Université de Moncton, CANADA E1A-3E9
Logic and Functional Programming Group,
Department of Computing Sciences, Simon Fraser University,
KOEN DE BOSSCHERE,
Vakgroep Elektronica en Informatiesystemen,
Universiteit Gent, E-mail: email@example.com.
We describe a set of programming patterns used for implementing a scalable infrastructure which supports remote execution mechanisms, mobile code and agents in a distributed logic programming framework. The particular focus of this paper is on the use of BinProlog's strong metaprogramming abilities. Some advanced logic programming constructs, as intuitionistic implication and high-order call/N, cooperate with encapsulated socket-level constructs for maximum configurability and efficiency. We show that strong metaprogramming is not a security threat if used through a set of filtering interactors which allow source level implementation of arbitrary security policies. Mobile code is implemented in a scalable way through a set of distributed client+server pairs interconnected through a master server acting only as an address exchange broker for peer-to-peer interactors. We have thoroughly tested our programming patterns and design principles through a realistic implementation in a widely used, freely available Prolog system (http://clement.info.umoncton.ca/BinProlog) as well as with its Java peers built on top of our unification enhanced Java based Linda implementation (http://clement.info.umoncton.ca/LindaInteractor). Keywords: mobile code, remote execution, metaprogramming, agents, Linda coordination, blackboard-based logic programming, Prolog, distributed programming, virtual worlds, scalable and secure Internet programming, Prolog/Java embedding
Although the Internet has been designed to survive nuclear war and its underlying packet switching technology is intrinsically peer-to-peer, fault-tolerant and scalable, successive higher level networking and programming layers have given up (often too easily) these abilities. Among the most annoying and at the same time the most pragmatically well thought decisions dominating the world of the Internet, at various programming and application development layers:
In this paper, we will show how some standard and some non-standard Logic Programming language tools can be used to elegantly recover the full potential of the Internet for building scalable, peer-to-peer, programmable multi-user communities. The virtual layer we will build is based on a small set of very-high level programming constructs making essential use of meta-programming, which is seen here as the ability to view information either as code or as data, and efficiently switch between these views, on demand. In particular, we will show how powerful remote execution mechanisms, agents and mobile code can be expressed in this framework.
With important resources at their disposal, operating system/computer/network companies have found their own (often complex and costly) ways to get back some of the original power, hidden under various layers of software/hardware. Despite the existence of practical solutions to some of these problems, no unifying approach exists that can help clarify the logical structure of future peer-to-peer, scalable virtual communities, where the distinction between human and non-human agents will become less and less obvious. This article proposes one such approach. We are also convinced that the power of the knowledge/processing component of logic programming will help efforts towards making intelligent network/mobile agent programming more effective.
We will describe our constructs as built on the top of the popular Linda coordination framework, enhanced with unification improved pattern matching, although our actual implementation is based on a more general coordination logic. This allows negotiation of tuple removal as well as of suspension between user intents at definition time, when an object is stored on the (shared and synchronized) blackboard, as well as at reference time, when it is retrieved associatively.
Linda  based frameworks like Multi-BinProlog  offer a wide variety of blocking and non-blocking as well as non-deterministic blackboard operations (backtracking through alternative answers). For reasons of embeddability in multi-paradigm environments and semantic simplicity we have decided to drop non-determinism and return to a subset close to the original Linda operators (combined with unification), and simple client-server component melted together into a scalable peer-to-peer layer forming a `web of interconnected worlds'.
in(X) waits until it can take an object matching X from the server
all(X,Xs) reads the list Xs matching X currently on the server
run(Goal) starts a thread executing Goal
halt stops current thread
out(X) puts X on the server
in(X) waits until it can take an object matching X from the server
all(X,Xs) reads the list Xs matching X currently on the server
run(Goal) starts a thread executing Goal
halt stops current thread
The presence of the all/2 collector compensates for the lack of non-deterministic operations. Note that the only blocking operation is in/1. Notice that blocking rd/1 is easily emulated in terms of in/1 and out/1, while non-blocking rd/1 is emulated with all/2.
1: Basic Linda operations
A servant is one of the simplest possible agents:
The servant is started as a background thread with run(servant(Name)). No `busy wait' is involved, as the servant's thread blocks until in(todo(Task)) succeeds.
In an `Internet chat' application, for instance,
whisper/2, defined as
unblocks the matching in(todo(mes_to(To,Mes,From))) of a notifier servant which then outputs a message by executing mes_to(To,Mes,From).
More generally, distributed event processing is implemented by creating a `watching' agent for a given pattern.
Note the fundamental link between `event-processing' and the more general Linda protocol: basically an out/1 operation can be seen as generating an event and adding it to the event queue, while an in/1 operation can be seen as servicing an event. While usually event-loops switch on numeric or pseudo-numeric event constants in a rather rigid and non-compositional way, Linda-based event dispatching is extensible by adding new patterns. When (naturally) restricted to ground patterns, efficient indexing can be used with performance hits unnoticeable in a network lag context. Linda's out/blocking in combination can be seen as automating the complex if-then-else logic of (distributed) hierarchical message dispatching loops through unification.
Remote processing as well as simple security mechanisms are expressed easily, by creating a `command server' thread close to the following code:
% remote processing request
% a remote command processor
whoami(Me), % gets unique user identification information
( friend_of(Me,ForWho)->call(What),fail % trusted friends
Clearly, command_server and notifier threads can be seen as `behavior components' of a unique agent. Moreover, they actually might cooperate in a synchronized way if each please/2 command would trigger a whisper/2 action to be serviced by a notifier later. In this example security is based on allowing arbitrary execution for trusted (i.e. known to friend_of/2) clients of existing server-side code.
Note that by restricting command_server to execute only sequences of acceptable known commands, on top of code-as-data sent through out/1 transactions, arbitrary security mechanisms can be put in place at source level. We will describe this in more detail later.
The MOO inspired `Web of Worlds' metaphor [22, 24] as a set of Linda blackboards storing state information on servers connected over the the Internet allows a simple and secure remote execution mechanism through specialized server-side interpreters.
Our implementation is integrated in the freely available BinProlog compiler, starting with version 5.60. To ensure interoperability between Windows 95/NT PCs and Unix machines we have conservatively implemented a generic socket package with operations specialized towards support for Linda operations and remote execution. A master server on a `well-known' host/port is used to exchange identification information among peers composed of a client and a `multiplexing' server (Fig. 1)
1: A Web of Worlds
The code for our `generic' server, with various components which are overridden through the use of intuitionistic implication to obtain customized special purpose servers at user level, follows. Higher order call/N , combined with intuitionistic assumptions, are used to pass arbitrary interactors to this generic server code.
( call(Interactor,Server)->true % higher-order call to interactor
; \+ errmes(failed_interaction(Server),Interactor)
Note the use of a specialized server-side interpreter server_loop, configurable through higher-order `question/answer' closures we have called interactors.
We have found out that the use of intuitionistic implications (pioneered by Miller's work [15, 14, 9]) helps to overcome (to some extent) Prolog's lack of object oriented programming facilities, by allowing us to `inject' the right interactor into the generic (and therefore reusable) interpreter. BinProlog's =>> temporarily assumes a clause in `asserta' order, i.e. at the beginning of the predicate. The assumption is scoped to be only usable to prove its right side goal and vanishes on backtracking. We refer to [19, 21, 7] for more information on assumptions and their applications.
The following example shows how the same client-side code is used for both for point-to-point and `broadcast' communication, depending on a to_all(Pattern) assumption allowing to select matching targets.
Slightly more complex code (see file extra.pl in the BinProlog distribution) also handles proxy forwarding services transparently.
Implementation of arbitrary remote execution is easy in a Linda + Prolog system due to Prolog's metaprogramming abilities. No complex serialization or remote procedure/method call packages are needed. In BinProlog (starting with version 5.60) code fetched lazily over the network is cached in a local database and then dynamically recompiled on the fly if usage statistics indicate that it is not volatile and it is heavily used locally.
allow remote predicate calls without/with returned answer on the calling site.
allows us to use BinProlog to remotely execute shell commands on a remote machine and collect/print the answers at the calling site.
As an example, high performance file-transfer (actually known to be faster than well-known http as ftp protocols) is implemented by orchestrating Prolog based remote_run control operations, which are used to trigger direct full speed transfers entirely left to optimized C built-ins like sock2file/2:
Once the basic Linda protocol is in place, and Prolog terms are sent through sockets, a command filtering server loop simply listens and executes the set of `allowed' commands.
Despite Prolog's lack of object oriented features, we have implemented code reuse with intuitionistic assumptions [21, 7].
For instance, to iterate over the set of servers forming the receiving end of our `Web of Worlds', after retrieving the list from a `master server' which constantly monitors them making sure that the list reflects login/logout information, we simply override host/1 and port/1 with intuitionistic implication:
To specialize a generic server into either a master server or a secure `chat-only' server which merges messages from BinProlog users world-wide, we simply override the filtering step in the generic server's main interpreter loop.
On an Intranet of trusted users and computers, or in different windows of an unconnected PC or workstation, a user might want to experiment with a server allowing arbitrary Prolog command execution, i.e. full remote metaprogramming. One can start a local remote predicate call server with:
BinProlog's convention is that if the name returned by default_host/1 is different from localhost it assumes that the user wants to get connected and interoperate with other BinProlog users.
To override this default setting, one can use either
or temporarily override the default host by assuming it as a host/1 fact with BinProlog's intuitionistic assumption. Such assumptions are scoped and forgotten on backtracking. With
a given server is registered on a master server given by (also overridable):
An unrestricted server can run benign commands like
as well as more malicious ones like:
to collect password information ready for one's favorite cracking algorithm.
Note that registered servers are in principle accessible to other users and therefore not secure.
Next, we will give two simple approaches to implement security.
Here is the code of the chat server:
By overriding the default server_interactor with our chat_server_interactor we tell to chat_server that only commands matching known, secure operations (i.e. mes/2 and ping/1) have to be performed on behalf of remote users.
This might look very simple but some of BinProlog's key features (intuitionistic implication and higher-order call/N) are used inside the `generic' server code (see file extra.pl in the BinProlog distribution) to achieve this form of configurability.
Security is achieved by having specialized meta-interpreters `filtering' requests on the server side. Intuitionistic assumption is used to override the generic server's interactor, i.e. the inner server operation filtering allowed requests, while allowing reuse of most of the generic server's code in an object oriented programming style.
Here is some interactor code for two other `secure' servers.
The first one (BinProlog's current default) is a Linda Term server (to be started with run_server/0) enhanced with moderately weak remote management commands. It offers a good combination of security and convenience: it only allows modifying its dynamic database remotely through Linda operations and displaying messages. It can be shut down remotely and checked if alive and of course it can be subject to resource attacks by malicious users. However it can do no harm to one's file system and cannot be used to collect information about one's computer.
Note that this interactor even supports proxy forwarding to a given host/port. Clearly, the secure execution of forwarded queries is the responsibility of the proxy target.
Keeping one's host/port information secret from other users can be achieved by starting a master server local to a secure physical or virtual Intranet.
For users behind a firewall, this might actually be the only way to try out these operations as the default master server might be unreachable.
A local master server on, e.g. port 7788 is started with something like:
Client and server programs intended to be managed by a local master server will have simply to assume a fact:
or set the master server with:
To keep the workload of the master server minimal, only when an error is detected by a client, the master server is asked to refresh its information and possibly remove dead servers from its database.
BinProlog starting from version 5.40 communicates with our recently released Java based Linda Interactors special purpose trimmed down pure Prolog engines written in Java which support the same unification based Linda protocol as BinProlog. The natural extension was to allow Java applets to participate to the rest of our `peer-to-peer' network of BinProlog interactors. As creating a server component within a Java applet is impossible due to Java's (ultra)-conservative security policies we have simply written a receiving-end servant close to our example in subsection 2, which relies on a proxy server on the site where the applet originates from, for seamless integration in our world of peer-to-peer interactors.
Here is the code for a more realistic servant, multiplexed among multiple servers and usable inside a Java applet.
out(servant_id(ID)), % registers this servant
in(query(ServerId,Query)), % waits for a query
% sends back an answer, if needed
( ServerId=->true % no reply sent to anonymous servers
functor(Query,stop,_), % stops when required
Note the presence of an overridable client-side interactor, allowing the generic servant code to be easily reused/specialized. Multiplexing is achieved by having each server's in/1 and out/1 data marked with unique server_id/1 records. The server side code for a single query is simply:
% identifies the server the query comes from
out(query(ID,Query)), % asks servant
in(answer(ID,R)), % gets back answer
To integrate multiple Java applet based clients in our `Web of Worlds', we use a more complex forwarding server, also available as equivalent Java code, to be run as a BinProlog and/or Java daemon on the same machine as the HTTP server the applet comes from.
Note that the forwarding server has the ability to interact with multiple servants, in particular with multiple Java applets.
We will shortly discuss here the basic Mobile Code facilities we have implemented. Future work will focus on building intelligent agent applications on top of them.
The following operations
allow fetching remote files/operator definitions rload/1, fetch_remote_operators/0 or on-demand fetching of a predicate at a time from a remote host during execution of TopGoal. This is basically the same mechanism as the one implemented for Java applet code fetching, except that we have also implemented a caching mechanism, both at predicate level (predicates are cached as dynamic code on the server to efficiently serve multiple clients) and at file level, on the client side.
Dynamic recompilation is used on the client side to speed-up heavily used relatively non-volatile predicates. With dynamically recompiled consulted code, listing of sources and dynamic modification to any predicate is available while average performance stays close to statically compiled code (usually within a factor of 2-3). Although when code comes over the network, code fetching time becomes more significant, the combination of remote code fetching and dynamic recompilation is a powerful accelerator for distributed network applications comparable with Java's just-in-time (JIT) compilation technology.
Implementing agents `roaming over' a set of servers is a simple and efficient high-level operation. First, we get the list of servers from the master server. Then we iterate through the appropriate remote-call negotiation with each site. Our agent's behavior is constrained through security and resource limitations of participating interpreters, having their own command filtering policies.
In particular, on a chat-only server, our roaming agent can only display a message. If the local interpreter allows gathering user information then our agent can collect it. If the local interpreter allows question/answering our agent will actually interact with the human user through the local server window.
Note that `mobile agents' do not have to be implemented as code physically moved from one site to another. In this sense we can talk about virtual mobile agents which are actually sets of synchronized remote predicate calls originating from a unique site, where most of the code is based/executed, while code actually moved can be kept to strict minimum, i.e. only a few remotely asserted clauses.
Our mobile agents are seen as `connection brokers' between participating independent server/client sites. For instance if two sites are willing to have a private conversation or code exchange they can do so by simply using the server/port/password information our agent can help them to exchange.
Note that full metaprogramming combined with source-level mobile-code have the potential of better security than byte-code level verification as it happens in Java, as meaningful analysis and verification of program properties is possible.
To give a flavor of how agent scripts are crafted in our framework we will describe a complete example emphasizing the idea of Virtual Mobile Agents described in section 6.3.
Each file begins with predicate main/0 which is BinProlog's user-defined auto-start point.
The first file (script.pro) mostly dispatches Unix processes creating the other tasks in separate cmdtool windows (for demo purposes) of appropriate sizes.
Each file contains the code describing the behavior of an agent. The file master.pl initiates a local master server which keeps an updated list of all the existing servers and their properties (unique id, host, port). A secret pasword is used when asking the master server to trigger remote action which, for instance, will shut down all the servers.
The file agent.pl contains the code shared among different agents. Main/0 starts by registering an unrestricted server, allowing execution of remote commands to the local master server. It is programmed in object oriented style with default actions to be performed by more specialized agents.
make_cmd(["I am ",I," !"],Cmd),
write('mobile code not welcome, intruder attempted to attach:'),nl,
The predicates for accepting/rejecting mobile `visitors' can be used to enforce such policies on a server.
The first agent has default initialization init/0 and its run/1 action step set to reject visiting mobile agents.
Note the use of the builtin login/1 predicate giving a unique identity to each agent.
The second agent is similar, except that its run/1 action step is set to accept mobile agents.
The mobile agent itself uses BinProlog's built-in iteration over the servers registered on the master server as a powerful remote execution mechanism. In the presence of the to_all/1 assumption, BinProlog's client code reacts by iterating over the set of servers matching a given pattern (everything in this case). This applies to every command, and in particular to remote execution requests specified by do/1.
The predicate animate/1 iterates over the set of servers, sending a small chunk of mobile code to be assumed on each server and then executed. Note that more sophisticated code migration schemes can be set up quite easily by assuming code(SourceFile), which has the effect to lazily fetch to a given server the code needed for the execution of a remote query.
Note the use of the built-in predicate yell/1, which broadcasts a message over the set of registered servers. It is actually defined as
On top of this, BinProlog 5.75 implements a built-in Internet/Intranet chat mechanism chat/0, allowing users sharing a master server to exchange messages. For one-server communication, the talk/0 built-in allows Java applets on the same machine to exchange messages among them and with BinProlog users.
A very large number of research projects have recently started on mobile agent programming. Among the pioneers, Kahn and Cerf's Knowbots  Among the most promising recent developments Luca Cardelli's Oblique project at Digital and mobile agent applications  and IBM Japan's aglets . We share their emphasis on going beyond code mobility as present in Java, for instance, towards control mobility. We think that distributed containers with ability to negotiate with agents the resulting local and global behavior can offer a secure and flexible approach to Internet aware distributed programming. A growing number of sophisticated Web-based applications and tools are on the way to be implemented in LP/CLP languages. Among them, work with a similar emphasis on remote execution, agents, virtual worlds can be found in [3, 13, 4, 2, 18, 12, 24].
Our remote execution mechanisms are based on a set of filtering interpreters which can be customized to support arbitrary negotiations with remote agents and are plugged in generic servers. The practical implementation is built on proven client/server technology, on top of a generic socket package, while giving the illusion of a `Web of MOOs' with roaming mobile agents at the next level of abstraction.
A Java based Linda implementation, using a minimal set of logic programming components (unification, associative search) has been recently released (the Java TermServer, available at http://clement.info.umoncton.ca/BinProlog). It allows to communicate bidirectionally with BinProlog, allowing creation of combined Java/Prolog mobile-agent programs. In particular, Java applets can be used as front end in browsers instead of the more resource consuming CGIs most of our applications are currently based on. It holds promise for smooth cooperation with existing Java class hierarchies as well as various BinProlog based components with intelligence and flexible metaprogramming on the logic programming side combined with visualization and WWW programming abilities on the Java side.
Future work will focus on intelligent mobile agents integrating knowledge and controlled natural language processing abilities, following our previous work described in [6, 17, 22].
We thank for support from NSERC (grants OGP0107411 and 611024), and from the FESR of the Université de Moncton. Koen De Bosschere is research associate with the Fund for Scientific Research - Flanders.