Merge pull request #11959 from hmaarrfk/pad_fill_before_benchmark

MAINT: Explicitely cause pagefaults to happen before starting the benchmarks

5 files changed, 25 insertions, 6 deletions

diff --git a/benchmarks/README.rst b/benchmarks/README.rst
index f4f0b0de9..b67994ce0 100644
--- a/benchmarks/README.rst
+++ b/benchmarks/README.rst
@@ -60,3 +60,11 @@ Some things to consider:
 - Preparing arrays etc. should generally be put in the ``setup`` method rather
   than the ``time_`` methods, to avoid counting preparation time together with
   the time of the benchmarked operation.
+
+- Be mindful that large arrays created with ``np.empty`` or ``np.zeros`` might
+  not be allocated in physical memory until the memory is accessed. If this is
+  desired behaviour, make sure to comment it in your setup function. If
+  you are benchmarking an algorithm, it is unlikely that a user will be
+  executing said algorithm on a newly created empty/zero array. One can force
+  pagefaults to occur in the setup phase either by calling ``np.ones`` or
+  ``arr.fill(value)`` after creating the array,
diff --git a/benchmarks/benchmarks/bench_lib.py b/benchmarks/benchmarks/bench_lib.py
index 83f26c9d1..e6c91a27c 100644
--- a/benchmarks/benchmarks/bench_lib.py
+++ b/benchmarks/benchmarks/bench_lib.py
@@ -19,7 +19,10 @@ class Pad(Benchmark):
     ]
 
     def setup(self, shape, pad_width, mode):
-        self.array = np.empty(shape)
+        # avoid np.zeros or np.empty's lazy allocation.
+        # np.full causes pagefaults to occur during setup
+        # instead of during the benchmark
+        self.array = np.full(shape, 0)
 
     def time_pad(self, shape, pad_width, mode):
         np.pad(self.array, pad_width, mode)
diff --git a/benchmarks/benchmarks/bench_ma.py b/benchmarks/benchmarks/bench_ma.py
index d313f01dc..aff78df0a 100644
--- a/benchmarks/benchmarks/bench_ma.py
+++ b/benchmarks/benchmarks/bench_ma.py
@@ -89,7 +89,9 @@ class Concatenate(Benchmark):
     ]
 
     def setup(self, mode, n):
-        normal = np.zeros((n, n), int)
+        # avoid np.zeros's lazy allocation that cause page faults during benchmark.
+        # np.fill will cause pagefaults to happen during setup.
+        normal = np.full((n, n), 0, int)
         unmasked = np.ma.zeros((n, n), int)
         masked = np.ma.array(normal, mask=True)
 
diff --git a/benchmarks/benchmarks/bench_reduce.py b/benchmarks/benchmarks/bench_reduce.py
index 353eb980c..ffc148cd2 100644
--- a/benchmarks/benchmarks/bench_reduce.py
+++ b/benchmarks/benchmarks/bench_reduce.py
@@ -29,8 +29,10 @@ class AddReduceSeparate(Benchmark):
 
 class AnyAll(Benchmark):
     def setup(self):
-        self.zeros = np.zeros(100000, bool)
-        self.ones = np.ones(100000, bool)
+        # avoid np.zeros's lazy allocation that would
+        # cause page faults during benchmark
+        self.zeros = np.full(100000, 0, bool)
+        self.ones = np.full(100000, 0, bool)
 
     def time_all_fast(self):
         self.zeros.all()
diff --git a/benchmarks/benchmarks/bench_shape_base.py b/benchmarks/benchmarks/bench_shape_base.py
index b05ea8263..6edad2ea3 100644
--- a/benchmarks/benchmarks/bench_shape_base.py
+++ b/benchmarks/benchmarks/bench_shape_base.py
@@ -23,7 +23,9 @@ class Block(Benchmark):
         self.four_1d = np.ones(6 * n)
         self.five_0d = np.ones(1 * n)
         self.six_1d = np.ones(5 * n)
-        self.zero_2d = np.zeros((2 * n, 6 * n))
+        # avoid np.zeros's lazy allocation that might cause
+        # page faults during benchmark
+        self.zero_2d = np.full((2 * n, 6 * n), 0)
 
         self.one = np.ones(3 * n)
         self.two = 2 * np.ones((3, 3 * n))
@@ -31,7 +33,9 @@ class Block(Benchmark):
         self.four = 4 * np.ones(3 * n)
         self.five = 5 * np.ones(1 * n)
         self.six = 6 * np.ones(5 * n)
-        self.zero = np.zeros((2 * n, 6 * n))
+        # avoid np.zeros's lazy allocation that might cause
+        # page faults during benchmark
+        self.zero = np.full((2 * n, 6 * n), 0)
 
     def time_block_simple_row_wise(self, n):
         np.block([self.a_2d, self.b_2d])