 951005 398 2555005 1463 890

10005 2985 165520005 4810 281530005 6630 402240005 7466 524650000 9263 6565

Table6. Computational Time using all the parallel function involved and the Open/Close strategy.

Complete Parallel - Open/Close

0

1000

2000

3000

4000

5000

6000

7000

8000

9000

10000

105 1005 5005 10005 20005 30005 40005 50000

MH Runs

Com

puta

tiona

l Tim

e (s

ec)

SerialParallel

Chains Length

Comp Time Serial

Comp Time Parallel

Figure: Computational Time (s) versus Metropolis length, running all theparallelized functions in DYNARE and the basic parallel implementation(the ‘Open/Close’ strategy). (Ratto et al., 2009).

Parallel DYNARE


Test 4

14

105 66 601005 273 1175005 871 332

10005 1588 46020005 2791 47030005 3963 49240005 5292 47950000 6624 498

Table8. Computational Time with Always Open strategy.

Open/Close vs AlwaysOpenComplete Parallel

0

1000

2000

3000

4000

5000

6000

7000

105 1005 5005 10005 20005 30005 40005 50000

MH Runs

Com

puta

tiona

l Tim

e (s

ec)

Open/CloseAlways Open

Figure: Comparison of the ‘Open/Close’ strategy and the ‘Always-open’strategy. Computational Time (s) versus Metropolis length, running allthe parallelized functions in DYNARE (Ratto et al., 2009).

Parallel DYNARE

Conclusions

Conclusions

The methodology identified for parallelizing MATLAB codes withinDYNARE proved to be effective in reducing the computationaltime of the most extensive loops. This methodology is suitable for‘embarrassingly parallel’ codes, requiring only a minimalcommunication flow between slave and master threads. Theparallel DYNARE is built around a few ‘core’ routines, that act as asort of ‘parallel paradigm’. Based on those routines, parallelizationof expensive loops is made quite simple for DYNARE developers.A basic message passing system is also provided, that allows themaster thread to monitor the progress of slave threads. The testmodel ls2003.mod is available in the folder \tests\parallel ofthe DYNARE distribution, that allows running parallel examples.

Parallel DYNARE

Conclusions

Bibliography I

B Barney. Introducing To Parallel Computing (Tutorial). LawrenceLivermore National Laboratory, 2009.

W.L. Goffe and M. Creel. Multi-core CPUs, clusters, and gridcomputing: A tutorial. Computational economics, 32(4):353–382, 2008.

T. Lubik. Investment spending, equilibrium indeterminacy, and theinteractions of monetary and fiscal policy. Technical ReportEconomics Working Paper Archive 490, The Johns HopkinsUniversity, 2003.

Parallel DYNARE

Conclusions

Bibliography II

Marco Ratto, Werner Roeger, and Jan in ’t Veld. QUEST III: Anestimated open-economy DSGE model of the euro area withfiscal and monetary policy. Economic Modelling, 26(1):222 –233, 2009. doi: DOI:10.1016/j.econmod.2008.06.014. URLhttp://www.sciencedirect.com/science/article/B6VB1-4TC8J5F-1/2/7f22da17478841ac5d7a77d06f13d13e.

M. Russinovich. PsTools v2.44, 2009. available at MicrosoftTechNet, http://technet.microsoft.com/en-us/sysinternals/bb896649.aspx.

http://www.sciencedirect.com/science/article/B6VB1-4TC8J5F-1/2/7f22da17478841ac5d7a77d06f13d13e

http://www.sciencedirect.com/science/article/B6VB1-4TC8J5F-1/2/7f22da17478841ac5d7a77d06f13d13e