diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 8b999421c87273e3a7ca4c68a20b792acabcc101..358c7fe8cae37792215d2fe541c08645620faadc 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -24,7 +24,6 @@ Python 3.5 K40:
     - nvidia-k40
     except:
     - tags
-    - distributed-fmm-global
   
 Python 2.7 POCL:
     script:
@@ -38,7 +37,6 @@ Python 2.7 POCL:
     - pocl
     except:
     - tags
-    - distributed-fmm-global
 
 Python 3.5 POCL:
     script:
@@ -52,7 +50,6 @@ Python 3.5 POCL:
     - pocl
     except:
     - tags
-    - distributed-fmm-global
 
 Python 3.6 POCL:
     script:
@@ -66,7 +63,20 @@ Python 3.6 POCL:
     - pocl
     except:
     - tags
-    - distributed-fmm-global
+
+Python 3.6 POCL MPI:
+    script:
+    - export PY_EXE=python3.6
+    - export PYOPENCL_TEST=portable
+    - export EXTRA_INSTALL="numpy mako mpi4py"
+    - export PYTEST_ADDOPTS="-k mpi --capture=no"
+    - curl -L -O -k https://gitlab.tiker.net/inducer/ci-support/raw/master/build-and-test-py-project.sh
+    - ". ./build-and-test-py-project.sh"
+    tags:
+    - python3.6
+    - pocl
+    except:
+    - tags
 
 Documentation:
     script:
diff --git a/setup.py b/setup.py
index 32cca6a23d46884c398fe0f31c363f7e8f4e7c76..9b424e191ead0357b59a0382926e832c8e073a9c 100644
--- a/setup.py
+++ b/setup.py
@@ -3,7 +3,7 @@
 
 
 def main():
-    from setuptools import setup
+    from setuptools import setup, find_packages
 
     version_dict = {}
     init_filename = "boxtree/version.py"
@@ -39,7 +39,7 @@ def main():
               'Topic :: Utilities',
               ],
 
-          packages=["boxtree"],
+          packages=find_packages(),
           install_requires=[
               "pytools>=2018.4",
               "pyopencl>=2013.1",
@@ -47,7 +47,12 @@ def main():
               "pytest>=2.3",
               "cgen>=2013.1.2",
               "six",
-              ])
+              ],
+          package_data={
+              "boxtree": [
+                  "distributed/default_perf_model.json"
+              ]
+          })
 
 
 if __name__ == '__main__':
diff --git a/test/test_distributed.py b/test/test_distributed.py
index 42ca946d31d838804a4cceb995ba561d8b710247..c4a22107db94ccbf2fbb1796f5f05bfc68c604ff 100644
--- a/test/test_distributed.py
+++ b/test/test_distributed.py
@@ -1,8 +1,5 @@
 import numpy as np
 import pyopencl as cl
-from mpi4py import MPI
-from boxtree.distributed.calculation import DistributedFMMLibExpansionWrangler
-from boxtree.distributed import DistributedFMMInfo
 import numpy.linalg as la
 from boxtree.pyfmmlib_integration import FMMLibExpansionWrangler
 from boxtree.tools import ConstantOneExpansionWrangler as \
@@ -17,6 +14,7 @@ logging.getLogger("boxtree.distributed").setLevel(logging.INFO)
 
 
 def _test_against_shared(dims, nsources, ntargets, dtype):
+    from mpi4py import MPI
 
     # Get the current rank
     comm = MPI.COMM_WORLD
@@ -73,9 +71,13 @@ def _test_against_shared(dims, nsources, ntargets, dtype):
     # Compute FMM using distributed memory parallelism
 
     def distributed_expansion_wrangler_factory(tree):
+        from boxtree.distributed.calculation import \
+                DistributedFMMLibExpansionWrangler
+
         return DistributedFMMLibExpansionWrangler(
             queue, tree, helmholtz_k, fmm_level_to_nterms=fmm_level_to_nterms)
 
+    from boxtree.distributed import DistributedFMMInfo
     distribued_fmm_info = DistributedFMMInfo(
         queue, trav, distributed_expansion_wrangler_factory, comm=comm)
     pot_dfmm = distribued_fmm_info.drive_dfmm(sources_weights)
@@ -99,13 +101,15 @@ def test_against_shared(num_processes, dims, nsources, ntargets):
     newenv["dims"] = str(dims)
     newenv["nsources"] = str(nsources)
     newenv["ntargets"] = str(ntargets)
+    newenv["OMP_NUM_THREADS"] = "1"
 
     import subprocess
     import sys
     subprocess.run([
         "mpiexec", "-np", str(num_processes),
         "-x", "PYTEST", "-x", "dims", "-x", "nsources", "-x", "ntargets",
-        sys.executable, __file__],
+        # https://mpi4py.readthedocs.io/en/stable/mpi4py.run.html
+        sys.executable, "-m", "mpi4py.run", __file__],
         env=newenv,
         check=True
     )
@@ -123,6 +127,7 @@ class ConstantOneExpansionWrangler(ConstantOneExpansionWranglerBase):
 
 
 def _test_constantone(dims, nsources, ntargets, dtype):
+    from mpi4py import MPI
 
     # Get the current rank
     comm = MPI.COMM_WORLD
@@ -136,7 +141,6 @@ def _test_constantone(dims, nsources, ntargets, dtype):
     import pyopencl as cl
     ctx = cl.create_some_context()
     queue = cl.CommandQueue(ctx)
-    print(queue.context.devices)
 
     if rank == 0:
 
@@ -189,13 +193,15 @@ def test_constantone(num_processes, dims, nsources, ntargets):
     newenv["dims"] = str(dims)
     newenv["nsources"] = str(nsources)
     newenv["ntargets"] = str(ntargets)
+    newenv["OMP_NUM_THREADS"] = "1"
 
     import subprocess
     import sys
     subprocess.run([
         "mpiexec", "-np", str(num_processes),
         "-x", "PYTEST", "-x", "dims", "-x", "nsources", "-x", "ntargets",
-        sys.executable, __file__],
+        # https://mpi4py.readthedocs.io/en/stable/mpi4py.run.html
+        sys.executable, "-m", "mpi4py.run", __file__],
         env=newenv,
         check=True
     )