NOTE: This transcription was contributed by Martin P.M. van der Burgt, who has devised a process for producing transcripts automatically. Although its markup is incomplete, we believe it serves a useful purpose by virtue of its searchability and its accessibility to text-reading software. It will be replaced by a fully marked-up version when time permits. —HR


In reaction to Ernest Chang’s “Deadlock Detection”.

We consider a finite directed graph, without edges of which both ends coincide, and introduce the following terminology. Let there be an edge directed from node A to node B ; then
—     this edge will be denoted as “edge AB ”
—     it is an “outgoing edge” of node A
—     it is an “incoming edge” of node B
—     node A is a “predecessor” of node B
—     node B is a “successor” of node A .

Let “node B is reachable from node A” mean “node B is node A or node B is a successor of a node that is reachable from node A ” .

Let “nodes A and B belong to the same strong component” mean “node A is reachable from node B and node B is reachable from node A ”.

Let a “knot” be a strong component of more than one node, such that all successors of all nodes of the knot belong to the knot as well.

*              *
*

In each node the identities of its predecessors and of its successors are recorded; a node contains no further information about the shape of the graph. Furthermore each node can send messages to its successors and to its predecessors —note that, in spite of its being directed, each edge accommodates two-way traffic— .

We can now ask ourselves the question whether we can design a diffusing computation, to be fired by an arbitrary node D , that will enable node D to detect whether or not it belongs to a knot. Because the answer is clearly “No” when node D has no successors —a circumstance that would be recorded in node D itself— we can restrict ourselves to the case that D has at least one successor. In the latter case the statement “node D belongs to a knot” is equivalent to the statement “node D is reachable from each node that is reachable from node D”.

From the last formulation it follows quite clearly that the determination that node D does belong to a knot involves all nodes that are reachable from node D . To establish for each node whether it is reachable from node D or not, is the first task to which we turn our attention.

*              *
*

For the sake of brevity we introduce the predicate Q1, defined as Q1(X) = “node X is reachable from node D ” .

In order to record for each node whether Q1 has been established, we associate with each node X a variable —of type “natural number”— that we denote by Xn ; the “significance” of its value follows from the relation

P1(X):Xn < 1 or Q1(X) ,
a relation that the algorithm shall keep invariant for each node X . Because, to start with, nothing about reachability from node D has been established, we assume the —in view of P1 : only safe!— initialization Xn = 0 for each node.

Nodes can send to their successors so-called M1-messages, meaning “the recipient of an M1 message is reachable from node D ”. We can justify this meaning by guarding the sending of an M1-message from node A to node B by An ≥ 1 , because this guard implies Q1(A) on account of P1(A) . In order to suppress (logically harmless, but superfluous) further transmission of M1-messages along that edge, the guard is further strengthened by a boolean —associated with that edge and denoted by M1AB — which is reset to false with the first transmission of an M1-message along that edge. For each edge AB , the boolean variable M1AB is assumed to be stored in node A and to be initialized at false.

The sending of an M1-message from node A to node B can now be represented by the guarded command

              An ≥ 1 and M1AB → {Ql(A) and Ql(B)}
                  MlAB:= false;
if Bn > l → skip
  ▯ Bn = 0 → Bn:= 1; (A X: X is a successor of B: MlBX:= true)
fi {Pl(B)}
The first two lines describe the sending; the following alternative construct describes the reception .

Note that —thanks to our initialization and the second alternative— we have also the invariant

An ≥ 1 or non M1AB
for each edge AB (and that, as a result, the sending of an M1-message from node A to node B could have been guarded by M1AB all by itself).

The diffusing computation can be started by node D performing

Dn:= 1; (A X: X is a successor of D : M1DX:= true)
an act, which maintains all invariants. Because the graph is finite and each edge carries at most one message, the diffusing computation will die out; from the fact that then no more messages are sent, we conclude that in the final state we have again non M1AB for each edge AB . Observing for the edge AB the invariant
An = 0 or Bn ≥ 1 or M1AB
we conclude that in the final state we have for each edge AB
An = 0 or Bn ≥ 1 .
In the final state we can now conclude for each node X
Xn ≥ 1 or non Q1(X)
as follows. For a node X such that non Q1(X) holds it is obviously true. For a node X such that Q1(X) holds, there exists a path from D to X , and our previous result tells us that on that path a node with n ≥ 1 can only be followed on that path by a successor with n ≥ 1 as well. Hence, from Dn ≥ 1 we conclude Xn ≥ 1 , thereby establishing our latest result. Combining our latest result with P1 we conclude that upon termination we have for all nodes X
(Xn ≥ 1) = Q1(x) .

*              *
*

Recalling our interest in the truth of “node D is reachable from each node that is reachable from node D ” and having dealt with the reachability from node D , we now turn our attention to the question whether node D is reachable from a node.

For the sake of brevity we introduce the predicate Q2 , defined as Q2(X) = “node D is reachable from node X ” and because we are only interested in recording for nodes whether Q2 has been established provided Q1 holds for them as well, we can use for each node X the same variable Xn by virtue of the relation

P2(X): Xn < 2 or Q2(x) .
a relation that the algorithm shall keep invariant for each node. In view of our initialization Xn = 0 relation P2 holds to start with for each node. Because our algorithm as developed so far only assigns the value 1 to Xn’s , that earlier part of the algorithm leaves P2 invariant. We change, however, the way in which node D starts the diffusing computation into
Dn:= 2; (A X: X is a successor of D : M1DX:= true) ,
i.e. at the start node D records that node D is reachable from itself.

Nodes can send to their predecessors so-called M2-messages, meaning “node D is reachable from the recipient of an M2-message”. We can justify this meaning by guarding the sending of an M2-message from node B to its predecessor A by Bn ≥ 2 , because that guard implies Q2(B), from which Q2(A) follows on account of the existence of the edge AB . In order to suppress logically harmless but superfluous transmissions of M2-messages, the guard is further strengthened by a boolean —associated with that edge and denoted by M2BA— which is reset to false with the first transmission of an M2-message along that edge. For edge AB the boolean variable M2BA is assumed to be stored in node B and to be initialized at false. The boolean M2BA is also used to suppress a superfluous M2-message from node B to a predecessor A for which Q1(A) has not (yet) been established. Because the establishement of Q1(A) takes place in node A and M2BA is stored in node B , we change in the transmission of an M1-message from node A to node B the reaction of the recipient by including the setting of M2BA.

The transmission of an M1-message from node A to node B can now be represented by the guarded command

            An ≥ 1 and MlAB →
                  MlAB:= false;
M2BA:= true;
if Bn ≥ 1 → skip
  ▯ Bn = 0 → Bn:= 1; (A X: X is a successor of B: MlBX:= true)
fi .
The transmission of an M2-message from node B to node A can now be represented by the guarded command
            Bn > 2 and M2BA →
                M2BA:= false;
if An ≥ 2 → skip
  ▯ An = 1 → An:= 2
fi .
The above traffic keeps clearly P1(X) and P2(X) invariant for all nodes. Because the graph is finite and each edge transmits at most one M1-message and hence at most one M2-message, the diffusing computation will die out; from the fact that then no more messages are sent we conclude that in the final state we have for each edge AB
non M1AB and (Bn < 2 or non M2BA) .
Observing for the edge from node A to node B the invariant
An ≠ 1 or M1AB or M2BA .
we conclude, combining the last two relations, that in the final state we have
An ≠ 1 or Bn < 2
for each edge AB . In the final state we can now conclude
Xn ≥ 2 or non (Q1(X)) and Q2(x))
as follows. For a node X such that non (Q1(X) and Q2(X)) holds it is obviously true. For a node X such that (Q1(X) and Q2(X)) holds,there exists a path from X to D , all nodes of which have n ≥ 1 . Our latest result tells us that on that path a node with n ≥ 2 can only be preceded by a node with n ≠ 1 , hence n ≥ 2 . Because the path ends in node D with Dn ≥ 2 , we conclude Xn ≥ 2 , thereby establishing our latest result. Combining our latest result with P2 we conclude for the final state
non Q1(X) or ((Xn ≥ 2) = Q2(X)) .

Summing up: in the final state we have
a node with Xn = 0 satisfies non D1(X) ,
a node with Xn = 1 satisfies Q1(X) and non Q2(X)
a node with Xn = 2 satisfies Q1(x) and Q2(x) .

Note that after firing —i.e. autonomously setting its Dn to 2 and the M1DX’S for its outgoing edges to true— node D acts just like any other node of the graph. Note also that all the ways in which the M1- and M2-message traffic may be mixed in time —a node X with Xn = 2 may still receive M1-messages— did not need to be mentioned in our argument.

*              *
*

We now turn our attention to the third part of the algorithm, viz. the detection that D is reachable from all nodes that are reachable from D . For the sake of the inductive argument required we introduce a rooted tree T with node D as its root, with edges from the given graph and subspanning all nodes reachable from D . (The definition of “reachable from node D ” implies the existence of such a tree,) Its edges will be called the “T-edges”.

Let “the subtree of node X ” be defined as that subtree of T of which node X is the root. For the sake of brevity we introduce the predicate Q3, defined as

Q3(X) =“node X is reachable from node D and node D is reachable from all nodes of the subtree of node X ” .

Because the subtree of node D is T itself, and T subspans by definition all nodes reachable from node D , the value of D3(D) is the answer we are looking for. Because Q3(X) ⇒ (Q1(X) and Q2(X)) , we can use for the recording whether Q3 has been established for each node the same variable Xn by virtue of the relation

P3(X):Xn < 3 or Q3(X)
which is clearly an invariant of our algorithm developed thus far.

Nodes can send to their predecessors so-called M3-messages meaning “Q3 holds for the sender of this M3-message”. A node will send an M3-message with this meaning at most once, and only along its one and only incoming T-edge. With each T-edge we associate a boolean variable that will be used to record whether it has carried an M3-message. More precisely, with each T-edge AB we associate a boolean variable that is denoted by M3BA and assumed to be stored in node A . For the T-edge AB we shall keep the relation

non M3BA or (Bn ≥ 3)(1)
invariant; it is initially true because M3BA is initialized with the value false.

For a node X we may conclude that Q3(X) holds, provided
1)     Q3(Y) holds for all “T-successors Y of X” —where “Y is a T-successor of X” means “the edge XY is a T-edge”— and
2)     Q1(X) and Q2(X) holds —note that X could be a leaf of T!—

The sending of an M3-message from B to A along its one an only incoming T-edge can now be represented by the guarded command

      Bn = 2 and (A X: X is a T-successor of B: M3XB) →
                Bn:= 3;
M3BA:= true .
For node D , which has no incoming T-edge, this reduces to recording that Q3(D) has been established:
      Dn = 2 and (A X: x is a T-successor of D: M3XD) →
                Dn:= 3 .
It is clear that with this addition the algorithm leaves P1(X) and P2(X) and P3(X) invariant for each node X . Note that for the M3BA associated with the T-edge AB we have the invariant stronger than (1):
M3BA = (Bn ≥ 3) .
The proof that upon completion we have
(Dn > 3) = Q3(D)
is left to the reader.

*              *
*

We are now only left with the task of choosing the rooted tree T . We choose for the T-edges those edges that carry an M1-message that is accompanied by the transition from n = 0 to n = 1 for its recipient. It is clear that each node X with Q1(X) is reachable from D via such edges; it is also clear that each X with Q1(X) that differs from D has exactly one such incoming edge. Hence they satisfy all properties required from the edges of the rooted tree T .

With this choice for the tree T , the component of the guard controlling the M3-traffic

(A X: X is a T-successor of B: M3XB)
presents some implementation problems, because the identities of their T-successors are not recorded in nodes. We therefore associate also a boolean variable M3BA with each non-T-edge AB , and replace the above guard component by
(A X: X is a successor of B: M3XB)
and introduce a further mechanism that can set the M3BA to true for each non-T-edge AB . Because upon arrival of an M1-message its recipient can decide whether it received the M1-message via a T-edge or via a non-T-edge, we can give each recipient of an M1-message via a non-T-edge the obligation to return along that edge an M3-message meaning “the edge along which this M3-message is carried is not a T-edge”. The recipient of that M3-message need not distinguish between the two meanings of an M3-message: in both cases it sets the M3BA for that edge to true. With this protocol we would eventually establish for each non-T-edge AB
non Q1(A) or (M3BA = (Bn ≥ 1)) .

In the case Q3(D) each edge that has carried an M1-message has to carry an M2- and an M3-message as well. In the case non Q3(D) —characterized by the existence of a node X with Q1(X) and non Q2(X)— it would be permissible to suppress all M3-traffic, and, from the point of efficiency, the more we suppress, the better. We can achieve suppression by strengthening the guards controlling the M3-signalling via T-edges by establishing for a non-T-edge AB eventually

non Q1(A) or (M3BA = (Bn ≥ 2))
Also in the case Q3(D) this strengthening of the guards is permissible: in that case —characterized by the absence of a node X with Q1(X) and non Q2(X)— each M3BA of an outgoing non-T-edge of a reachable node will be set to true, and our initially strengthened component of the guard controlling the M3-traffic along T-edges becomes equivalent to the original
(A X: X is a T-successor of B: M3XB) .
This means, however, that along non-T-edges the M2- and M3-messages can be combined into M23-messages.

Incorporating this optimization we get the following solution. For a non-T-edge it maintains the invariant

An = 0 or (M3BA = (Bn ≥ 2 and non M1AB and non M2BA)
The start of the computation by node D :
Dn:= 2; (A X: X is a successor of D: M1DX:= true)
The sending of an M1-message from A to its successor B :
          An ≥ 1 and M1AE →
          M1AB:= false;
M2BA:= true;
if Bn ≥ 1 → skip
  ▯ Bn = 0 → “record that AB is B’s incoming T-edge”;
        Bn:= 1; (A X: X is a successor of B: M1BX:= true)
fi
The sending of an M2-message from E to A via B’s incoming T-edge:
          Bn ≥ 2 and M2BA →
          M2BA:= false;
if An ≥ 2 → skip
  ▯ An = 1 → An:= 2
fi
The sending of an M23-message from B to A via an incoming non-T-edge:
          Bn ≥ 2 and M2BA →
          M2BA:= false;
if An ≥ 2 → skip
  ▯ An = 1 → An:= 2
fi ; M3AB:= true
The sending of an M3-message from B to A via B’s incoming T-edge:
          Bn = 2 and (A X: X is a successor of E: M3XB) →
  Bn:= 3;
M3BA:= true
For node D , which has no incoming T-edge, this reduces to recording that Q3(D) has been established:
            Dn = 2 and (A X: X is a successor of D: M3XD) →
  Dn:= 3

Note that we did not prescribe that along T-edges the M3-messages are preceded by the M2-messages.

*              *
*

The algorithm described above is a diffusing computation that dies out. In the case action from D is only required if it belongs to a knot, it suffices —although it leaves “residues” in all other nodes— . Alternatively we can superimpose the detection mechanism for diffusing computations [1]: upon receipt of the termination signal in D , the truth of

(Dn ≠ 3) = Q3(D)
is guaranteed. (Note that each directed edge of the above, carrying messages of the diffusing computation possibly in both directions, corresponds to two edges —in opposite directions— in the terminology of [1]. Note further that node D of the above, which can receive messages, is not the “environment” of [1], and that a (virtual) “environment” has to be added to the graph that fires D . Note further that the “engagement edges” from [1] are not restricted to the T-edges from the above.) I shall not carry out the superposition.

*              *
*

The above has been prompted by Ernest Chang [2], in which the author —following R.C.Holt— reduces deadlock detection to the question whether a node —in a bipartite directed graph of “processes” and “resources”— belongs to a knot. Chang restricts himself to a diffusing computation of which termination is not necessarily detected in the case non Q3(D). (Chang’s algorithm —if correct— recognizes non Q3(D) explicitly in the trivial case when there exists a node X with Q1(X) , but without successors at all.)

The main incentive for writing the above was that while reading [2] I could not decide whether its author had solved the problem or not; neither could I decide —if he has solved it— whether all complexity of his solution is really required.

Warning. Those that would like to study Chang’s [2] should be warned that in his programs he attaches —without explaining his conventions— a syntactic role to indentation. (End of warning.)

After spending a number of hours studying [2] I had placed mentally so many question marks in margine that I decided that there was only one way to understand his problem: I had to put his text aside and try to solve the problem myself. Hence this text.

After solving the problem (without pencil and paper, within a few hours) I spent two days on its first presentation, which was rejected because —though much less so than [2]— parts of my argument were still too operational. (I had also made the mistake of borrowing some of Chang’s —operationally defined! terminology, such as “primary” and “secondary” edges.) The writing of the above text —my second effort— took me another three full days, days that I consider as extremely well-spent. It forced me to separate my four main concerns —Q1(X), Q2(X), Q3(X), and the definition of the rooted tree T — even more rigorously than I had done before; but even after that had been achieved, finding the proper notations and deciding upon the proper amount of essential detail turned out to be a major challenge. (I hope that I have met it well.)

Note. For the operationally minded I point out that there is no objection to regarding the tree T as completely defined, all through the computation. The objection that, at the beginning of the computation, T is not “known” yet, is totally irrelevant: with a modest amount of clairvoyance it is. (End of Note.)

I must admit that, in the meantime I have lost all incentive to return to [2] in order to try, once again, whether I can convince myself that Chang’s solution is correct. If I had to referee the paper, I would give the author on this point the benefit of the doubt, but would recommend unconditional rejection of the paper in its current form, because it is much faster to solve the problem yourself than to try to understand his text. I am also not tempted to investigate to what extent the complexity of his program is due to a clumsy coding or to further optimizations that I did not incorporate —such as the combination of an M2- and an M3-message into a single M23-message along a T-edge, whenever possible— , but he did not disentangle.

Acknowledgement. I am grateful to the members of the Tuesday Afternoon Club for explaining with refreshing clarity to me why my first version had to be rejected.

[1]Dijkstra, Edsger W. and Scholten, C.S., Termination detection for diffusing computations. (EWD687a, submitted to ACTA INFORMATICA)
[2]Chang, Ernest, Decentralized Deadlock Detection in Distributed Systems. (University of Toronto)

21 February 1979

Plataanstraat 5prof.dr.Edsger W.Dijkstra
5671 AL NUENENBurroughs Research Fellow
The Netherlands

Transcribed by Martin P.M. van der Burgt
Last revision 2015-02-06 .