[or-cvs] r15260: Completed presentation and discussion of message transfer ti (projects/hidserv/trunk/doc)

[kloesing at seul.org]

Author: kloesing
Date: 2008-06-14 17:03:51 -0400 (Sat, 14 Jun 2008)
New Revision: 15260

Added:
   projects/hidserv/trunk/doc/reqres.png
Modified:
   projects/hidserv/trunk/doc/report.tex
Log:
Completed presentation and discussion of message transfer times and connection durations.

Modified: projects/hidserv/trunk/doc/report.tex
===================================================================
--- projects/hidserv/trunk/doc/report.tex	2008-06-14 20:28:58 UTC (rev 15259)
+++ projects/hidserv/trunk/doc/report.tex	2008-06-14 21:03:51 UTC (rev 15260)
@@ -5,8 +5,9 @@
 \usepackage{url}
 \begin{document}
 \title{Tor Hidden Service Performance Analysis}
-\author{The Tor Project\footnote{Please direct questions and comments to
-tor-assistants at torproject.org.}}
+\author{The Tor Project\footnote{Please direct questions and comments
+either to \texttt{tor-assistants at torproject.org} or
+\texttt{or-dev at freehaven.net}.}}
 \date{June 15, 2008}
 \maketitle
 
@@ -666,34 +667,42 @@
 
 \subsection{Message Transfer}
 
-\begin{figure}
-\centering
-\includegraphics[width=0.8\textwidth]{reqtime.png}
-\begin{tabular}{rrrrrr}
-Min. & 1st Qu. & Median & Mean & 3rd Qu. & Max.\\\hline
-0.061 & 1.447 & 2.561 & 11.350 & 5.164 & 2752.000\\
-\end{tabular}
-\caption{Request transmission times}
-\label{fig:reqtime}
-\end{figure}
+Figure~\ref{fig:reqres} contains a plot of mean request times versus mean
+response times for each test. Data points on the diagonal represent tests
+with equal mean request and mean response time. Most points seem clustered
+around the line. However, there is a tendency for requests to be longer
+than responses which is also supported by mean and median.
 
 \begin{figure}
 \centering
-\includegraphics[width=0.8\textwidth]{restime.png}
-\begin{tabular}{rrrrrr}
-Min. & 1st Qu. & Median & Mean & 3rd Qu. & Max.\\\hline
-0.061 & 1.500 & 2.545 & 8.568 & 4.720 & 2931.000\\
+\includegraphics[width=0.8\textwidth]{reqres.png}
+\begin{tabular}{lrrrrrr}
+& Min. & 1st Qu. & Median & Mean & 3rd Qu. & Max.\\\hline
+Mean request time & 0.312 & 2.302 & 4.154 & 7.596 & 8.234 & 143.300\\
+Mean response time & 0.298 & 2.184 & 3.682 & 6.560 & 6.911 & 135.000
 \end{tabular}
-\caption{Response transmission times}
-\label{fig:restime}
+\caption{Mean request and response times}
+\label{fig:reqres}
 \end{figure}
 
-\emph{TODO Karsten: What to do with the results in
-Figures~\ref{fig:reqtime} and \ref{fig:restime}? are messages ordered? what
-about the phenomenon of large response times and small request times for
-the same connection? Any insights from variable correlations here (add
-diagram)?}
+Mean request or response times of 100 seconds or more are almost
+inacceptable for most applications. A look into the log statements
+confirmed that these mean values result from normal connections and not
+from short-lived connections with few, very high message transfer times.
 
+Further, in some cases mean request times are up to two orders of magnitude
+higher than mean response times and vice versa. This is rather surprising,
+because both messages are routed via the same Tor circuit.
+
+At the moment, message transfer times raise more questions than can be
+answered. However, these questions are not directly related to hidden
+services, but also apply to regular Tor circuits. A solution to improve
+message transfer times would probably require changes to the core Tor
+design including its path selection algorithm. Therefore, message transfer
+times will not be further considered in this attempt to speed up hidden
+services. The presented data may nevertheless be helpful for later attempts
+to speed up Tor message transfer times in general.
+
 \subsection{Connection Durability}
 
 Figure~\ref{fig:duration} shows connection durations.
@@ -701,6 +710,10 @@
 maximum allowed test time of four hours after a connection has been
 established. The unexpectedly high number of 650 out of 962 (67.57\%)
 established connections did not break within the first 3:45 hours.
+Considering the set of connections that broke before 3:45 hours elapsed,
+there is no noticeable point of time at which connections break, but
+connection breaks seem to be equally distributed over the whole time
+interval.
 
 \begin{figure}
 \centering
@@ -713,22 +726,10 @@
 \label{fig:duration}
 \end{figure}
 
-The remaining part of breaking connections is displayed in
-Figure~\ref{fig:broken}. There is no noticeable point of time at which
-connections break, but connection breaks seem to be equally distributed
-over the whole time interval.
+Hidden service connections appear to be quite stable, so that there is no
+need to put special focus on it in the attempt to improve the hidden
+service protocol.
 
-\begin{figure}
-\centering
-\includegraphics[width=0.8\textwidth]{broken.png}
-\caption{Connection durations of breaking connections}
-\label{fig:broken}
-\end{figure}
-
-Altogether, hidden service connections appear to be quite stable, so that
-there is no need to put special focus on it in the attempt to improve the
-hidden service protocol.
-
 \section{Discussion}
 
 Ideas what changes are most likely to improve the overall performance.

Added: projects/hidserv/trunk/doc/reqres.png
===================================================================
(Binary files differ)


Property changes on: projects/hidserv/trunk/doc/reqres.png
___________________________________________________________________
Name: svn:mime-type
   + application/octet-stream