From ddd5c6feddd7bc7a1914d2a91eef2f3159d94cb8 Mon Sep 17 00:00:00 2001 From: bdemsky Date: Mon, 25 Jul 2016 13:02:10 -0700 Subject: [PATCH] notes --- doc/iotcloud.tex | 16 +++++++++------- 1 file changed, 9 insertions(+), 7 deletions(-) diff --git a/doc/iotcloud.tex b/doc/iotcloud.tex index ed78c2a..d4f8045 100644 --- a/doc/iotcloud.tex +++ b/doc/iotcloud.tex @@ -686,11 +686,11 @@ $\tuple{ck,\tuple{k, v}} \in KV_s \wedge \item Equality: Two messages $t$ and $u$ are equal if their sequence numbers, senders, and contents are exactly the same. \item Message: A message $t$, is the tuple $t = (i(t), s(t), contents(t))$ containing the sequence number, machine ID of the sender, and contents of $t$ respectively. \item Parent: A parent of a message $t$ is the message $A(t)$, unique by the correctness of HMACs, such that $HMAC_C(t) = HMAC_P(A(t))$. -\item Chain: A chain of messages with length $n \ge 1$ is a message sequence $(t_i, t_{i+1}, ..., t_{i+n-1})$ such that for every index $i < k \le i+n-1$, $t_k$ has sequence number $k$ and is the parent of $t_{k-1}$. Note that no two entries in a chain can have the same sequence number. +\item Chain: A chain of messages with length $n \ge 1$ is a message sequence $(t_i, t_{i+1}, ..., t_{i+n-1})$ such that for every index $i < k \le i+n-1$, $t_k$ has sequence number $k$ and is the parent of $t_{k-1}$. Note that no two entries in a chain can have the same sequence number.\note{This definition is never used, instead you refer to paths..} \item Partial message sequence: A partial message sequence is a sequence of messages, no two with the same sequence number, that can be divided into disjoint chains. \item Total message sequence: A total message sequence $T$ with length $n$ is a chain of messages that starts at $i = 1$. The path of a message $t$ is the total message sequence whose last message is $t$. \item Consistency: A partial message sequence $P$ is consistent with a total message sequence $T$ of length $n$ if for every message $p \in P$ with $i(p) < n$, $T_{i(p)} = p$. This implies that $\{p \in P | i(p) \le n\}$ is a subsequence of T. -\item Transitive closure set at index $i$: A set $\mathscr{S}$ of clients comprising a connected component of an undirected graph, where two clients are connected by an edge if they both received the same message $t$ with index $i(t) > i$. +\item Transitive closure set at index $i$: A set $\mathscr{S}$ of clients comprising a connected component of an undirected graph, where two clients are connected by an edge if they both received the same message $t$ with index $i(t) > i$.\note{Confusing to reuse i} \end{enumerate} @@ -702,7 +702,7 @@ $\tuple{ck,\tuple{k, v}} \in KV_s \wedge \begin{prop} If a rejected message entry is added to the RML at index $i$, the message will remain in the RML until every client has seen it. \end{prop} \begin{proof} Every RML entry $e$ remains in the queue until it reaches the tail, and is refreshed by the next sender $J$ at that time if $min(MS) > i(e)$; that is, until every client has sent a message with sequence number greater than $i(e)$. Because every client who sends a message with index $i$ has the state of the queue at $i - 1$, this client will have seen the message at $i(e)$. \end{proof} -\begin{lem} If two packets $t$ and $u$, with $i(t) \le i(u)$, are received without errors by a client $C$, then $t$ is in the path of $u$. \end{lem} +\begin{lem} If two packets $t$ and $u$, with $i(t) \le i(u)$, are received without errors by a client $C$, then $t$ is in the path of $u$. \note{path never defined}\end{lem} \begin{proof} Assume that $t$ is not in the path of $u$. Take $u$ to be the packet of smallest index for which this occurs, and $t$ be the packet with largest index for this $u$. We will prove that an error occurs upon receipt of $u$. @@ -713,14 +713,14 @@ Let $J = s(R_1)$, and $K = s(S_1)$. Because no client can send two messages with There are two cases: \begin{itemize} -\item Case 1: $J$ did not send a message in $S$. Then $v_J(t) > v_J(q) = v_J(u)$. $C$ will throw an error, because the latest index of $J$ changes in the opposite direction of the sequence number: $v_J(u) < v_J(t)$ but $i(u) > i(t)$. +\item Case 1: $J$ did not send a message in $S$. Then $v_J(t) > v_J(q) = v_J(u)$. $C$ will throw an error, because the latest index of $J$ changes in the opposite direction of the sequence number: $v_J(u) < v_J(t)$ but $i(u) > i(t)$.\note{Need more contet as to why C would throw an error...You are assuming that it reads two disjoint windows...one with u and the other with t...other case is that they are contiguous, in which case the hashes can't match...} \item Case 2: $J$ sent at least one message in $S$. Call the first one $p$. We know that $i(p) > i(S_1)$, since $J \neq K$. $R_1$ must be sent either before or after $p$. \begin{itemize} -\item Case 2.1: Client $J$ sends $p$, and then $R_1$. When $p$ was sent, whether it was accepted or rejected, $i(J, p) \geq i(p)$. Since $i(p) > i(S_1)$, $i(J, p) > q$. So $i(q) < i(J, p)$, which would cause $J$ to fail to send $R_1$, a contradiction. +\item Case 2.1: Client $J$ sends $p$, and then $R_1$. When $p$ was sent, whether it was accepted or rejected, $i(J, p) \geq i(p)$. Since $i(p) > i(S_1)$, $i(J, p) > q$. So $i(q) < i(J, p)$, which would cause $J$ to fail to send $R_1$, a contradiction.\note{Never defined i(J,p). don't you mean i(q)+1 instead of q?} \item Case 2.2: Client $J$ sends $R_1$, and then $p$. Let $X = (R_1, \dots )$ be the list of messages $J$ sends starting before $R_1$ and ending before $p$. \begin{itemize} \item Case 2.2.1: Some message in $X$ was accepted. In this case, before sending $p$, $J$'s value for its own latest index would be strictly greater than $v_J(q)$. ($J$ could not have sent a message with index less than $i(q)$ after receiving $q$). When preparing to send $p$, $J$ would have received its own latest index as at most $v_J(q)$. $J$ throws an error before sending $p$, because its own latest index decreases. -\item Case 2.2.2: All messages in $X$ were rejected. Client $J$ will always add the latest rejected message to the rejected-message list in the next update; so for every $i$, $1 \leq i < |X|$, the $i$th element of $X$ will be recorded in the RML of all further elements of $X$; and every element of $X$ will be recorded in $RML(p)$. Since every rejected message in $RML(p)$ will be in $RML(C, u)$, and $u$ is the first message that $C$ sees which does not have $t$ in its path, $R_1$ will be recorded in $RML(C, p)$. When $C$ receives $u$, $C$ will throw an error from the match $(J, iq+1)$ in $RML(C, p)$. +\item Case 2.2.2: All messages in $X$ were rejected. Client $J$ will always add the latest rejected message to the rejected-message list in the next update; so for every $i$, $1 \leq i < |X|$, the $i$th element of $X$ will be recorded in the RML of all further elements of $X$; and every element of $X$ will be recorded in $RML(p)$. Since every rejected message in $RML(p)$ will be in $RML(C, u)$, and $u$ is the first message that $C$ sees which does not have $t$ in its path, $R_1$ will be recorded in $RML(C, p)$. When $C$ receives $u$, $C$ will throw an error from the match $(J, iq+1)$ in $RML(C, p)$.\note{Missing case...what if C switches branches before J? It will see a version rollback...Also need more context...disjoint reads...etc...} \end{itemize} \end{itemize} \end{itemize} @@ -730,13 +730,15 @@ There are two cases: Suppose that there is a transitive closure set $\mathscr{S}$ of clients, at index $n$. Then there is some total message sequence $T$ of length $n$ such that every client $C$ in $\mathscr{S}$ sees a partial sequence $P_C$ consistent with $T$. \end{theorem} \begin{proof} -The definition of consistency of $P_C$ with $T$ is that every message $p \in P_C$ with index $i(p) \le n$ is equal to the message in that slot in $T$. Let $C_1$ be some client in the transitive closure set, with partial message sequence $P_{C_1}$, and let $u$ be some message with $i(u) > i$ that $C_1$ shares with another client. Then let $T$ be the portion of the path of $u$ ending at index $i$ and $t$ be the message at that index. Clearly, by Lemma 1, $P_{C_1}$ is consistent with $T$, and furthermore. We will show that, for every other client $D$ with partial sequence $P_D$, $P_D$ has some message whose path includes $t$. Because $D$ is in the transitive closure, there is a sequence of edges from $C_1$ to $D$. Call this $\mathscr{C} = (C_1, C_2, ..., D)$. I will prove by induction that $D$ has a message whose path includes $t$. +The definition of consistency of $P_C$ with $T$ is that every message $p \in P_C$ with index $i(p) \le n$ is equal to the message in that slot in $T$. Let $C_1$ be some client in the transitive closure set, with partial message sequence $P_{C_1}$, and let $u$ be some message with $i(u) > i$ that $C_1$ shares with another client. Then let $T$ be the portion of the path of $u$ ending at index $i$ and $t$ be the message at that index. Clearly, by Lemma 1, $P_{C_1}$ is consistent with $T$, and furthermore. \note{typo in previous sentence?} We will show that, for every other client $D$ with partial sequence $P_D$, $P_D$ has some message whose path includes $t$. Because $D$ is in the transitive closure, there is a sequence of edges from $C_1$ to $D$. Call this $\mathscr{C} = (C_1, C_2, ..., D)$. I will prove by induction that $D$ has a message whose path includes $t$. For the base case, $P_{C_1}$ includes $u$, whose path includes $t$. For the inductive step, suppose $P_{C_k}$ has an message $w$ with a path that includes $t$, and shares message $x$ with $P_{C_{k+1}}$ such that $i(x) > i$. If $i(w) = i(x)$, then $w = x$. If $i(w) < i(x)$, then, by Lemma 1, $w$ is in the path of $x$. If $i(w) > i(x)$, $x$ is in the path of $w$; note again that its index is greater than $i$. In any case, $t$ is in the path of $u_k+1$. Let $w$ the message of $D$ whose path includes $t$. By Lemma 1, every message in $P_D$ with index smaller than $i(w)$ is in the path of $w$. Since $t$ is in the path of $w$, every message in $P_D$ with smaller index than $i(t)$ is in $T$. Therefore, $P_D$ is consistent with $T$. \end{proof} +\note{General comments...some diagrams could help...theorem 1 is pretty hard to follow...} + \subsection{Future Work} \paragraph{Support Messages} A message is dead once receiving machine sends an entry with a newer -- 2.34.1