# 
# 
# patch "mtn_benchmark/random_content.py"
#  from [654582744bd8a67d13964429f5330cab51e306ae]
#    to [9ae8b38da49a0de5c5737a6d3259057a286309e2]
# 
============================================================
--- mtn_benchmark/random_content.py	654582744bd8a67d13964429f5330cab51e306ae
+++ mtn_benchmark/random_content.py	9ae8b38da49a0de5c5737a6d3259057a286309e2
@@ -94,3 +94,23 @@
             replacement = self.genbytes(r, insert_size)
             text = text[:pivot1] + replacement + text[pivot2:]
         return text
+
+
+# Directory topology generation:
+#   While there are directories to add:
+#     Pick a directory
+#       Sample a Bernoulli distribution (p)^depth, where depth=0 for the root
+#         dir, 1 for a child of the root, etc.
+#       If this comes up true, attach a new dir here
+# This has a reasonable-looking empirical fit to real trees.  The linux kernel
+# has p = 0.4, uclinux has p = 0.5, xfree86 (which is rather pathological) has
+# something like p = 0.8, but even this has a lower density of deep
+# directories than the actual tree (while getting the maximum deepness right).
+# ACE is ~0.6, my other test trees are two sparse.  These are all eyeball
+# fits, looking only at distribution of depths, I didn't try any other forms
+# of generative models, and there are some systematic irregularities visible
+# even in this one (mostly that there needs to be some sort of "bottleneck"
+# control, some trees like X have about the same number of depth 5 dirs as
+# depth 3 dirs, for others this drops off much quicker).
+#
+# But this should be good enough for now.