more edits

Author: James Stevens

gossip.tex

@@ -56,7 +56,7 @@ \subsection{Gossip Results}
 To analyze the effectiveness of this gossip protocol we created a
 swarm like the ones we used to test the effectivenes of local, global
 and omniscient knowledge. The gossip swarm results are presented in
-Table X. From these results we can clearly see that the gossip
+Table \ref{tab:noseedresults}. From these results we can clearly see that the gossip
 protocol replicates the critically rare pieces much more efficiently
 than local and global knowledge do. This allows it significantly extend the
 lifetimes of three of the five test swarms. The gossip protocol does

intro.tex

@@ -61,27 +61,21 @@ \section{Introduction}
 construct an improved global view of the swarm at each peer.  We then
 modify the next piece selection routine to leverage the information
 about the globally rarest blocks to request the blocks most in need of
-replication from its neighbors. We verified the effectiveness of this approach 
-by running a series of simulations that approximate a
-range of swarms. We also analyze the gossip protocol's overhead to show
-that gossip is practical in terms of network usage and node storage.
-
-Additionally, we also propose an alternative gossip protocol to assist
+replication from its neighbors. We verified the effectiveness of this
+approach by running a series of simulations that approximate a unhealthy
+swarms. We also analyze the gossip protocol's overhead to show that gossip
+is practical in terms of network usage and node storage.  Additionally,
+we also propose an alternative low-overhead gossip protocol to assist
 seeds that are altruistic enough to want the swarm to survive after they
-leave. This protocol uses gossip to let the seed know when it is safe
-to leave the swarm by ensuring that enough replicas of each pieces are
-available. We also show that the alternative gossip protocol uses even
-less network resources than the previous gossip protocol.
-
+leave by letting these seeds know when it is safe to leave.
 
 The primary insight of this work is that the lack of a global view is
 a contributing factor to the last block problem in unhealthy swarms
 and can be improved with a global picture of the swarm.  Furthermore,
-the use of gossip to communicate information about the swarm as a whole
+the use of gossip to communicate global information about the swarm 
 allowed us to do efficiently what had previously been considered too
 costly to implement \cite{cohen:1}.
 
-
 The remainder of this paper is organized as follows. Section II discusses
 the related work. Section III describes our simulation design.  Section IV
 is a study of local knowledge vs. global knowledge in BitTorrent.  Section

swarm.tex

@@ -15,13 +15,13 @@ \subsection{Experimental Setup}
 node's peers. The goal of this section is to explore the potential for
 a gossip algorithm to improve BitTorrent. The first set of simulations
 use swarms that start with a single seed and a number of leechers.
-The seed is altruistic and therefore will stay the entire time. This 
+The seed is altruistic and will stay the entire time. This 
 simulates a ``healthy'' swarm that will not die, but may benefit from 
 better replication of pieces. The second type is a swarm with no seeders, 
 although a complete copy of the file exists within the swarm between a
 few nodes. This simulates a situation in which the seeders have left
 and a swarm is in an ``unhealthy'' state. None of the nodes are
-altruistic and therefore leave as soon as possible. For this swarm,
+altruistic and leave as soon as possible. For this swarm,
 there are two goals. The primary goal is to simply make sure the swarm
 survives, because it is possible for a node to leave with pieces that
 only it holds, which means no other node can complete. The second goal
@@ -52,9 +52,9 @@ \subsection{Swarms with Seeders}
 number of peers per node, and the initial seeder's upload rate will be varied.
 In the healthiest swarms with a seeder, there should be many nodes, many peers per node, and
 a high seeder upload rate. Likewise, the least healthy swarms have fewer nodes,
-fewer peers per node, and a lower upload rate for the seeder. However, if too fewer
+fewer peers per node, and a lower upload rate for the seeder. However, if too few
 nodes exist than the peers allowed per node, then each node will have perfect global
-knowledge by default, therefore, the number of nodes must always be kept above
+knowledge by default, so the number of nodes must always be kept above
 the number of peers per node. 
 
 
@@ -94,7 +94,7 @@ \subsection{Swarms with Seeders}
 replicated. We also ran some even slower seeders such as an as uprate
 of 20 that were not finishing in any reasonable amount of time so we
 did not collect the run times. One interesting result was that even if
-the numer of nodes doubles, the performance penalty for node downloading
+the number of nodes doubles, the performance penalty for node downloading
 was essentially negligible. This shows that BitTorrent scales very well
 assuming the initial seeder can get the pieces replicated.  We can also
 see that as the number of peers increases, performance improves somewhat
@@ -119,12 +119,9 @@ \subsection{Swarms with No Seeders}
 large number of nodes join, complete the file, and then leave. The
 four nodes that start this simulation would then be the remainder of
 that initial flash crowd while the new nodes joining would represent
-the steady stream of nodes that often join after a flash crowd. For
-more information on this behavior in torrents, please see 
-``The Bittorrent  P2P File-Sharing System: Measurements and Analysis''
-(**CITE HERE**). 
+the steady stream of nodes that often join after a flash crowd. 
 
-Using this ``unheatly'' swarm simulation setup we then ran the
+Using this ``unheathly'' swarm simulation setup we then ran the
 simulation five times for the three levels of node knowledge discussed
 in section 4.1: local knowledge, global knowledge and
 omniscience. Five runs were performed for each level of knowledge
@@ -146,14 +143,16 @@ \subsection{Swarms with No Seeders}
 \multicolumn{11}{|c|}{Experiments}\\ \hline
 Runs & 1 & 2 & 3 & 4 & 5 & 1 & 2 & 3 & 4 & 5\\ \hline
 Knowledge Type & \multicolumn{5}{c|}{nodes stranded} & \multicolumn{5}{c|}{average times}\\ \hline
-local      & 99&  2& 99& 99& 99&1387&2540&1710&2019&1433\\
-global     & 99& 99& 99& 99& 46&1850&1464&1369&1535&2391\\
-omniscient &  1&  2&  2& 99&  2&1711&2370&1646&1465&2386\\ \hline \hline
+Local      & 99&  2& 99& 99& 99&1387&2540&1710&2019&1433\\
+Global     & 99& 99& 99& 99& 46&1850&1464&1369&1535&2391\\
+Omniscient &  1&  2&  2& 99&  2&1711&2370&1646&1465&2386\\ 
+Gossip      & 11& 99&  1&  7& 99&1981&1254&2293&1901&2130\\ \hline \hline
 \multicolumn{11}{|c|}{Summary}\\ \hline
 Knowledge Type & \multicolumn{5}{c|}{Average Nodes Stranded} & \multicolumn{5}{c|}{Average Times}\\ \hline
 Local & \multicolumn{5}{c|}{80} & \multicolumn{5}{c|}{1818}\\
 Global & \multicolumn{5}{c|}{88} & \multicolumn{5}{c|}{1727}\\
 Omniscient & \multicolumn{5}{c|}{21} & \multicolumn{5}{c|}{1916}\\
+Gossip      & \multicolumn{5}{c|}{43} & \multicolumn{5}{c|}{1912}\\
 \hline\end{tabular}
 \label{tab:noseedresults}
 \end{table*}
@@ -179,11 +178,11 @@ \subsection{Swarms with No Seeders}
 global view of the swarms causes the nodes with global knowledge to all
 download the same piece from the critical nodes so many copies are
 created of one critically rare piece while other critically rare
-pieces aren't copied at all. The local knowledge nodes were therefore
+pieces aren't copied at all. The local knowledge nodes were 
 more efficient because their notion of the rarest piece was more
-diverse. This lead them to replicate different pieces and achieve more
+diverse. This led them to replicate different pieces and achieve more
 efficeint replication overall. The primary conclusion to be drawn from
-these results is, therefore, that while local knowledge can
+these results is while local knowledge can
 occaisionally save a swarm, a more intelligent piece replication scheme
 would be more effective. This opens the door for BitTorrent gossip to
 achieve real gains over the traditional implementation by extending