Limit the number of threads to 1 per process in order to prevent MKL from...

Limit the number of threads to 1 per process in order to prevent MKL from creating numerous threads leading to a strong oversubscription of the CPU/RAM as pcigale already does its own parallelisation.

CHANGELOG.md

@@ -11,6 +11,7 @@
 - When the pcigale.ini file was missing, pcigale would crash and display a fairly cryptic backtrace. Now it explicitly states that the file could not be found. (Médéric Boquien)
 
 ### Optimised
+- By default the MKL library created many threads for each for the parallel processes. Not only was this not necessary as a high-level parallelisation already exists, but it generated a strong oversubscription on the CPU and on the RAM. The slowdown was over a factor of ~2 in some cases. Now we mandate KML to use only 1 thread fo each process. (Médéric Boquien & Yannick Roehlly)
 
 ## 0.10.0 (2016-09-15)
 ### Added

pcigale/analysis_modules/pdf_analysis/workers.py

@@ -12,6 +12,7 @@ from astropy.cosmology import WMAP7 as cosmology
 import numpy as np
 import scipy.stats as st
 
+from ..utils import nothread
 from .utils import (save_best_sed, save_pdf, save_chi2, compute_chi2,
                     weighted_param)
 from ...warehouse import SedWarehouse
@@ -44,6 +45,10 @@ def init_sed(params, filters, analysed, fluxes, variables, t_begin, n_computed):
     global gbl_params, gbl_previous_idx, gbl_filters, gbl_analysed_variables
     global gbl_warehouse
 
+    # Limit the number of threads to 1 if we use MKL in order to limit the
+    # oversubscription of the CPU/RAM.
+    nothread()
+
     gbl_model_fluxes = np.ctypeslib.as_array(fluxes[0])
     gbl_model_fluxes = gbl_model_fluxes.reshape(fluxes[1])
 

pcigale/analysis_modules/savefluxes/workers.py

@@ -11,6 +11,7 @@ import numpy as np
 
 from ...warehouse import SedWarehouse
 from ..utils import OUT_DIR
+from ..utils import nothread
 
 
 def init_fluxes(params, filters, save_sed, variables, fluxes, info, t_begin,
@@ -41,6 +42,10 @@ def init_fluxes(params, filters, save_sed, variables, fluxes, info, t_begin,
     global gbl_params, gbl_previous_idx, gbl_filters, gbl_save_sed
     global gbl_warehouse, gbl_variables
 
+    # Limit the number of threads to 1 if we use MKL in order to limit the
+    # oversubscription of the CPU/RAM.
+    nothread()
+
     gbl_model_fluxes = np.ctypeslib.as_array(fluxes[0])
     gbl_model_fluxes = gbl_model_fluxes.reshape(fluxes[1])
 

pcigale/analysis_modules/utils.py

@@ -69,3 +69,25 @@ def save_fluxes(model_fluxes, model_parameters, filters, names,
     out_table.write("{}/computed_fluxes.fits".format(directory))
     out_table.write("{}/computed_fluxes.txt".format(directory),
                     format='ascii.fixed_width', delimiter=None)
+
+
+def nothread():
+    """Some libraries such as Intel's MKL have automatic threading. This is good
+    when having only one process. However we already do our own parallelisation.
+    The additional threads created by the MKL increase in excess the pressure on
+    the CPU and on the RAM slowing everything down.
+
+    Parameters
+    ----------
+    None
+
+    Returns
+    -------
+    None
+
+    """
+    try:
+        import mkl
+        mkl.set_num_threads(1)
+    except ImportError:
+        pass