Download optimisation

- new data structure to handle download tracking, uses less CPU

src/network/bmproto.py

@@ -546,7 +546,6 @@ class BMProto(AdvancedDispatcher, ObjectTracker):
         for connection in network.connectionpool.BMConnectionPool().inboundConnections.values() + \
                 network.connectionpool.BMConnectionPool().outboundConnections.values():
             try:
-                with connection.objectsNewToMeLock:
                 del connection.objectsNewToMe[hashId]
             except KeyError:
                 pass

src/network/connectionpool.py

@@ -40,7 +40,6 @@ class BMConnectionPool(object):
             if not i.fullyEstablished:
                 continue
             try:
-                with i.objectsNewToMeLock:
                 del i.objectsNewToMe[hashid]
             except KeyError:
                 with i.objectsNewToThemLock:

src/network/downloadthread.py

@@ -3,10 +3,8 @@ import threading
 import time
 
 import addresses
-#from bmconfigparser import BMConfigParser
 from debug import logger
 from helper_threading import StoppableThread
-#from inventory import Inventory
 from network.connectionpool import BMConnectionPool
 import protocol
 from state import missingObjects
@@ -49,32 +47,15 @@ class DownloadThread(threading.Thread, StoppableThread):
             for i in connections:
                 now = time.time()
                 timedOut = now - DownloadThread.requestTimeout
-                # this may take a while, but it needs a consistency so I think it's better to lock a bigger chunk
-                with i.objectsNewToMeLock:
                 try:
-                        downloadPending = len(list((k for k, v in i.objectsNewToMe.iteritems() if k in missingObjects and missingObjects[k] > timedOut and not v)))
+                    request = i.objectsNewToMe.randomKeys(requestChunk)
                 except KeyError:
                     continue
-                    if downloadPending >= DownloadThread.minPending:
-                        continue
-                    # keys with True values in the dict
-                    try:
-                        request = list((k for k, v in i.objectsNewToMe.iteritems() if k not in missingObjects or missingObjects[k] < timedOut))
-                    except KeyError:
-                        continue
-                    random.shuffle(request)
-                    if len(request) > requestChunk - downloadPending:
-                        request = request[:max(1, requestChunk - downloadPending)]
-                    if not request:
-                        continue
-                    # mark them as pending
-                    for k in request:
-                        i.objectsNewToMe[k] = False
-                        missingObjects[k] = now
                 payload = bytearray()
                 payload.extend(addresses.encodeVarint(len(request)))
                 for chunk in request:
                     payload.extend(chunk)
+                    missingObjects[k] = now
                 i.append_write_buf(protocol.CreatePacket('getdata', payload))
                 logger.debug("%s:%i Requesting %i objects", i.destination.host, i.destination.port, len(request))
                 requested += len(request)

src/network/objectracker.py

@@ -5,6 +5,7 @@ from threading import RLock
 from debug import logger
 from inventory import Inventory
 from network.dandelion import Dandelion
+from randomtrakcingdict import RandomTrackingDict
 from state import missingObjects
 
 haveBloom = False
@@ -32,8 +33,7 @@ class ObjectTracker(object):
     initialTimeOffset = 60
 
     def __init__(self):
-        self.objectsNewToMe = {}
-        self.objectsNewToMeLock = RLock()
+        self.objectsNewToMe = RandomTrackingDict()
         self.objectsNewToThem = {}
         self.objectsNewToThemLock = RLock()
         self.initInvBloom()
@@ -61,9 +61,6 @@ class ObjectTracker(object):
                 self.initAddrBloom()
             else:
                 # release memory
-                with self.objectsNewToMeLock:
-                    tmp = self.objectsNewToMe.copy()
-                    self.objectsNewToMe = tmp
                 deadline = time.time() - ObjectTracker.trackingExpires
                 with self.objectsNewToThemLock:
                     self.objectsNewToThem = {k: v for k, v in self.objectsNewToThem.iteritems() if v >= deadline}
@@ -88,8 +85,7 @@ class ObjectTracker(object):
             if hashId in Dandelion().hashMap:
                 Dandelion().fluffTrigger(hashId)
             if hashId not in missingObjects:
-                missingObjects[hashId] = time.time() - ObjectTracker.initialTimeOffset
-            with self.objectsNewToMeLock:
+                missingObjects[hashId] = True
             self.objectsNewToMe[hashId] = True
 
     def hasAddr(self, addr):

src/randomtrackingdict.py

@@ -0,0 +1,118 @@
+import random
+from threading import RLock
+from time import time
+
+class RandomTrackingDict(object):
+    maxPending = 10
+    pendingTimeout = 60
+    def __init__(self): # O(1)
+        self.dictionary = {}
+        self.indexDict = []
+        self.len = 0
+        self.pendingLen = 0
+        self.lastPoll = 0
+        self.lock = RLock()
+
+    def __len__(self):
+        return self.len
+
+    def __contains__(self, key):
+        return key in self.dictionary
+
+    def __getitem__(self, key):
+        return self.dictionary[key][1]
+
+    def __setitem__(self, key, value):
+        with self.lock:
+            if key in self.dictionary:
+                self.dictionary[key][1] = value
+            else:
+                self.indexDict.append(key)
+                self.dictionary[key] = [self.len, value]
+                self.len += 1
+
+    def __delitem__(self, key):
+        if not key in self.dictionary:
+            raise KeyError
+        with self.lock:
+            index = self.dictionary[key][0]
+            self.indexDict[index] = self.indexDict[self.len - 1]
+            self.dictionary[self.indexDict[index]][0] = index
+            # if the following del is batched, performance of this single
+            # operation can improve 4x, but it's already very fast so we'll
+            # ignore it for the time being
+            del self.indexDict[-1]
+            del self.dictionary[key]
+            self.len -= 1
+            if index >= self.len - self.pendingLen:
+                self.pendingLen -= 1
+
+    def setMaxPending(self, maxPending):
+        self.maxPending = maxPending
+
+    def setPendingTimeout(self, pendingTimeout):
+        self.pendingTimeout = pendingTimeout
+
+    def randomKeys(self, count=1):
+        if self.lastPoll + self.pendingTimeout < time():
+            with self.lock:
+                self.pendingLen = 0
+        if self.len == 0 or self.pendingLen >= self.maxPending:
+            raise KeyError
+        with self.lock:
+            available = self.len - self.pendingLen
+            if count > available:
+                count = available
+            retval = random.sample(self.indexDict[:self.len - self.pendingLen], count)
+            for i in retval[::-1]:
+                # swap with one below lowest pending
+                self.pendingLen += 1
+                swapKey = self.indexDict[-self.pendingLen]
+                curIndex = self.dictionary[i][0]
+                self.indexDict[-self.pendingLen] = i
+                self.indexDict[curIndex] = swapKey
+                self.dictionary[i][0] = self.len - self.pendingLen
+                self.dictionary[swapKey][0] = curIndex
+            self.lastPoll = time()
+            return retval
+
+if __name__ == '__main__':
+    def randString():
+        retval = b''
+        for _ in range(32):
+            retval += chr(random.randint(0,255))
+        return retval
+
+    a = []
+    k = RandomTrackingDict()
+    d = {}
+    
+#    print "populating normal dict"
+#    a.append(time())
+#    for i in range(50000):
+#        d[randString()] = True
+#    a.append(time())
+    print "populating random tracking dict"
+    a.append(time())
+    for i in range(50000):
+        k[randString()] = True
+    a.append(time())
+    print "done"
+    while len(k) > 0:
+        retval = k.randomKeys(1000)
+        if not retval:
+            print "error getting random keys"
+        #a.append(time())
+        try:
+            k.randomKeys(100)
+            print "bad"
+        except KeyError:
+            pass
+        #a.append(time())
+        for i in retval:
+            del k[i]
+        #a.append(time())
+    a.append(time())
+
+    for x in range(len(a) - 1):
+        print "%i: %.3f" % (x, a[x+1] - a[x])