Using Attributed Variables in the Implementation of

Using Attributed Variables in the Implementation of
Proceedings of the 1995 International Conference on Logic Programming
Using Attributed Variables in
the Implementation of Concurrent
and Parallel
M. Hermenegildo, D. Cabeza, M. Carro
Computer Science Department
Technical University of Madrid (UPM), Spain
Incorporating the possibility of attaching attributes to variables in a logic programming system has been shown to allow the addition of general constraint solving
capabilities to it. This approach is very attractive in that by adding a few primitives any logic programming system can be turned into a generic constraint logic
programming system in which constraint solving can be user dened, and at source
level { an extreme example of the \glass box" approach. In this paper we propose a
dierent and novel use for the concept of attributed variables: developing a generic
parallel/concurrent (constraint) logic programming system, using the same \glass
box" avor. We argue that a system which implements attributed variables and
a few additional primitives can be easily customized at source level to implement
many of the languages and execution models of parallelism and concurrency currently proposed, in both shared memory and distributed systems. We illustrate this
through examples and report on an implementation of our ideas.
Keywords: Implementation Techniques, Concurrency, Parallelism, Logic Programming, Attributed Variables, Generic Implementations.
1 Introduction
A number of concepts and implementation techniques have been recently introduced
which allow extending unication in logic languages in a exible and user-accessible
way. One example is that of meta-structures, introduced by Neumerkel [22], which
allow the specication by the user of how unication should behave when certain
types of terms, called meta-structures and marked as such by the user, are accessed
during unication. More or less at the same time, the data type attributed variable was introduced by Le Houitouze [18] with the purpose of implementing various
memory management optimizations. Although the behavior of attributed variables
during unication was not specied in this work, a number of applications were
proposed including the implementation of delayed computations, reversible modication of terms, and variable typing. Earlier, Carlsson [5] used a data type called
suspension, which was incorporated into SICStus Prolog [6] for the implementation
of coroutining facilities [10]. \Attributed variables" and \suspension variables" are
essentially the same objects. Le Houitouze's contribution was to put some emphasis on the data type as such and on memory management. He also used attributed
1 The work described in this paper is supported in part by ESPRIT project \ACCLAIM", and
by CICYT contract TIC93-0737-C02-01. The authors would like to thank C. Holzbaur and the
anonymous referees for useful discussions in the context of the paper.
variables as a low level primitive for the implementation of mechanisms that necessitated the specication of the behavior of the data type during unication.
A rened version of the concept of meta-structures and attributed variables
was used by Holzbaur in [16] for the specication and implementation of a variety
of instances of the general CLP scheme [19]. By enhancing the SICStus Prolog
system with attributed variables a generic system is provided which is basically a
SICStus Prolog \clone" where the unication mechanism has been changed in such
a way that the user may introduce interpreted terms and specify their unication
through Prolog predicates. This approach is very attractive in that it shows that
by adding a few primitives any logic programming system can be turned into a
generic constraint logic programming system in which constraint solving can be
user dened, at the source level { an extreme example of the \glass box" approach.
Another system which implements constraint solving using similar techniques is the
ECLi PSe system developed at ECRC [11].
While this approach in principle has drawbacks from the performance point of
view (when fully interpreted up to an order of magnitude slowdown is possible w.r.t.
native CLP systems) the convenience and generality of the approach can make it
very worthwhile in many cases. Furthermore, the speed can be easily increased in
interesting cases by writing the unication handlers in a lower-level language. The
potential for achieving both genericity and reasonable speed is illustrated by the
relatively good performance exhibited by the ECLi PSe system, which has been used
in many practical applications.
Inspired by the previously discussed use of attributed variables we propose a different and novel use for such variables in a completely dierent context: developing
generic parallel/concurrent (constraint) logic programming systems, using the same
\glass box" avor. Attributed variables have already been used to implement the
coroutining (delay) facilities present in many Prolog systems { often what is actually being done is in fact restoring such capabilities after having \cannibalized" the
delay mechanism support for implementing the attributed variables. However, we
argue that a system which implements both support for attributed variables and a
few additional primitives related to concurrency and parallelism can do much more
than simply restoring the delay mechanism. In fact, it is our thesis that using the
primitives mentioned above it is possible to easily implement many of the languages
and execution models of parallelism and concurrency currently proposed. We illustrate this through examples and we discuss how quite complex concurrent languages
and parallel execution models can be implemented using only such primitives. Furthermore, we argue that this can be done in a seamless and user-transparent way
in both shared memory and distributed systems. Also, one additional advantage of
our technique is that it relates and unies the two main approaches currently used
in concurrent logic programming, and which are seen traditionally as unrelated:
\shared variable" systems, in which communication among parallel tasks is done
through variables, and \distributed" or \blackboard" systems in which communication is done through explicit built-ins which access shared channels or global data
It should be noted that the use that we propose of attributed variables in the
implementation of concurrency and parallelism does not necessarily prevent their
simultaneous use also for other purposes, such as the original one of constraint
solving. Also, note that the approach proposed, although having some similarities,
diers from that of \generic objects," recently and independently discussed by the
KL1 implementors [8]. The idea in generic objects is to provide an interface at the
\C" level for a particular class of extensions. Our approach diers in both the level
at which the extensions are made (which is completely at the source level in our
case, thus really oering a reective, glass box approach), and the nature and extent
of the extensions proposed, which goes beyond those that are related to supporting
Space limitations force the presentation to cover only some basic cases and give
incomplete descriptions of the implementations. For more details the reader is
referred to [14]. Also, the implementations described can be obtained by contacting
the authors.
2 Attributed Variables and Related Primitives
We provide a brief introduction to attributed variables. For concreteness, we follow a
stylized version of Holzbaur's rst implementation of attributed variables in SICStus
Prolog. The reader is referred to [16, 17] for more detailed information.
2.1 General Concepts
Attributed variables are variables with an associated \attribute." Attributes are
terms which are attached to variables, and which are accessed in a special way during unication and also through special built{in predicates. As far as the rest of a
given Prolog implementation is concerned, attributed variables behave like variables.
Special treatment for attributed variables does apply mainly during unication: as
will be described later, when an attributed variable is to be unied with another
attributed variable or some other non-variable term, user-dened predicates specify
how this unication has to be performed. The following is a list of typical predicates which provide for the introduction, detection, and manipulation of attributed
variables. In general, attributed variable related operations are correctly undone
upon backtracking.
get attr(X,C): if X is an attributed variable, unify the corresponding attribute with C, otherwise fail.
attach attr(X,C): turn the free variable X into an attributed variable with
attribute C.
detach attr(X): remove the attribute from an attributed variable, turning it
into a normal variable.
Attributed variables are dealt with specially during unication. Essentially, the
dierent possible cases are handled as follows. A unication between an unbound
variable and an attributed variable binds the unbound variable to the attributed
variable. When an attributed variable is about to be bound during unication
to a non-variable term or another attributed variable, the attributed variable and
the value it should be bound to are collected. At the next inference step, the
pending attributed variable-value pairs are supplied to user-dened handlers which
are dened by the user by means of the following predicates:
verify attr(C,T): invoked when an attributed variable with an attribute
which unies with C is about to be unied with the non-variable term T.
combine attr(C1,C2): invoked when two attributed variables with attributes
C1 and C2 are about to be unied.
Note that the two predicates are not called with the attributed variables involved, but with the corresponding attributes instead. The is done for reasons of
simplicity and eciency (e.g. indexing). Note also, however, that if access to the actual attributed variable is needed the variable itself can be included in the attribute
when it is attached. In general, a number of other primitives are often provided
which allow pretty printing and dumping of the results in a user{understandable
2.2 Attributed Variables And Coroutining { an Example
The following example, due to [17] serves both to illustrate the use of the primitives
introduced in the previous section and also to recover the functionality of freeze/2
[10] since attribute variables are, as mentioned before, most easily implemented in
practice by \cannibalizing" an existing implementation of freeze:
freeze(X, Goal) :attach_attr(V, frozen(V,Goal)),
X = V.
verify_attr(frozen(V,Goal),Val) :detach_attr(V),
V = Val,
frozen(V2,G2)) :detach_attr(V1),
V1 = V2,
The call to attach attr ties the term representing the frozen goal to the relevant
variable. When the variable is bound the unication routine escapes to the userdened generic handler verify attr which in turn performs the meta-call. Note
the denition of combine attr needed for handling the case where two variables
which have frozen goals attached are unied: a conjunction of the goals is attached
to the resulting variable.
Note that the explicit encoding of delay primitives such as freeze/2 and their
incorporation into the attributed variable handling mechanism is not to be understood as a mere substitute for the original C code. The true motivation for explicit
encodings is that it enables the user to freely dene the combination and interaction
of such delay primitives with other uses of the attributed variables such as the implementation of a constraint solver. Note that such a solver may also itself perform
some delaying, for example when dealing with non-linear constraints.
3 Kernel Concurrent Primitives
We now introduce a simple concurrent/parallel language that we call \Kernel &Prolog" (K&P). The purpose of this language is to provide a small set of basic
operators which will allow the implementations that we would like to propose. This
language is essentially identical to the kernel language used in the shared memory
[15] and distributed [13] implementations of the &-Prolog system, but it is described
here for the rst time.
Essentially, the K&P language subsumes Prolog and includes all the attributed
variable primitives described in Section 2. In addition, it provides the following
operators which provide for creation of processes, assignment of computational resources to them, and synchronization:
{ Standard fork/join parallel conjunction operator (the one used, for
example, by the &-Prolog parallelizing compiler [4]). It performs a parallel
\fork" of the two literals involved and waits for the execution of both literals
to nish (i.e. the join). If no processors are available, then the two literals
may be executed in the same processor and sequentially, i.e. one after the
other. For example, ..., p(X) & q(X), r(X), ... will fork a task p(X)
in parallel with q(X). The continuation r(X) will wait until both p(X) and
are completed.2 The implementation of this primitive at the abstract
machine level is well understood [15].
&&/2 { \fair" fork/join parallel conjunction operator. It performs a parallel
fork of the two literals involved and waits for the execution of both literals to
nish (join). A \thread" is assigned to each literal. The execution of the two
literals will be interleaved either by executing them on dierent processors
(if they are available) or by multiplexing a single processor. Thus, even if
no processors are available, the two literals will be executed with (apparent)
simultaneity in a fair way.
&/1 { Standard fork operator. It performs a parallel fork of the literal involved.
No waiting for its return is performed (unless explicitly expressed using the
wait primitive { see below). For example, ..., p(X) &, q(X), r(X), ...
will fork a task p(X) in parallel with the rest of the computation.
&&/1 { \fair" version of the fork operator.
&@/2 { \Placement" standard fork operator. It performs a parallel fork of
the literal involved, assigning it to a given node. No waiting for its return is
involved. If that node is busy, then the literal will eventually be executed in
that node when it becomes idle. For example, ..., p(X) &@ node , q(X),
... will fork the task p(X) in parallel with the rest of the computation and
assign it to node node.3 The second argument can be a variable. If the variable
is instantiated at the time the literal is reached, its value is used to determine
its placement. If the variable is unbound at that time, then the goal is not
assigned to any particular node and the variable is bound to the node id. of
the node that picks up the task, when it does so.
&&@/2 { \fair" placement fork operator. It performs a parallel fork of the
literal involved, assigning it to a given node and nding (or, if not available,
creating) a thread for it in that node.
wait(X): This primitive suspends the current execution thread until X is
bound. X can also contain a disjunction of variables, in which case execution waits for either one of such variables to be bound.
lock(X,L)/unlock(L): This primitive gets/releases a lock L (on the term X).
Note that in the discussion above a (parallel) conjunction of literals can always
be used in place of a literal, i.e. the expression ..., (a,b) & (c, d & e, f), ...
is supported. Also, note that the \placement" primitives (&@/2 and &&@/2) and the
wait/lock primitives are sucient to express all the other primitives.
In addition to the \placement" operators described above, which can be directly
used in distributed environments, the language also provides as base primitives a
Linda-like [7] library, and a lower-level Unix socket interface both of which reproduce
the functionality of those of SICStus Prolog. In fact, in distributed environments the
primitives described above are implemented using the Linda library [13]. However,
the Linda interface can also be used directly: there is a server process which handles
the blackboard. Prolog client processes can write (using out/1), read (using rd/1),
2 Note that the goals do not need in any way to be independent { this is only necessary if certain
eciency properties of the parallel execution are to hold. However, unlike in the source language,
in the kernel language care must be taken to lock properly concurrent accesses to shared variables
(see locking primitives).
3 This is implemented by having a private goal stack for each agent, from which other nodes
cannot pick work, and putting the goal being scheduled on the private goal stack of the appropriate
and remove (using in/1) data (i.e. Prolog terms) to and from the blackboard. If
the data is not present on the blackboard, the process suspends until it is available.
Alternatively, other primitives (in noblock/1 and rd noblock/1) do not suspend
if the data is not available { they fail instead and thus allow taking an alternative
action if the data is not in the blackboard. The input primitives can also wait on
disjunctions of terms.
4 Implementing Concurrent (Constraint) Languages in Distributed Environments
We now describe a concrete application of our ideas. Our objective in this example
is to combine the two main approaches currently used in concurrent logic programming, and which are seen traditionally as unrelated: \shared variable" systems, in
which communication among parallel tasks is done through logical variables (e.g.
Concurrent-Prolog [25], PARLOG [12], GHC [30], Janus [24], AKL [20], Oz [27],
etc.), and \distributed" or \blackboard" systems, in which communication is done
through explicit built-ins which access shared channels or global data areas (e.g.
Multi-Prolog [2], Shared Prolog [3], and Prologs incorporating Linda [7], being one
of the most popular Linda implementations the one bundled with SICStus [1]). In
order to do that, we will sketch a method for implementing communication through
shared variables by means of a blackboard. We assume the availability of the primitives introduced in the rst sections. We also assume that we want to implement a
simple concurrent (constraint) language which basically has a sequential operator,
a parallel operator (which, since we are in a distributed environment could actually
mean execution in another node of the net), and \ask" and \tell" unication primitives. The sort of system that we have in mind could perhaps be a local area net,
where the nodes are workstations. The incorporation of the sequential operator (to
mark goals that should not be \farmed out") and the special marking of \(remote)
communication variables" that will be mentioned later is relevant in the environment being considered. Note that it would be extremely inecient to blindly run
a traditional concurrent logic language (creating actual possibly remote tasks for
every parallel goal and allowing for all variables to be possibly shared and worked
on concurrently by goals in dierent nodes) in such a distributed environment. A
traditional concurrent language can of course be compiled to run eciently in such
an environment after granularity analysis [9, 32, 21] | in fact, this can be seen as
a source level transformation to a language of the type we are considering.
To implement this language on K&P we start by observing that the sequential
and parallel operators of the source language map directly into the sequential (\,")
and &@ (or &&@/2, if fairness is needed) operators of K&P. However, while this allows creating remote tasks, it does not by itself implement the communication of
values between nodes through shared variables. We propose to do this by placing
before the concurrent call a call to a predicate which will attach an attribute to
the shared variables marking them as \communication variables". Also, a unique
identier is given to each communication variable. All bindings to these variables
are posted on the blackboard (using the out/1 primitive) as (variable id,value) pairs,
where if values contain themselves new variables, such variables are represented by
their identiers. Thus, substitutions are represented as explicit mappings. When
bound to a communication variable, a non-communication variable is turned into
a communication variable. Tell and ask operations on ordinary variables, which
are handled in the standard way, are distinguished from tell and ask operations to
(remote) communication variables by the fact that the latter have the corresponding attribute attached to them. Thus, tell and ask unications to communication
variables will be handled by the attributed variable unication. A tell will be implemented by actually performing the binding to the variable in the manner explained
above using the out/1 blackboard primitive. An ask will wait until a binding for
the variable is posted on the blackboard. This will be commonly implemented using
the blocking rd/1 blackboard primitive, since in general a variable can have multiple readers and thus in/1 cannot be used. On the other hand, if a threadedness
analysis is performed and a variable is determined to have only one producer and
one consumer then in/1 can be used performing on the y garbage collection on the
blackboard.4 Else, when a remote goal nishes, a call to a tidying-up predicate can
be used to erase the entries in the blackboard corresponding to the bindings of variables which are not used as communication variables any more (and are not linked
to other active communication variables) and creating the corresponding term in
the heap of the process which continues with the execution.
A Concrete Implementation in SICStus Prolog
In order to be more concrete we sketch our implementation of the ideas outlined
above in a widely available environment: SICStus Prolog, enhanced with attributed
variables, and using the Linda library provided with recent versions of the system.
We hope that this detailed presentation of a concrete implementation will clarify
the issues that appear in practice when using the techniques proposed.
The implementation of the basic operators such as & and &@ in a Linda based
environment is not our current subject but is in any case relatively straightforward
(details of a particular implementation, also available by ftp, can be found in [13]):
a number of Prolog processes running in dierent network nodes are started as
Linda clients and thus share the blackboard, which is accessible through the normal
Linda primitives. Goals that are followed by & are simply posted on the blackboard.
Idle processes are waiting for work to be posted, which they then execute. Goals
that are followed by &@ x are posted on the blackboard with an identier that
indicates they are meant to be run on a given machine x. This allows, for example,
the following query to start a \producer" goal in a remote machine \alba" and a
consumer locally:
?- N=10, producer(N,L) &@ alba, consumer(L).
As mentioned before, in order to implement communication between nodes
through the variable L we would like to mark that variable as shared by attaching
an attribute to it. In general L may be bound to a complex term with intervening
variables, and then each such variable has to be marked in turn. On the other hand,
in the blackboard implementation we are considering, variables posted to the blackboard lose their identity. Thus, a unique identier needs to be given to each one.
Note that since attribute attachment operations are local to each process, identifying the shared variables and giving them identiers (which can be done once) is
best separated from the action of actually attaching attributes to them (which has
to be repeated in each node sharing the variable).
We implement a predicate var ids(LVars, Pairs) which given a set of lexical
variables appearing in the forked goal(s) returns the set of intervening run-time
variables, assigns a unique identier to each of them, and returns the information
in the form of (
) pairs. In our example, a call to this predicate is placed
before the call to producer(L) as follows:
V ariable; I d
?- var_ids([L],Ps), ( assign_ids(Ps), producer(N,L) ) &@ alba,
assign_ids(Ps), consumer(L).
4 This illustrates how the attributed variable approach also allows performing low-level optimizations as source to source transformations.
(this is handled automatically by a simple lexical expansion of the original query).
Only one pair would be generated in this case, since L is a free variable (of course,
this important case can be treated specially, but the general purpose primitive is
used for illustration purposes). Note that assign ids(Ps), dened by
assign_ids([(X,Id)|Ps]) :assign_id(X,Id),
assign_id(X,Id) :attach_attr(X,shv(Id)).
does the actual attachment of the attribute shv(Id) to each shared variable, and
that this is done both in the local and the remote machine (alba, in this case).
Once suitably marked, all the unications involving these communication variables
are handled through the blackboard. The appropriate handlers are given in the
following \blackboard unication" code (recall that verify attr/2 is called when
an attributed variable is unied with a term, and combine attr/2 is called when
two attributed variables are bound to each other):
verify_attr(shv(Id),Term) :trans_shterm(Term, NewTerm),
combine_attr(shv(I1), shv(I2)) :shv_unify('$shv'(I1),'$shv'(I2)).
The predicate trans shterm/2 transforms a term into its blackboard representation:
trans_shterm(X,'$shv'(Id)) :var(X), !,
( get_attr(X,shv(Id)) -> true ;
attach_attr(X,shv(Id)) ).
trans_shterm(Term,NewTerm) :functor(Term,F,N),
trans_shterm(0,_,_) :- !.
trans_shterm(N,Term,NewTerm) :N > 0,
N1 is N-1,
This predicate uses the primitive new shv id/1, which returns a new shared variable
identier dierent from any other in any process participating in the computation.
The predicate shv unify/2 performs the actual unication of terms that are
already in the blackboard (we have left out all explicit locking in the unication for
shv_unify(A, B) :dereference(A, VA),
dereference(B, VB),
dereference(X, V) :X = '$shv'(_),
sh_get_bind(X,Binding), !,
dereference(Binding, V).
shv_unify_values(X1,X2) :X1='$shv'(Id1),
( Id1=Id2 ; sh_bind(X1,X2) ), !.
shv_unify_values(X1,X2) :X1='$shv'(_), !,
shv_unify_values(X1,X2) :X2='$shv'(_), !,
shv_unify_values(X1,X2) :functor(X1,F,N),
shv_unify_args(N, X1, X2).
shv_unify_args(0, _, _) :- !.
shv_unify_args(N, X1, X2) :N > 0,
arg(N, X1, A1),
arg(N, X2, A2),
shv_unify(A1, A2),
N1 is N-1,
shv_unify_args(N1, X1, X2).
The unication routine uses the following primitive operation, which returns the
immediate binding of a shared variable or fails if it does not exist:
sh_get_bind(Id,T) :- linda:rd_noblock(shbinding(Id,T)).
The following operation is used when writing out a binding for a variable:
sh_bind(Id,T) :- linda:out(shbinding(Id,T)).
For example, given the query
?- var_ids([L],Ps), assign_ids(Ps), producer(3,L).
and the following denition of a simple producer, the contents of the blackboard
after execution are listed to its right:
producer(0,T):- !, T = [].
producer(N,T) :- N>0,
T = [N|Ns],
N1 is N-1,
shbinding($shv(3),[ ])
In order to support synchronization some blocking (\ask") primitive has to be
provided. For simplicity, we only describe the implementation of the K&P wait
primitive in this context, which is in any case sucient for most purposes:
wait(X) :get_attr(X,shv(Id)), !,
wait_shnonvar(X) :X = '$shv'(_), !,
sh_wait_bind(X, Binding),
The following primitive is used above:
sh_wait_bind(Id,T) :- linda:rd(shbinding(Id,T)).
A simple stream communication based consumer using these primitives can be constructed as follows:
consumer(T) :wait(T),
consumer_body([H|T]) :consumer(T).
Note that the above producer and consumer can also be seen as the result of a
straightforward compilation of the following fragment of GHC code:
producer(0,T) :- T = [].
producer(N,T) :N>0 | T = [N|Ns],
N1 is N-1,
consumer([H|T]) :| consumer(T).
Of course, an equivalent distributed producer-consumer situation (in which the
elements are consumed in the same order as they are produced, as in the program
above) can be easily implemented making direct use of the Linda primitives using
the query:
?- N=10, lproducer(N) &@ alba, lconsumer.
and the following program:
lproducer(N) :- lproducer(N,1).
lconsumer :- lconsumer(1).
lproducer(0,C) :- !,
lproducer(N,C) :N>0,
N1 is N-1,
C1 is C+1,
lconsumer(C) :linda:rd(message(C,T)),
lconsumer_data(number(N),C) :C1 is C+1,
However, arguably this program lacks the elegance of the shared variable communication based program: for example, if we want to run simultaneously several
instances of producers and consumers, the generation of new identiers for the messages must be explicitly encoded. The shared variable communication approach
sketched allows in some ways having the best of both worlds or, in any case, being
able to choose between them. It certainly provides the expected functionality. Its
performance of course depends heavily on the performance of the blackboard implementation. However, this is certainly also the case if the Linda primitives are
used directly.
5 Implementing Other Models Using Attributed
Lack of space does not allow elaborating further but we argue that using techniques
similar to those that we have proposed it is possible to implement many other parallel and concurrent models at the source level. For example, while and-parallelism
can be supported in or-parallel implementations by folding it into or-parallelism, no
communication among and-parallel tasks is possible. Our technique could be used
to provide this communication, for example by \escaping" shared variable unications and asserting them to the common database. We believe it is as well quite
possible to encode the determinacy driven synchronization and-parallelism of the
Andorra-I system [23] in terms of our wait primitive and the concurrency operators.
We also believe it is quite possible to implement languages with deep guards and/or
those based on the Extended Andorra Model [31], such as AKL [20].
For example, one of the most characteristic features of deep guard languages is
precisely the behavior of the guards, and one of the main complications in implementing such languages is in implementing the binding rules that operate within
such guards. If the Herbrand domain is used, the guard binding rules require in
principle that no bindings to external variables be made. Thus, it is necessary to
keep track of the level of nesting of guards and assign to each variable the guard
level at which it was created. Note that this can be done by assigning to each
guard a hierarchical identier and attaching to each variable such an identier as
(part of) its attribute. Unications in the program are labeled with the identier
of the guard in which they occur (the level is passed down recursively through an
additional argument). Such unications are handed over to the attributed variable
handler which makes computation suspend unless the variable and the binding have
the appropriate relative identiers. The binding rules for domains other than Herbrand can be more complex because they often use the concept of entailment. But
note that in the proposed approach all constraint solving would possibly be implemented through attributed variables anyway. Thus, it is not dicult to imagine
that a correct entailment check can be written at the source level using the same
primitives and wait. Some models are more involved: in AKL, for example, there is
a notion of local bindings and there is an additional rule controlled by the concept of
\stability" (closely related to that of independence) which allows non-deterministic
bindings to propagate at \promotion" time. We believe however that there is also
potential for the use of attributed variables for the implementation of AKL. For
example, promotion rules can also be implemented by updating the identiers (the
attributes) of all the local variables to higher levels.
Another model for parallel execution of Prolog is the DDAS (\Dynamic Dependent And-parallel Scheme") model of K. Shen [26]. In a very simplied form
the DDAS model is an extension to (goal level) independent and-parallel models
which allows ne grained synchronization of tasks, implementing a form of \dependent" and-parallelism. Parallelism in this model is controlled by means of \Extended Conditional Graph Expressions" (ECGE for short) which are of the form:
( conditions => goals ). As such, these expressions are identical to those used in
standard independent and-parallelism: if the conditions hold, then the goals can be
executed in parallel, else, they are to be executed sequentially. The main dierence
is that a new builtin is added, dep/1. This builtin can appear as part of the conditions of an ECGE. Its eect is to mark the variable(s) appearing in its argument
specially as \shared" or \dependent" variables. This character is in eect during
the execution of the goals in the ECGE and disappears after they succeed. Only the
leftmost active (i.e. non nished) goal in the ECGE (the \producer") is allowed to
bind such variables. Other goals which try to bind such variables (the \consumers")
must suspend until the variable is bound or they become leftmost (i.e. all the goals
to their left have nished). In order to support this model in K&P we assume a
source to source transformation (using term expansion/2) of ECGEs: an ECGE is
turned into a Prolog if-then-else such that if the conditions succeed then execution
proceeds in parallel (using the &/2 operator, which directly encodes the fork-join
parallelism implemented by the ECGEs), else it proceeds sequentially. Dependent
variables shared by the goals in a ECGE are renamed. The dep/1 annotation is
transformed into a call to a predicate that marks the variables as dependent by
attaching attributes to them. Such attributes also encode whether a variable is
in a producer or a consumer position. Unication is handled in such a way that
bindings to variables whose attribute corresponds to being in the producer position
are actually bound. Note that if the variable is being bound to a complex term
with variables, these variables also have to be marked as dependent. Bindings to
variables whose attribute corresponds to being in a consumer position suspend the
execution of the associated process (using wait/1). The change from producer to
consumer status is implemented as follows: each parallel goal containing a dependent variable (except the last one) is replaced by the sequential conjunction of the
goal itself and a call to a predicate which will \pass the token" of being leftmost
to the next goal (or short-circuiting the token link if it is an intermediate goal).
This predicate also takes care of restoring the connection lost due to the variable
6 Performance
The objective of the technique presented is achieving a certain functionality through
the use of attributed variables, rather than any increase in performance. However,
it is still interesting to make some observations regarding the resulting implementations. Table 1 presents some results obtained with the concurrent extension to
SICStus Prolog described in Section 4. A set of programs involving producers and
consumers was run with each process running in a dierent workstation (Sun IPC)
connected over an Ethernet network. The performance of our implementation of
concurrent logic programming is compared with equivalent programs written di-
shared var.
shared var./linda
Incomplete message protocol (30 messages)
17 5
3 +1
2 +1
Bounded buer protocol (30 messages, 3 places)
21 0
6 ?5
2 +4
One to many communication with acknowledge (20 messages, two readers)
25 1
5 45
10 + 1
2 +1
Table 1: Distributed Shared Variable Communication
qs iap
qs conc 1
qs conc 2
1 Proc. 2 Proc. 3 Proc. 5 Proc.
90 (1) 50 (1.8) 50 (1.8) 50 (1.8)
320 (1) 170 (1.8) 120 (2.6) 80 (4.0)
400 (1) 210 (1.9) 150 (2.6) 100 (4.0)
120 (1) 80 (1.5) 60 (2.0) 30 (4.0)
7 Proc.
50 (1.8)
70 (4.5)
90 (4.4)
30 (4.0)
9 Proc.
50 (1.8)
60 (5.3)
80 (5.0)
30 (4.0)
Table 2: &-Prolog Performance for Concurrent Benchmarks
rectly by hand in Linda in the most ecient way possible. \Time" gives the execution time in seconds. \Space" is an expression which gives the number of blackboard
items generated by the programs with an input of size (experimental results conrm this expression). \Operations" is the number of operations performed in the
blackboard during the execution of the programs. This number was measured by
instrumenting the SICStus implementation of the Linda library. The incomplete
message program is the standard program implementing a two-way communication
between processes [12, 28, 29]. The bounded buer program is also the standard
one. The one to many communication with acknowledge program allows several
processes to read a stream produced by another, the latter being informed of which
process read each element. We argue that, despite the amount of metainterpretation
of terms happening when using shared variables for communication the resulting
performance is still reasonable, specially if we take into account that the current implementation using attributes is very nave, and many optimizations can be made to
improve performance, both in the low{level implementation and in the compilation
In order to also see if actual performance improvements from parallelism are attainable with the technique, we have performed some measurements on a prototype
implementation of communication through shared variables in the &-Prolog system
using similar techniques. It should be noted that the results are based on rather
inecient implementations of the wait/1, lock/2 and unlock/1 primitives.
Table 2 shows the results. Times are in ms. and speedups between parentheses.
qs iap is the independent and{parallel version of quicksort, where the two recursive
calls are executed in parallel, provided for reference. A small list is used in order
to somewhat limit the parallelism available using (goal level) independent andparallelism. The benchmark uses append rather than dierence lists. qs conc 1
is again quicksort, in which the list splitting is performed concurrently with the
recursive calls, acting as a producer. In qs conc 2 the concurrency is extended also
to the append call, and all the goals in the recursive clause are run concurrently.
Finally, nrev is the standard nave reverse benchmark. The results are interesting in
that they show that even with a nave implementation of the concurrency primitives
reasonable speedups can be achieved with our techniques in programs with small
granularity (for example, nrev), even if not (yet) speed, when compared to sequential
Prolog. Again, our objective herein is simply to point out and substantiate to
some extent the great potential that in our view the concept of attributed variables
has in the implementation of generic parallel, concurrent, and distributed logic
programming systems.
7 Conclusions
We have proposed a dierent and novel use for the concept of attributed variables:
developing a generic parallel/concurrent (constraint) logic programming system,
using the same \glass box" avor as that provided by attributed variables and
meta-terms in the context of constraint logic programming implementations. We
argue that a system which implements attributed variables and the few additional
primitives which have been proposed constitutes a kernel language which can be
easily customized at source level to implement many of the languages and execution
models of parallelism and concurrency currently proposed, in both shared memory
and distributed systems. We have illustrated this through a few examples.
While the wide applicability of the ideas presented is very attractive, a clear issue
is the performance of the systems built using them. Of course, such performance
is bound to be slower than that of the corresponding native implementations. It
is clear that the native implementation approach is both sensible and practical,
and simply the way to go in most cases. On the other hand we also feel there it
is interesting to be able to have a generic system which can be easily customized
to emulate many implementations. On one hand, it can be used to study in a
painless way dierent variations of a scheme or to make quick assessments of new
models. On the other hand the loss in performance is compensated in some ways
by the exibility (a tradeo that has been found acceptable in the implementation
of constraint logic programming systems), and such performance can be improved
in a gradual way by pushing the implementation of critical operations down to C.
[1] J. Almgren, S. Andersson, L. Flood, C. Frisk, H. Nilsson, and J. Sundberg.
Sicstus Prolog Library Manual. Po Box 1263, S-16313 Spanga, Sweden, October
[2] K. De Bosschere. Multi{Prolog, Another Approach for Parallelizing Prolog.
In Proceedings of Parallel Computing, pages 443{448. Elsevier, North Holland,
[3] A. Brogi and P. Ciancarini. The Concurrent Language, Shared Prolog. ACM
Transactions on Programming Languages and Systems, 13(1):99{123, 1991.
[4] F. Bueno, M. Garca de la Banda, and M. Hermenegildo. Eectiveness of Global
Analysis in Strict Independence-Based Automatic Program Parallelization. In
International Symposium on Logic Programming, pages 320{336. MIT Press,
November 1994.
[5] M. Carlsson. Freeze, Indexing, and Other Implementation Issues in the Wam.
In Fourth International Conference on Logic Programming, pages 40{58. University of Melbourne, MIT Press, May 1987.
[6] M. Carlsson. Sicstus Prolog User's Manual. Po Box 1263, S-16313 Spanga,
Sweden, February 1988.
[7] N. Carreiro and D. Gelernter. Linda in Context. Communications ACM, 32(4),
[8] T. Chikayama, T. Fujise, and D. Sekita. A Portable and Ecient Implementation of KL1. In Springer Verlag, editor, Proc. 6th. Int'l Symposium on
Programming Language Implementation and Logic Programming, September
[9] S. K. Debray, N.-W. Lin, and M. Hermenegildo. Task Granularity Analysis in
Logic Programs. In Proc. of the 1990 ACM Conf. on Programming Language
Design and Implementation, pages 174{188. ACM Press, June 1990.
[10] A. Colmerauer et Al. Prolog II: Reference Manual and Theoretical Model.
Groupe D'intelligence Articielle, Faculte Des Sciences De Luminy, Marseilles,
[11] European Computer Research Center. Eclipse User's Guide, 1993.
[12] S. Gregory. Parallel Logic Programming in PARLOG: the Language and its
Implementation. Addison-Wesley Ltd., Wokingham, England, 1987.
[13] M. Hermenegildo. A Simple, Distributed Version of the &-Prolog System. CLIP
TR, School of Computer Science, Technical University of Madrid (UPM), 1994.
Available from
[14] M. Hermenegildo, D. Cabeza, and M. Carro. On the Uses of Attributed Variables in Parallel and Concurrent Logic Programming Systems. CLIP TR,
School of Computer Science, Technical University of Madrid (UPM), 1994.
Available from
[15] M. Hermenegildo and K. Greene. The &-prolog System: Exploiting Independent And-Parallelism. New Generation Computing, 9(3,4):233{257, 1991.
[16] C. Holzbaur. Metastructures vs. Attributed Variables in the Context of Extensible Unication. In 1992 International Symposium on Programming Language
Implementation and Logic Programming, pages 260{268. LNCS631, Springer
Verlag, August 1992.
[17] C. Holzbaur. SICStus 2.1/DMCAI Clp 2.1.1 User's Manual. University of
Vienna, 1994.
[18] Serge Le Houitouze. A New Data Structure for Implementing Extensions to
Prolog. In P. Deransart and J. Maluszynski, editors, Proceedings of Programming Language Implementation and Logic Programming, number 456 in Lecture Notes in Computer Science, pages 136{150. Springer, August 1990.
[19] J. Jaar and J.-L. Lassez. Constraint Logic Programming. In ACM Symp.
Principles of Programming Languages, pages 111{119. ACM, 1987.
[20] S. Janson and S. Haridi. Programming Paradigms of the Andorra Kernel
Language. In 1991 International Logic Programming Symposium, pages 167{
183. MIT Press, 1991.
[21] P. Lopez and M. Hermenegildo. Ecient Term Size Computation for Granularity Control. In Proc. of the Twelfth International Conference on Logic
Programming. The MIT Press, June 1995.
[22] U. Neumerkel. Extensible Unication by Metastructures. In Proceeding of the
META'90 workshop, 1990.
[23] V. Santos-Costa, D.H.D. Warren, and R. Yang. Andorra-I: A Parallel Prolog
System that Transparently Exploits both And- and Or-parallelism. In Proceedings of the 3rd. ACM SIGPLAN Symposium on Principles and Practice of
Parallel Programming. ACM, April 1990.
[24] V. Saraswat. Concurrent Constraint Programming Languages. PhD thesis,
Carnegie Mellon, Pittsburgh, 1989. School of Computer Science.
[25] E. Y. Shapiro. A Subset of Concurrent Prolog and Its Interpreter. Technical
Report TR-003, ICOT, 1-4-28 Mita, Minato-ku Tokyo 108, Japan, January
[26] K. Shen. Exploiting Dependent And-Parallelism in Prolog: The Dynamic,
Dependent And-Parallel Scheme. In Proc. Joint Int'l. Conf. and Symp. on
Logic Prog. MIT Press, 1992.
[27] G. Smolka. The Denition of Kernel Oz. DFKI Oz documentation series,
German Research Center for Articial Intelligence (DFKI), November 1994.
[28] S. Taylor. Parallel Logic Programming Techniques. Prentice-Hall, Englewood
Clis, NY, 1989.
[29] E. Tick. Parallel Logic Programming. MIT Press, Boston, Mass., 1991.
[30] K. Ueda. Guarded Horn Clauses. In E.Y. Shapiro, editor, Concurrent Prolog:
Collected Papers, pages 140{156. MIT Press, Cambridge MA, 1987.
[31] D.H.D. Warren. The Extended Andorra Model with Implicit Control. In
Sverker Jansson, editor, Parallel Logic Programming Workshop, Box 1263, S163 13 Spanga, SWEDEN, June 1990. SICS.
[32] X. Zhong, E. Tick, S. Duvvuru, L. Hansen, A.V.S. Sastry, and R. Sundararajan. Towards an Ecient Compile-Time Granularity Analysis Algorithm. In
Proc. of the FGCS'1992 International Conference. Institute for New Generation Computer Technology (ICOT), June 1992.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF