test: change the way to calculate the normalized euclidean distance to see if...

test: change the way to calculate the normalized euclidean distance to see if two outputs are close enough

test/distance_batsim_output.py

@@ -49,8 +49,20 @@ def open_and_compare(file1, file2):
     out1 = out1[out1.success == 1]
     out2 = out2[out2.success == 1]
 
+    runtimes1 = out1.finish_time - out1.starting_time
+    runtimes2 = out2.finish_time - out2.starting_time
+    
+    if not np.allclose(runtimes1, runtimes2, atol=1):
+        warnings.warn(f"Some jobs in {file1} and {file2} don't have the same runtime (+/- 1sec). It is unusual, as runtime are normally an input of the simulation.")
+
     return out1, out2
 
+def sum_execution_times(out):
+    """Returns the sum of all job execution times from a clean batsim output"""
+
+    runtimes = out.finish_time - out.starting_time
+    return runtimes.sum()
+
 
 def euclidean_distance(s1, s2):
     """Returns the Euclidean distance between two series s1 and s2"""
@@ -64,21 +76,6 @@ def lateness_distance(s1, s2):
 
     return float(np.sum([y-x for x, y in zip(s1, s2)]))
 
-def normalized_euclidian_distance(s1, s2):
-    """Return the euclidien distance normalized by the l2 norm of the vectors, 
-    or None if one of the vectors is the null vector (undefined)"""
-
-    n1, n2 = l2_norm(s1), l2_norm(s2)
-    if n1==0 or n2==0:
-        return None
-    eucl_dist = euclidean_distance(s1, s2)
-    return float( eucl_dist**2 / (n1 * n2) )
-
-def l2_norm(s):
-    """Return the l2 norm of the series s"""
-
-    return float( np.sqrt(np.sum([x * x for x in s])) )
-
 
 def distances(file1, file2, euclidean=True, lateness=False, norm_eucl=False,
               field=["finish_time"]):
@@ -87,13 +84,20 @@ def distances(file1, file2, euclidean=True, lateness=False, norm_eucl=False,
 
     out1, out2 = open_and_compare(file1, file2)
 
+    sum_runtimes1 = sum_execution_times(out1)
+    sum_runtimes2 = sum_execution_times(out2)
+
     dist = {}
     for f in field:
         dist[f] = {}
+        eucl = euclidean_distance(out1[f], out2[f])
         if euclidean:
-            dist[f]["euclidean"] = euclidean_distance(out1[f], out2[f])
+            dist[f]["euclidean"] = eucl
         if norm_eucl:
-            dist[f]["normalized_euclidean"] = normalized_euclidian_distance(out1[f], out2[f])
+            if sum_runtimes1 == 0 or sum_runtimes2 == 0:
+                dist[f]["normalized_euclidean"] = None
+            else:
+                dist[f]["normalized_euclidean"] = eucl / np.sqrt(sum_runtimes1 * sum_runtimes2)
         if lateness:
             dist[f]["lateness"] = lateness_distance(out1[f], out2[f])
 

test/test_fb_users.py

@@ -99,7 +99,8 @@ def simu_output_are_close_enough(out_dir1, out_dir2):
     """Returns true if the jobs.csv in both directories are epsilon-close (for 
     the normalized euclidean distance) for each of the fields submission time, 
     start time and finish time."""
-    epsilon = 1e-6 # precision of 0.0001%
+    epsilon = 1e-4 # threshold under which we consider that the distance is 
+                   # neglectable compared to the summed execution times
 
     dis = distances(f"{out_dir1}/_jobs.csv", f"{out_dir2}/_jobs.csv", 
                     euclidean=False, norm_eucl=True,