From: Brian Warner <warner@allmydata.com>
Date: Fri, 9 Jan 2009 02:41:16 +0000 (-0700)
Subject: dirnode deep-traversal: remove use of Limiter, stick with strict depth-first-traversa... 
X-Git-Url: https://git.rkrishnan.org/components/architecture.txt?a=commitdiff_plain;h=39a089dc7e2d49e1186ff2761c9e010769730931;p=tahoe-lafs%2Ftahoe-lafs.git

dirnode deep-traversal: remove use of Limiter, stick with strict depth-first-traversal, to reduce memory usage during very large (300k+ dirnode) traversals
---

diff --git a/src/allmydata/dirnode.py b/src/allmydata/dirnode.py
index 9c38f86d..375af2a5 100644
--- a/src/allmydata/dirnode.py
+++ b/src/allmydata/dirnode.py
@@ -15,7 +15,6 @@ from allmydata.monitor import Monitor
 from allmydata.util import hashutil, mathutil, base32, log
 from allmydata.util.assertutil import _assert, precondition
 from allmydata.util.hashutil import netstring
-from allmydata.util.limiter import ConcurrencyLimiter
 from allmydata.util.netstring import split_netstring
 from allmydata.uri import NewDirectoryURI, LiteralFileURI, from_string
 from pycryptopp.cipher.aes import AES
@@ -472,42 +471,55 @@ class NewDirectoryNode:
         directories) I can reach. Most work should be done here.
 
         I avoid loops by keeping track of verifier-caps and refusing to call
-        each() or traverse a node that I've seen before.
+        walker.add_node() or traverse a node that I've seen before. This
+        means that any file or directory will only be given to the walker
+        once. If files or directories are referenced multiple times by a
+        directory structure, this may appear to under-count or miss some of
+        them.
 
         I return a Deferred that will fire with the value of walker.finish().
         """
 
         # this is just a tree-walker, except that following each edge
-        # requires a Deferred. We use a ConcurrencyLimiter to make sure the
-        # fan-out doesn't cause problems.
+        # requires a Deferred. We used to use a ConcurrencyLimiter to limit
+        # fanout to 10 simultaneous operations, but the memory load of the
+        # queued operations was excessive (in one case, with 330k dirnodes,
+        # it caused the process to run into the 3.0GB-ish per-process 32bit
+        # linux memory limit, and crashed). So we use a single big Deferred
+        # chain, and do a strict depth-first traversal, one node at a time.
+        # This can be slower, because we aren't pipelining directory reads,
+        # but it brought the memory footprint down by roughly 50%.
 
         monitor = Monitor()
         walker.set_monitor(monitor)
 
         found = set([self.get_verify_cap()])
-        limiter = ConcurrencyLimiter(10)
-        d = self._deep_traverse_dirnode(self, [],
-                                        walker, monitor, found, limiter)
+        d = self._deep_traverse_dirnode(self, [], walker, monitor, found)
         d.addCallback(lambda ignored: walker.finish())
         d.addBoth(monitor.finish)
         d.addErrback(lambda f: None)
 
         return monitor
 
-    def _deep_traverse_dirnode(self, node, path,
-                               walker, monitor, found, limiter):
+    def _deep_traverse_dirnode(self, node, path, walker, monitor, found):
         # process this directory, then walk its children
         monitor.raise_if_cancelled()
-        d = limiter.add(walker.add_node, node, path)
-        d.addCallback(lambda ignored: limiter.add(node.list))
+        d = defer.maybeDeferred(walker.add_node, node, path)
+        d.addCallback(lambda ignored: node.list())
         d.addCallback(self._deep_traverse_dirnode_children, node, path,
-                      walker, monitor, found, limiter)
+                      walker, monitor, found)
         return d
 
     def _deep_traverse_dirnode_children(self, children, parent, path,
-                                        walker, monitor, found, limiter):
+                                        walker, monitor, found):
         monitor.raise_if_cancelled()
-        dl = [limiter.add(walker.enter_directory, parent, children)]
+        d = defer.maybeDeferred(walker.enter_directory, parent, children)
+        # we process file-like children first, so we can drop their FileNode
+        # objects as quickly as possible. Tests suggest that a FileNode (held
+        # in the client's nodecache) consumes about 2440 bytes. dirnodes (not
+        # in the nodecache) seem to consume about 2000 bytes.
+        dirkids = []
+        filekids = []
         for name, (child, metadata) in children.iteritems():
             verifier = child.get_verify_cap()
             # allow LIT files (for which verifier==None) to be processed
@@ -516,12 +528,18 @@ class NewDirectoryNode:
             found.add(verifier)
             childpath = path + [name]
             if IDirectoryNode.providedBy(child):
-                dl.append(self._deep_traverse_dirnode(child, childpath,
-                                                      walker, monitor,
-                                                      found, limiter))
+                dirkids.append( (child, childpath) )
             else:
-                dl.append(limiter.add(walker.add_node, child, childpath))
-        return defer.DeferredList(dl, fireOnOneErrback=True, consumeErrors=True)
+                filekids.append( (child, childpath) )
+        for (child, childpath) in filekids:
+            d.addCallback(lambda ignored, child=child, childpath=childpath:
+                          walker.add_node(child, childpath))
+        for (child, childpath) in dirkids:
+            d.addCallback(lambda ignored, child=child, childpath=childpath:
+                          self._deep_traverse_dirnode(child, childpath,
+                                                      walker, monitor,
+                                                      found))
+        return d
 
 
     def build_manifest(self):