INLA crashing when num.threads

pierre...@hotmail.fr

unread,

Nov 14, 2016, 11:49:15 AM11/14/16

to R-inla discussion group

Hello,

I try to run the model below with 5 covariates (with one interaction) with one random effect and with one a SPDE model.

formula = y ~ -1 + intercept + a + b + x + c + a:b + f(u, model="iid) + f(spatial.field, model=spde)

model<-inla(formula,data=inla.stack.data(stack, spde=spde), family="gaussian", control.predictor=list(A=inla.stack.A(stack)), control.compute=list(dic=TRUE,cpo=TRUE))

The model consists of 76,147 observations and the mesh has 373 vertices.

This model is running when I add num.threads = 1, but it kept crashing when num.threads > 1. I added control.inla= list(strategy = "gaussian", int.strategy = "eb") to fix the problem but it kept crashing when num.threads was superior to 1.

Does the model will run faster if num.threads > 1 ? If so, is there a way to run the model with more threads ?

I work on Mac OS Sierra 10.12.1, 4 Core, with 16Gb RAM, and below I add the error message and the last lines of the message when verbose=TRUE.

Thanks,

Pierre Barry

The error message is :

Error in inla.inlaprogram.has.crashed() :

The inla-program exited with an error. Unless you interupted it yourself, please rerun with verbose=TRUE and check the output carefully.

If this does not help, please contact the developers at <he...@r-inla.org>.

And here are the last lines the message when I add verbose = TRUE

Index table: number of entries[9], total length[100340]

tag start-index length

Predictor 0 99919

u 99919 42

spatial.field 99961 373

intercept 100334 1

a 100335 1

b 100336 1

c 100337 1

x 100338 1

a:b 100339 1

parse section=[11] name=[INLA.Parameters] type=[INLA]

inla_parse_INLA...

section[INLA.Parameters]

lincomb.derived.only = [Yes]

lincomb.derived.correlation.matrix = [No]

global_node.factor = 2.000

global_node.degree = 2147483647

reordering = -1

Contents of ai_param 0x7fcfcbd6bc70

Optimiser: DEFAULT METHOD

Option for domin-BFGS: epsx = 0.005

Option for domin-BFGS: epsf = 1e-05 (rounding error)

Option for domin-BFGS: epsg = 0.005

Option for GSL-BFGS2: tol = 0.1

Option for GSL-BFGS2: step_size = 1

Option for GSL-BFGS2: epsx = 0.005

Option for GSL-BFGS2: epsf = 0.000353553

Option for GSL-BFGS2: epsg = 0.005

Restart: 0

Mode known: No

Gaussian approximation:

abserr_func = 0.0005

abserr_step = 0.0005

optpar_fp = 0

optpar_nr_step_factor = -0.1

Gaussian data: Yes

Strategy: Use a mean-skew corrected Gaussian by fitting a Skew-Normal

Fast mode: On

Use linear approximation to log(|Q +c|)? Yes

Method: Compute the derivative exact

Parameters for improved approximations

Number of points evaluate: 9

Step length to compute derivatives numerically: 0.000100002

Stencil to compute derivatives numerically: 5

Cutoff value to construct local neigborhood: 0.0001

Log calculations: On

Log calculated marginal for the hyperparameters: On

Integration strategy: Automatic (GRID for dim(theta)=1 and 2 and otherwise CCD)

f0 (CCD only): 1.100000

dz (GRID only): 0.750000

Adjust weights (GRID only): On

Difference in log-density limit (GRID only): 4.000000

Skip configurations with (presumed) small density (GRID only): On

Gradient is computed using Central difference with step-length 0.010000

Hessian is computed using Central difference with step-length 0.100000

Hessian matrix is forced to be a diagonal matrix? [No]

Compute effective number of parameters? [Yes]

Perform a Monte Carlo error-test? [No]

Interpolator [Auto]

CPO required diff in log-density [3]

Stupid search mode:

Status [On]

Max iter [1000]

Factor [1.05]

Numerical integration of hyperparameters:

Maximum number of function evaluations [100000]

Relative error ....................... [1e-05]

Absolute error ....................... [1e-06]

To stabilise the numerical optimisation:

Minimum value of the -Hessian [-inf]

CPO manual calculation[No]

Laplace-correction is Disabled.

inla_build: check for unused entries in[/var/folders/5w/jb_mh_dx7p95wnjmrfrx09_0 0000gp/T//Rtmpu4whHv/file10b661425d56/Model.ini]

inla_INLA...

Strategy = [DEFAULT]

Size is [100340]

Chose OpenMP-strategy [HUGE]

Chose density-strategy [LOW]

Size of graph=[100340] constraints=[0]

Found optimal reordering=[amdbarc] nnz(L)=[39790117] and use_global_nodes(user)=[no]

List of hyperparameters:

theta[0] = [Log precision for the Gaussian observations]

theta[1] = [Log precision for u]

theta[2] = [Theta1 for spatial.field]

theta[3] = [Theta2 for spatial.field]

Optimise using DEFAULT METHOD

INLA help

unread,

Nov 14, 2016, 2:51:55 PM11/14/16

to pierre...@hotmail.fr, R-inla discussion group

On Mon, 2016-11-14 at 08:49 -0800, pierre...@hotmail.fr wrote:
> Hello,
>
> I try to run the model below with 5 covariates (with one interaction)
> with one random effect and with one a SPDE model.
>
> formula = y ~ -1 + intercept + a + b + x + c + a:b + f(u, model="iid)
> + f(spatial.field, model=spde)
> model<-inla(formula,data=inla.stack.data(stack, spde=spde),
> family="gaussian", control.predictor=list(A=inla.stack.A(stack)),
> control.compute=list(dic=TRUE,cpo=TRUE))
>
> The model consists of 76,147 observations and the mesh has 373
> vertices.
>
> This model is running when I add num.threads = 1, but it kept
> crashing when num.threads > 1. I added control.inla= list(strategy =
> "gaussian", int.strategy = "eb") to fix the problem but it kept
> crashing when num.threads was superior to 1.
>
> Does the model will run faster if num.threads > 1 ? If so, is there a
> way to run the model with more threads ?
>
> I work on Mac OS Sierra 10.12.1, 4 Core, with 16Gb RAM, and below I
> add the error message and the last lines of the message when
> verbose=TRUE.
>

Hi,

yes, it will run faster with more threads (up to some limit).

could you try adding

control.inla = list(reordering = "metis")

?

--
Håvard Rue
he...@r-inla.org

pierre...@hotmail.fr

unread,

Nov 15, 2016, 9:10:52 AM11/15/16

to R-inla discussion group, pierre...@hotmail.fr, he...@r-inla.org

Le lundi 14 novembre 2016 19:51:55 UTC, help help a écrit :

Hi,

yes, it will run faster with more threads (up to some limit).

could you try adding

control.inla = list(reordering = "metis")

?

--
Håvard Rue
he...@r-inla.org

Hello,

Thanks for the response.

No it kept crashing either when num.threads > 1 or when I remove num.threads.

Below the last lines when verbose = TRUE. The reordering parameter changed from -1 to 3.

Thanks,

Pierre Barry

output:

summary=[1]

return.marginals=[1]

nquantiles=[3] [ 0.025 0.5 0.975 ]

ncdf=[0] [ ]

Index table: number of entries[9], total length[100340]

tag start-index length

Predictor 0 99919

u 99919 42

spatial.field 99961 373

intercept 100334 1

a 100335 1

b 100336 1

c 100337 1

x 100338 1

a:b 100339 1

parse section=[11] name=[INLA.Parameters] type=[INLA]

inla_parse_INLA...

section[INLA.Parameters]

lincomb.derived.only = [Yes]

lincomb.derived.correlation.matrix = [No]

global_node.factor = 2.000

global_node.degree = 2147483647

reordering = 3

Contents of ai_param 0x7fddaae15db0

inla_build: check for unused entries in[/var/folders/5w/jb_mh_dx7p95wnjmrfrx09_0 0000gp/T//RtmpvHD4W1/file183a4227b405/Model.ini]

inla_INLA...

Strategy = [DEFAULT]

Size is [100340]

Chose OpenMP-strategy [HUGE]

Chose density-strategy [LOW]

Size of graph=[100340] constraints=[0]

INLA help

unread,

Nov 15, 2016, 9:52:06 AM11/15/16

to pierre...@hotmail.fr, R-inla discussion group

if you can send me stuff so I can rerun it here, I can check

H

> --
> You received this message because you are subscribed to the Google
> Groups "R-inla discussion group" group.
> To unsubscribe from this group and stop receiving emails from it,
> send an email to r-inla-discussion...@googlegroups.com
> .
> To post to this group, send email to r-inla-discussion-group@googlegr
> oups.com.
> Visit this group at https://groups.google.com/group/r-inla-discussion
> -group.
> For more options, visit https://groups.google.com/d/optout.

--
Håvard Rue
he...@r-inla.org

Reply all

Reply to author

Forward

INLA crashing when num.threads > 1

pierre...@hotmail.fr

INLA help

pierre...@hotmail.fr

INLA help