Distributed Digital Rule Processor for Single System Image on a Clustered Network and Method

1. The Field of the Invention

This invention relates generally to distributed parallel processing on a computer network, and more particularly to a method and apparatus for parallel processing digital rules in rule nets and slave translators which are interconnected by a global controller and bindery.

2. Background.

Many organizations are now dependent on computers for the management of their data. Because of the mission critical nature of computer applications like web servers, accounting, point of sale, inventory and customer information, even small organizations need their computer systems to be running virtually all the time. Computer users desire their systems to be fast and reliable, but current von Neumann processors are coming close to the limit of their capacity. In addition, von Neuman processors are sometimes unstable and can crash. The increased processing requirements for virtual reality graphics and client server network applications also justify the need for faster, affordable and more stable computers.

One way to speed up a computer is to increase a von Neumann single processor's speed. This has been occurring at a steady rate in the computer industry through miniaturization and circuit optimization. Another method of increasing computer performance is by adding more von Neumann processors to a single machine, also known as parallel processing. Unfortunately, conventional parallel processing machines have throughput bottlenecks and software which is difficult to debug. It is also known that von Neumann processors on separate machines or nodes can be connected in configurations such as client-server or neural networks. Some neural net models have even been designed which can process digital information and return a result.

Although adding more processors may help increase the speed of a computer it does not necessarily increase the reliability of the computer or its operating system. The popular operating systems on microcomputers such as Macintosh, Windows, Windows NT, Unix, and NetWare unpredictably fail from time to time. Overcoming single computer failures using multiple computer systems as a backup for systems that fail is known as redundancy.

One solution to single server failure is the concept of clustering. A cluster is a group of three or more servers (nodes) connected on a network which have high availability because they work together as one logical entity. When an independent server within the cluster fails, the workload of that server, and the services or applications it was providing, are distributed to the remaining servers in the cluster. Redundancy in a cluster provides high availability of data, network services and applications to users. The transfer of an application from a failed server to a remaining server is called failover. In clustering, the performance of the servers is also balanced as the servers allocate the network load to match the hardware capabilities of each server in the cluster. In order for the data on hard disk drives to be visible to the various nodes in the cluster, a shared disk sub-system is connected to all nodes in the cluster. If data were stored on a local hard drive of one of the servers, that data would be unavailable when that server crashed. Distributed file systems allow for all running servers to access the same data without corruption.

Outside users want to see the cluster as a single logical entity. This single system image (SSI) is achieved through certain characteristics of the cluster configuration. First IP (Internet Protocol) addresses which clients connect to are allowed to move from one server to another as part of failover. The virtual IP addresses, along with the software applications which use the IP address are moved from node to node when necessary.

Group membership software detects which nodes in a cluster are running, and cluster resource management software decides where the cluster resources reside (e.g. IP addresses, running application, disk subsystems). The decision as to which node gets a resource can be based on a cluster node preference list or some load balancing policy. The cluster resource management software does failover, failback and resource migration to adjust the load of the software on each node or server.

In a clustering system, the network client must have reconnection smarts so that the user cannot tell that behind the scenes a current connection to a server failed, and a new connection to the same IP address on another server has occurred. A major drawback with reconnection is the loss of gray data. An example of gray data loss is a transaction to a database on the server, which was not completed nor stored on the shared disk subsystem, thus the incomplete transaction will have to be started over. If the gray data remains in memory, then the transaction can automatically pickup where it stopped without restarting. If the gray data was erased from memory the transaction will have to be re-entered by the user when the application restarts on a remaining clustered server. Gray data recovery is a complex process and requires complex algorithms on the client side, in the database or on the servers. It is difficult for an application which runs on only one server at a time to prevent gray data loss because inside the cluster, the servers (nodes) are individually known to the group membership software and the cluster resource management software but they are not known to an individual application.

Current applications on a cluster are limited to running on one server. Of course an application will failover to another server if the server it is running on crashes, but it cannot automatically use the resources of all the servers in the cluster. Clusters provide a Single System Image (SSI) to users outside a group of network nodes, but a SSI from a software developer's perspective inside the cluster does not exist. So currently, programmers who desire to create applications which use the resources of more than one server in the cluster must organize their application to directly address the separate servers. This is a very complex and daunting task because of the necessary network knowledge and the high complexity of the clustering resource management software. What is needed is a parallel, distributed application execution environment on a cluster of von Neumann processors which appears as a single system image (SSI) to a software developer.

It is an object of the present invention to provide a distributed digital rule processor and method which create a true SSI for distributed software application execution, inside or outside a networked von Neumann processor cluster by broadcasting digital rules which are processed by rule nets and slave translators.

It is an object of the present invention to provide a distributed digital rule processor and method to allow the sharing of gray data between servers and the hiding of servers or von Neumann processors as slaves to a layer of digital rules.

It is another object of the invention to provide a distributed digital rule processor and method to create, store, and execute complex digital rules over a clustered network of von Neumann processors.

It is yet another object of the invention to provide such a distributed digital rule processor and method for executing digital rules and utilizing the high reliability benefits of a networked cluster of servers.

The distributed digital rule processor creates a single system image (SSI) and executes digital rules on a clustered von Neumann processor network. The processor has a plurality of rule nets each having an ordered list of rules and each rule has input variables and output variables. The rule nets can broadcast rules to other rule nets or slave translators. A plurality of slave translators executes the rules received from the rule nets and returns the results and data from the executed rules to the calling rule net. A global controller is coupled to the rule nets and slave translators which has a global bindery, a global data memory, and a current broadcast state. A global rule distribution queue is coupled to the global controller to store pending digital rules and broadcast rules to the rule nets and slave translators as signaled by the global controller.

In one embodiment of the invention, a computer programmer compiles source code through a compiler into digital rules. These digital rules are stored in the rule nets. Slave translators are also provided to convert rule calls from the rule nets, which may include data, into API calls to be executed by a von Neumann host where a slave translator resides. The results returned from the API call or von Neumann processor calls are then returned to the rule nets. A global controller has a global queue to arbitrate and store the rule calls between the rule nets and the slave translators. The global controller also stores global data and current processing state. The rules are a dynamic length group of variables connected together which include at least the Function State number of the rule and the inputs and outputs of the rule.

Additional objects and advantages of the invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by the practice of the present invention. The objects and advantages of the invention may be realized and obtained by means of the instruments and combinations particularly pointed out in the appended claims.

The above and other objects, features and advantages of the invention will become apparent from a consideration of the following detailed description presented in connection with the accompanying drawings in which:

FIG. 1 shows a distributed digital processor and the relationships between the global controller, global bindery and an individual rule net;

FIG. 2 shows a distributed digital processor and the relationships between the global controller and an individual slave translator; and

FIG. 3 shows a distributed digital processor with a communication bus and several connected processing nodes.

FIG. 4 shows a distributed digital processor on a clustered network;

FIG. 5 shows a distributed digital processor implemented in hardware using a PCI and ISA bus;

FIG. 6 shows a distributed digital processor with a hierarchy of connected global controllers;

FIG. 7 shows a section of a distributed digital processor with a detailed view of the local controller and the Other Node Buffer

FIG. 8 shows a local controller and rule net in a distributed digital processor and a partially abstracted view of the organization of the rule net;

FIG. 9 shows a slave local controller and a slave translator in a distributed processor and a partially abstracted view of the organization of the slave translator;

FIG. 10 is a view of one possible configuration of a digital rule stored in a rule net.

Reference will now be made to the drawings in which the various elements of the present invention will be given numeral designations and in which the invention will be discussed so as to enable one skilled in the art to make and use the invention. It is to be understood that the following description is only exemplary of the principles of the present invention, and should not be viewed as narrowing the appended claims.

Referring now to FIG. 1, this invention allows an application to be compiled through a specialized compiler which creates a number of digital rules to be executed in parallel over a number of processing nodes which are connected by a bus, network or a similar communications device. Even if one of the processing nodes crashes, the distributed rule processing environment is designed to continue executing the program without the loss of data through mechanisms described more fully herein.

The specialized compiler maps source code, such as a modified "C" or Java programming language, into sets of rules to be instantiated (loaded and stored) in a rule net storage 100. The compiler is designed to create rules which are unique and not minimizable. A digital rule can contain parameter information and as will be seen through this disclosure, may be compared to a function call or function in a conventional programming language. (Referring now to FIG. 1, these rule nets can call digital rules (i.e. functions) on other rule nets 100 or they may call the digital rules on slave translators 600 (FIG. 2). The digital rules have no direct data manipulation capabilities because they have no registers or data storage. A local controller 200 does all the data processing, taking orders from state chances and rules sent over the global bindery by the global controller 300 from the other rule nets. Return states to the rule nets are also broadcast over the global bindery 400.

FIGS. 1-2 illustrate a distributed digital processor and the relationships between the parts of the processor. Initially an overview of the relationships between each part of the distributed digital processor will be described, followed by a detailed explanation of each part and how they interrelate. A rule net storage 100 stores the digital rules. FIG. 1 shows one rule net 100 but multiple rule nets 100 can co-exist on the same node or distributed across several nodes. Source code created by a programmer is compiled by a specialized compiler into digital rules which are a conjunction of boolean variables to store states or values. These compiled rules are stored in the rule net storage 100 until they are executed. There are two types of rules in this invention: 1) a rule which is stored in a rule net 100 with its Function State and accompanying input variables and output variables, and 2) a broadcast rule which contains many of the elements of a rule in a rule net but it is broadcast without a Function State and has a different format as described in tables later in this description.

A local controller 200 is coupled to the rule net storage 100 and receives broadcast messages from the global controller 300 as to which rule in the rule net 100 should be recognized and what rule should then be broadcast to the global controller 300. The local controller 200 also broadcasts messages (calls) as to which digital rules should be recognized (i.e. executed) in other rule nets or slave translators as a result of the rules retrieved from the rule net storage 100 and executed in the local controller 200. Each local controller 200 has one or more von Neumann processors to allow it to process rules and make broadcasts.

A local bindery 500 is coupled to the local controller 200 to store the current state of a thread of execution from the program which is being executed in a local controller 200 through the rules stored in the rule net storage 100. The current state is called the Function State. The local bindery 500 for each rule net also stores parameters passed when a rule is received or broadcast. This allows each rule net to remember the parameter values which were broadcast and use them throughout the rule's operation. The parameters are buffer handle numbers or constants which the von Neumann processors use to process data passed from the global data memory 301 or from other rule nets 100 and their local controllers 200. Multiple instances of the local bindery 500 can exist to service multiple threads of execution and recursion for the application, or threads from different applications. Based on this disclosure, it can be seen that the local bindery contains similarities to a stack in conventional von Neumann architecture. Further references to a "rule net" will refer to a rule net storage 100 combined with a local controller 200 and its local binderies 500.

The local controller 200 is connected to a global controller 300 which stores the overall state of the digital processing network and controls the broadcast flow of the rules. The global controller 300 contains a global data memory 301 to store the global data for the executing application or applications. The global data could be stored in another area separate from the global controller as long as the data can be directly accessed by the global controller 300. The global data memory 301 must also be separate from the rule nets. This creates a separation between the code and data which is not found in von Neumann machines and it creates an inherently more stable system that cannot be easily corrupted. Von Neumann machines have the drawback that executing code can sometimes overwrite itself or its data thus crashing the executing code. For example, exceeding an array bound is likely to destroy code or other data.

A global bindery 400 is coupled to the global controller 300. The global bindery 300 has a queue to store digital rules which are individually broadcast to the local controllers 200 and slave local controllers 700 on the communications network. Digital rules which are broadcast by the local controller 200 for execution by other local controllers and their respective rule nets are received by the global controller 300 and stored in the global queue 400. The global queue 400 is preferably a first in first out (FIFO) queue which has a number of rule storage positions equal to the number of rule nets in the network.

FIG. 2 shows the configuration of an individual slave translator coupled to the global controller 300 and global bindery 400 in the distributed digital processor. The slave translators operate in parallel with the rule nets to process the compiled application. The slave translator storage 600 and the slave local controller 700 may be stored on the same node as the rule nets and their local controllers, or spread across other nodes on the bus or network. The slave translators differ from the rule nets in that they actually perform the requested work for the application as opposed to the rule nets which control the organization of the program. The slave translator also has digital rules stored in the slave translator storage 600. As mentioned, the digital rules have no direct data manipulation capabilities, so the von Neumann processors coupled to the slave local controller 700 do all the data processing. The slave local controller 700 takes orders from state changes and rule broadcasts sent by the rule nets through the global controller 300 over the global bindery 400 and it sends return states to the rule nets over the global bindery 400.

When a slave translator rule executes, it calls an application programmer interface (API) on the host von Neumann node which will actually perform work. For example, a database API may be called to do work on the global data passed from the global controller 300. The data is then returned to the rule net which called the slave translator rule. Slave rules do not call other rule nets and they do not call other slave translators directly because this would result in gray data which could be lost.

The slave local controller 700 operates in much the same manner as the rule net local controller 200 in that it receives broadcast rules from the global controller 300 and compares them to the rules in its slave translator storage 600. The difference between the slave local controller 700 and the rule net local controller 200 (FIG. 1) is that the slave local controller 700 calls slave functionality 900 through API calls and processes the actual work for the user's application. These API calls will be calls to applications such as database, email, hardware peripherals or operating system services. It should also be noted that the slave local bindery 800 is similar to the local bindery 500 because it stores the states of each thread being processed by the slave local controller 700. Multiple copies of the slave local bindery 800 can exist to process recursion or multiple threads of API execution. The slave local bindery 800 also stores the parameters that are passed with a particular broadcast rule and the source address of the calling rule net. It should also be mentioned that global controllers, local controllers, and API services must be serviced by at least one slave CPU on a given chassis. Every global controller 300 must have at least one von Neumann processor accompanying it. The slave local controllers 700 and rule net local controllers 200 and global controller 300 may share the same von Neumann processor or each have their own dedicated von Neumann processor.

The distributed digital processor of the current invention has a communication cycle. During each cycle, each local controller 200 and slave local controller 700 checks the rule that is being broadcast from the global bindery 400 by the global controller 300 to see if it recognizes the broadcast rule. In other words, each local controller 200 checks to see if the broadcast rule matches a rule listed in its rule net storage 100 or slave translator storage area 600. If the broadcast rule matches a rule in a specific controller's storage, the rule net or slave translator raises a flag that it desires to receive that rule and process it. This flag is actually broadcast on the network and is known as an Accept broadcast. The broadcasting rule net (or original calling rule) then loops in a Wait state at a wait rule and waits until it receives an Accept or it eventually times out. When an Accept is broadcast, the broadcasting rule net then picks the first rule net or slave translator who accepted the rule (broadcasted an Accept) and stores the From Node variable's value which is the address of the calling rule net, so further rules in this thread may be sent directly to that rule net. The accepting rule net or slave translator then processes the rule to completion and broadcasts a Complete rule broadcast to the calling rule net, including any data that may have been processed (by slave translators). Meanwhile the calling rule net has been in a Wait state to wait for the Complete. When the calling rule net receives the Complete from the first original accepting rule net for slave translator, it knows the rule is complete and it captures the completed data broadcast with the Complete broadcast.

Since this is a parallel system, many rule nets may Accept the broadcast call and they may run to completion for that first call. Then when they broadcast their Accept they are ignored by the calling rule net because it will only recognize the first accepting rule net whose address it has stored. Then the calling rule net will send the further rules in the current thread to the same rule net. Although many rule nets will finish one rule this is not as wasteful as it appears because usually the first call is just to initiate the transaction, in other words a "handshake" before more work is sent. Further rules are not sent to the rule nets who were not the first to Accept and they are ignored. Therefore, they actually do very little work for a general broadcast.

Rules will be sent with global data if there is any to be transferred. The local controller 200 will acknowledge that it received the rule and then process it along with global data transferred and then stored in the local bindery 500. Then the local controller 200 will broadcast a rule which needs to be processed by other rule nets. These broadcast rules will be stored by the global controller 300 on the global bindery 400 (FIFO queue). Then the communication cycle begins again as each local controller 200 and slave local controller 200 checks the global bindery to see if the rule broadcast by the global controller 300 matches a rule in its rule net storage 100 or slave translator storage 600.

It should be realized that there are many local controllers 200 and rule nets operating simultaneously to process the rules stored in the global bindery 400 and broadcast by the global controller 300. Depending on the processing power desired for the global controller 300, up to eight processors could be reasonably assigned to one global controller 300 and global bindery 400. Even more processors could be used depending on the size of the address variable defined to address the processors.

FIG. 3 shows a distributed digital processor with a communication bus 1006 and several connected individual processing nodes 1000. The communication bus 1006 may be a bus such as a network radio network or a bus using another common protocol such as ISA or PCI. The processing nodes may each reside within their own physical chassis or box or combined together on a single board inside a single chassis box. Each processing node 1000 is independent and will have at a minimum its own hardware 1002 with at least one von Neuman processor, a dynamic storage device such as RAM, and a static storage device such as a hard drive or other static storage array. Additional von Neumann processors may be added to the individual processing nodes in a known configuration such a symmetric multi-processing (SMP) to increase the speed of the individual processing node 1000. Of course other input and output hardware could also be connected to the individual processing node 1000, such as a conventional monitor, keyboard or mouse, etc. A conventional operating system 1004 also runs on the nodes 1000 such as NetWare, UNIX, Windows, Windows NT, DOS or some other type of operating system. Running on the operating system 1004 are the conventional slave application services 900 such as the operating system APIs, database applications, web services, email or other applications. The slave translators 600/700 call the slave application services 900 to perform work on the global data received from the global controller 300.

The global controller 300 and global bindery 400, are backed up (replicated) on each processing node 1000 connected by the communication bus 1006. Only one global controller 300 and global bindery 400 at a time will be active to control a given group of rule nets and slave translators. So if a processing node 1000 fails, then the backup copy of the global controller 300 and global bindery 400 on another processing unit 1000 will take over at the last point it was copied and no gray data will be lost. Then one of the remaining rule nets 100/200 which has not received an Accept or Complete back from a rule broadcast will re-broadcast that rule and rule processing will resume. Although the global controller 300 and the global bindery 400 will be replicated across several nodes, it only needs to be replicated enough times to create a stable system and not necessarily across every node on the system. The rule net storages 100, their local controllers 200 and local binderies 500 will also be copied to a selected number of nodes in case the executing rule net fails. Even though the time it takes to make the backup copies of the digital processor's components does reduce the efficiency of the invention, this a trade-off for greater stability. The use of a low latency, high bandwidth interconnect such as Virtual Interface Architecture (VIA), will reduce the backup overhead.

All of the active rule nets can reside on one node and then be copied to other nodes or the active rule nets can be divided between many nodes and then backed up to other nodes. For example, if there were 3 active rule nets in the system all three could reside on one node (A,B,C) as a group or they could each reside on separate nodes where the first node has active rule net layer A, the second node has active rule net layer B, and the third node has active rule net layer C. All of these rule nets can then send broadcast rules to each other and backup copies are also made of the active rule nets on other nodes.

Slave translators only exist in conjunction with the slave application services 900 provided by an API on a specific node. For example, a slave translator may be mapped to a specific database API. In this case, a slave translator can only be used on the specific processing node 1000 where an actual database resides. For the sake of stability, a slave translator mapped to a database API should exist on more than one processing node 1000, so that if one processing node crashes another slave translator can take over those API services. If a slave translator contains rules which cannot be processed on a specific node 100, those rules will be inactive.

FIG. 4 shows a distributed digital processor on a clustered network. The distributed digital rule processor exists on multiple nodes in a Local Area Network (LAN) or cluster. Network address variables are combined and broadcast with the digital rules to handle the addressing of multiple network nodes, multiple processors in each node, and multiple threads in each processor. The communication bus 1008 shown is preferably a high speed LAN. The communications bus 1008 could also be a wide area network (WAN) or even the Internet. Each individual processing node 1000 has a global rule engine, rule nets, slave translators and binderies 1012 to execute rules over the specified network and utilize the resources of each individual processing node 1000. The individual nodes 1000 on the network in FIG. 4 are network servers and have a network operating system such as Novell's NOS, Microsoft's Windows NT, UNIX or another networked operating system. A global address array, known as the other Node Array, is coupled to each rule net local controller 200 to map the value of each address variable's thread number and source group number to the slave translator's node address (chassis), von Neumann processor and thread address. This process will be described in detail later.

In addition to the networked software, there is a cluster resource manager 1014 which can take care of failover, failback and resource migration. An example of an application running on the cluster is shown as a database engine 901 or a HTTP server engine 902. If one server node 1000 were to fail, these slave application services 900 would be transferred to another server node 1000. The slave application services 900 can call operating system services 1004 or other application services which are not shown.

FIG. 5 shows a hardware implementation of the invention where each unit is a hardware card connected within the same physical chassis. Each node on the system shown in FIG. 5 requires its own von Neumann processor(s) with an ALU and its own conventional operating system. Each node also requires its own random access memory (RAM) and hard drive like the individual node 1000 described in FIG. 3. The main rule net 1100 is the rule net that would receive the first broadcast rule and then broadcast rules which would be received by the other rule nets 1300 and the slave translators 1200. The slave translators 1200 each have slave applications running on their operating system which would process the actual work requests of the rule nets 1300. This embodiment of the invention allows a whole digital rule processor to be configured in a desktop type system. In another embodiment of the invention, the rule nets 1300, slave translators 1200, global controller 300 and global bindery 400 can each be an application specific integrated circuit (ASIC) configured as shown in FIG. 5.

FIG. 6 shows a logical representation of multiple digital rule processors connected in a hierarchical configuration. The digital rule processor, logically represented as P, is distributed across the single board computer (SBC) nodes. This embodiment allows global controllers to be connected to other global controllers to divide the rules from multiple applications compiled under this invention between subservient global controllers. This hierarchy allows simultaneous application processing to be performed and farmed out to global controllers lower in the hierarchy and yet the rules are executed in an orderly fashion without having to timeshare scarce resources such as the global queue and bus. The fact that rules from an application are distributed across many individual nodes is a major difference between systems using only von Neumann processors (serial computation devices), and the digital parallel controller device of the current invention. Furthermore, networking of an application is a natural byproduct of the global controller and rule net behavior, rather than a contrived emulation done in software.

It should be mentioned again that the distributed digital processor trades speed for reliability. The distributed nature of the rule processor is not as fast as direct von Neumann processing but it is very stable and will not become indeterminate or crash. To aid in efficient processing, the rules or application instructions in the current invention should be less granular than a program written for a von Neumann processor. The coarser the granularity of rules in the target application, the faster the performance of the current invention because the actual work is divided into its granular parts by the APIs which are called. Having a coarse level of granularity also allows the digital processor to send instructions from the same thread directly to the same slave translator without having to broadcast a general broadcast rule which can be accepted by any slave translator. The application programmer must use the high level source code to overcome the processing inefficiency of this invention.

This invention is more stable than von Neumann processing systems for a number of reasons. First, there is a separation between code and data. In a normal von Neumann machine, the code and data are in the same data storage area. In the current invention, the executing application's data is stored in the global controller and the rules (code) are stored in the read only rule nets. Second, there is no instruction pointer to be corrupted. A von Neuman machine which has a corrupted instruction pointer will go into an indeterminate state and crash. The current invention cannot crash because when one of the individual processing nodes crashes, the remaining rule net(s) will realize that the last rule it broadcast was not completed by the rule net or slave translator who accepted the rule and the remaining rule net will rebroadcast the timed out rule and processing will resume.

The third reason that the current invention is more stable than von Neumann systems is the redundancy of the system. All the elements of the invention are backed up on other individual nodes. For example, the global data of the application is backed up on many nodes and if one node crashes, the data has not been lost and the state of the network has been stored. As stated above, the rule or thread that crashed can resume because the broadcasting rule net or even a copy of the broadcasting rule net will realize that it did not complete and will rebroadcast its rule and the processing will continue.

A fourth reason that the current invention is more stable than von Neumann architectures alone is that the local bindery contains similarities to a stack in a von Neumann machine, but it is not corruptible, or able to overflow from improper stack variable usage. In addition the plurality of local binderies allows for true parallelism.

A final reason for the stability of the current invention is that the compiler generates slave calls to allocate a separate buffer from the global data memory pool which is divided up specifically for local variables. Even if the buffer is corrupted, it will not affect other buffers on other memory pages or affect the local bindery. So the invention cannot lose control because of corrupt data. In contrast, local variables in von Neumann functions are allocated on the stack which can be corrupted.

Now each element of the invention described above will be discussed in further detail, and the specific states, storage location names, parameters and data passed through the digital rule processor will be discussed.

The variable storage locations in the distributed digital processor contain three segments: 1) a variable number, 2) the variable value, and 3) the variable type. The actual lengths of the variable segments are determined based on the available transmission speeds in the system. If the variable segments are larger then the transmission time will be longer and if the segment size is smaller then the transmission will be shorter. The preferred sizes of the variable parts are 16 bits for the variable number and variable value and 8 bits for the variable type. The segment lengths could be longer or shorter depending on the actual implementation of the invention.

The variable type has three sub-segments which are included in the variable. The variable type is composed of 3 sub-segments: an Indirect Flag, an Input/Output Variable, and a Variable Buffer type. The Indirect Flag indicates that the buffer holds another buffer number and offset.

The Variable Buffer Type indicates the size of the indirect buffer. This table shows the states of the variable buffer type:

The Output/Input variable has following 4 states as shown in the table below:

The Output.sub.-- to.sub.-- Input variable is essentially an output variable which is immediately converted by the global controller 300 into an input variable for the new rule being called.

Referring to FIG. 7, each local controller 200 has a rule net storage 100, and the local controller 200 is coupled to the global controller 300, global bindery and queue 400 and has access to the local bindery 800 (or local binderies). The local controller has logic to recognize a broadcast rule from the global controller 300 and global bindery 400 against the rules in the rule net 100 storage. If the destination is outside the broadcasting node's physical chassis, data buffers from the global data memory pool will be copied along with the rule. The destination address for each source region is stored and retrieved from the Other Node Buffer 210. The Other Node Buffer is essentially a hash table for storing the To.sub.-- Node and From.sub.-- Node and Destination.sub.-- Addr addresses. The hashing key for the Other Node Buffer is generated by an offset created from the source region (defined immediately below) and thread number. At any given time this table will be fairly sparse depending on the number of threads executing on a specific node.

At this point a source region number will be defined. Related functions are grouped together in a source code header (e.g. with a file with a .h suffix). These groups need to be recognized by the invention because they are related and will be sent to the same thread for processing. At compile time a table is generated which defines source region numbers of related function groups in the source headers.

In one embodiment of the invention, a specialized group of functions is designated which are the fundamental operations of an ALU as well as interfaces to I/O devices such as the keyboard, screen, disk files, mouse, serial ports, etc. All of the standard library functions such as string manipulation, file handling, floating point processing, etc. are also contained in this group. Variables stored in each rule net's local bindery, as well as the address mapping array, are in this group. It is possible to broadcast these fundamental operations across the bus or network communication device, but it is inefficient to send the fundamental operations across the network or bus. So, these fundamental operations are grouped together and recognized so they can be sent to the local von Neumann processor used by the local controller or slave local controller where the fundamental operation originated. Sending a fundamental operation to a von Neuman processor on the local chassis avoids the overhead of broadcasting it across the network and the processing associated with such a broadcast.

The local bindery 800 holds the status variables needed for: Function State 802, Counter Overflow 804, Rule Net Parameters 806, a Temporary Buffer 808, a From Function value 810, and a This Node buffer 812. Unlike a serial von Neumann processor which uses the instruction pointer as a state control for the code, a rule net requires a Function State variable to define which group of parallel rules may be recognized next. The address mapping array, known as the Other Node Buffer 210, is coupled to the local controller so the network chassis-processor-thread addresses are stored by thread-source code group in the address array. The thread number corresponds to a thread state stored in one of the local binderies. As the local controller 200 compares rules for different threads it switches between the copies of the local binderies corresponding to the thread. It should also be mentioned at this point that the slave translators have functionality and local bindery construction similar to a rule net, with the exception that a slave local controller can call slave functionality.

The Counter Overflow 802 is provided to limit the time a rule net should wait for a response to a digital rule it has broadcast for processing by another node. For example, an API call on a slave translator may fail to complete on time, and this counter sets a time limit after which the broadcasting rule net times out and either re-broadcasts the rule or enters an error recovery state. With potentially many rule nets asynchronously sending commands and receiving results, the potential for a failure may occur from time to time. This is especially true of a hardware failure in an I/O device (e.g. a disk drive). This Counter Overflow variable can be accessed for comparison to an internal clock and it can also be reset. It should also be mentioned that the Function State and Counter OverFlow are local variables only and never are broadcast through the global bindery 400.

When a rule net sends out a digital rule, there may be parameters which accompany the command. These parameters are stored in the Rule Net Parameters 806. When the local controller or slave translator recognizes that a call has been made to it, the parameters accompanying the call are saved in the Rule Net Parameters 806. Then the buffers supplied in the Rule Net Parameters 806 are used for further data manipulation. If the call came from another node or physical chassis, the data accompanying the rule is copied by the global controller 300 into the global data memory 301. Based on this disclosure it can be seen that the rule net parameters can be compared to parameters on a stack in a von Neumann machine.

Every rule net needs a work space to use as a stack for storage of temporary calculations and this work space is defined as the Temporary Buffer 808. When a rule net broadcasts a rule to a slave translator it tells the slave translator to use the Temporary Buffer 808 for storing the results of the broadcast rule. In order to optimize memory usage and not waste memory on idle rule net, which most will be at any given moment, permanent storage is not allocated to each rule net. The local controller 200 must allocate temporary memory at the start of each rule net execution, and free the memory at the end. The memory allocated depends on the storage requirements of rule net which is executing. The buffer handle number of the allocated temporary memory is stored in the Temporary Buffer register.

In a parallel digital rule processing environment any rule net can call another specific rule net. A race condition can occur where multiple nets call a specific net at the same time. This rule net must reply on the global bindery to acknowledge which caller rule net has won. The caller net number is stored in the From Function variable 810. Every rule broadcast in this invention has a From Function variable 810.

The Other Node Buffer 210 is the global array which stores the network chassis-processor node-thread number mapped by thread number and source region number. This array applies to all rule nets under the global controller and is essentially a listing of the address and source region of the calling rule net who created the specific threads in the local bindery. To.sub.-- Node, From.sub.-- Node, Destination.sub.-- Addr variables in rule broadcasts use this array of addresses as a source or sink. When the Destination.sub.-- Addr, To.sub.-- Node, From.sub.-- Node variables are used as a source, the local controller retrieves the data from the This.sub.-- Node register (which is unique for each node) to determine the address of the current node (chassis and processor) so that it can be sent with the rule broadcast. When these variables are a sink they contain the return addresses for the rules. For example, in slave translators the return address must be stored for the API calls that the slave translator has running. If there are database, email, or other services supported for the slave translator as well as the basic arithmetic, string, buffer, file I/O calls, then the return addresses are stored in the Other Node Buffer. The thread number comes from the data portion of the variable in the Other Node Buffer array. The preferred method of finding the array offset hashing key) is by using a 5 bit thread, and an 11 bit include number. In its preferred embodiment the data in the array offset is 8 bits for the chassis number (node in the cluster, with 0 not used), 3 bits for the processor number (single board computer or main CPU on the motherboard, with 0 for the global controller), and 5 bits for the thread, which can be ORed (a logical union operation, hereafter ORed) with the 5 bit thread from the data portion of the variable making the access. Of course, the size of this array could be larger or smaller depending on the overall size of network and number of von Neumann processors used.

This.sub.-- Node 812 can also be directly accessed by the global controller and is global to all rule nets and slave translators on a given physical chassis. It stores the current chassis-processor number. The This Node value is ORed with the particular thread number of the broadcast rule. The thread number comes from the data portion of the instance variable using this buffer. The preferred size of the thread data is 5 bits, and the group region is 11 bits. Again, the size of this value could be larger or smaller depending on the overall size of network and number of von Neumann processors used in each chassis.

Referring now to FIG. 8, the detailed operation of the local controller 200 will be discussed. When a rule broadcast is received, each local controller 200 attempts to recognize a rule from the rule net 100 by using at least one input variable 104, at least one output variable to set the next state 106, and a Function State 102. The Function State 102 for each rule is generated by the compiler and is strictly increasing. Several digital rules stored in the rule nets can have the same Function State 102 value in order to handle parallelism. For example, all the rules are sequentially numbered from 1 to N, and parallel rules will have the same Function State 102. The Function State 102 variable is both an input and an output variable because each rule needs an input Function State to help match the current state and an output Function State to determine the next Function State.

There are two types of rules in this invention, one which is stored in a rule net storage area 100 and can be matched with incoming broadcast rules. The second is broadcast rules which can be broadcast over the bus through the global controller 300 and global bindery 400. Broadcast rules can be grouped into four categories. First, Call instances which tell another rule net or slave translator to do something. This can be compared to a function call in von Neumann assembly code. Each call may have parameters, as well as return chassis-processor-thread information. The call is recognized by a local controller via a rule stored in a rule net storage or on a slave translator which then commences a sequence of Accept and then Completion or Failure messages. In the case of a slave translator call, the call must first be Accepted. After the call is accepted then work is done on a von Neumann processor called by the slave translator, to calculate an arithmetic operation, or process a database action. For example, after the slave translator rule's work is complete a Complete or Failure rule is broadcast.

The second category of broadcast rules are Accept rules, as mentioned above. Accept rule instances are broadcast to acknowledge that the receiving rule net or slave translator will process the requested action. The rule net or slave translator receiving the Accept broadcast stores who accepted the requested call and the address to be stored in the Other Node Buffer is returned to the calling rule net. If the calling rule is a general broadcast to any node, not a specifically addressed call, then the return Accept's From.sub.-- Node address will be saved in the Other Node Buffer at the index key number created by source region-thread so the address can be used by subsequent calls. The information regarding which rule net or slave translator accepted a call is very important because a rule net will then send all of the other rules which belong to that pre-defined source group and thread to the same slave translator to increase the throughput of the digital rule processing system. Whether or not a call is accepted by a rule net or slave translator will depend on the current processing load or hardware capability of the von Neumann processor(s) used by the slave translators or rule nets.

Completion instances are the third type of broadcast and are sent and recognized after the call has been processed and completed. Failure to successfully complete the task is also a completion. Often there is a return result in one of the parameters, or in the return value of the function.

The fourth potential broadcast rule is a Change of Execution instance. This type of instance will be broadcast when the compilation results in return statements, goto, else, and the internals of an if statement conditional, or for, while, or switch, case, break statements. These usually don't generate an external broadcast, but can be considered an internal broadcast within the rule net to change the internal state of the local controller and rule net. Because they are not broadcast, they do not contain From.sub.-- Function, To.sub.-- Node, From.sub.-- Node, From.sub.-- This.sub.-- Node, To.sub.-- This.sub.-- Node, or Destination.sub.-- Addr variables, which are contained in each of the other types of broadcast instances.

Referring to FIG. 9, the slave translators and API's 606 are called by the rule nets and executed by the slave local controller 700 to call the slave functionality 900 and complete the actual work in this invention. In order to make as efficient use as possible of the slave functionality 900, each one has its own von Neumann processor with an ALU, local memory, and disk drive upon which resides its own version of an operating system. The slave functionality has a separate, asynchronous relationship with the slave translator, and may talk to other von Neumann processors if desired. Ideally, a von Neumann processor will accompany each I/O device, such as a network card or disk drive. One processor should do the graphical user interface for the screen, keyboard and mouse. The user interface processor should be the main central processing unit of a conventional personal computer (PC).

Four additional von Neumann processors is a reasonable number to use for I/O device control without overburdening the bus capacity. The preferred implementation of this invention is able to address 7 von Neumann processors per slave unit. 3 bits are used for addressing the processors and 0 is the global controller. The use of more than 7 processors in a chassis could easily be accommodated by increasing the size of the address space, which means that the Other Node Buffer element size would have to be greater than 16 bits.

The slave functionality 900 can normally be allocated to specifics tasks. For example, in processor intensive tasks, such as virtual reality systems, the von Neumann processors have no I/O interaction, but strictly process data and return the result. Another example is a disk drive processor which can take all the file object calls, etc, or the network card processor. If only one von Neumann processor accompanies a slave local controller 700 and slave functionality 900, then all the predefined slave objects that are not supported by the von Neumann processor will be unavailable. Any user defined objects can be broadcast over a network to other rule nets or slave translators on nodes in the cluster.

The broadcast cycle of the invention and the broadcast instances which are dispatched will now be discussed in further detail. The order in which the previously compiled rules should execute must be broadcast in order to actually execute the digital rules stored on the rule nets or slave translators. The Function State number of the first rule to be executed by the invention is stored in the global controller in an initial register and is broadcast to initiate the target application. This would be equivalent to the function call of main() in a C program.

Because the invention is an asynchronous parallel machine several rule nets may recognize and broadcast rules. The global queue stores these rules. The rules are broadcast one at a time by the global bindery for all slave translators and rule nets to see and attempt to recognize. Once the broadcast rule and its parameters have been seen by all the rule nets and slave translators, the rule is discarded. Referring now to FIG. 10, if one of the rule nets or slave translators recognizes the Function State and the input variables 1400 of the broadcast instance, it stores the output variables 1500 from the broadcast and uses them as needed. Then each rule net can broadcast the rules for other rule nets or slave translators to recognize and execute. These broadcast rules are stored in the global queue. Each rule net can also buffer rules to be broadcast if the global queue is full and then broadcast the stored rules to the global queue later.

Referring again to FIG. 9, the slave translators can only recognize object calls. This call is denoted by a Function State 602 followed by the object number 604 which is an input variable. Then the called slave translator returns an Accept along with a From.sub.-- Function number (caller function number). These are both input variables. This From.sub.-- Function variable is necessary to guarantee that the correct caller rule net knows that the slave or rule net has accepted its net call in case more than one net calls this slave translator or rule net at the same time. All calls have a From.sub.-- Function output variable which stores data identifying the rule net number. This variable is also used to identify where this call came from so when the return data or return state is complete it can be returned to the correct calling rule net. After generating an Accept, the slave translator will call the slave functionality 900 which we defined by the API outputs 606 of the slave translator rule.

All calls have a From.sub.-- Node variable, of output type, which stores data identifying the thread number and the source region. This data is ORed with the This.sub.-- Node buffer which can be seen globally in each local bindery. This.sub.-- Node contains the chassis, and processor number (0 if it is the global controller), ORed with the thread number of the current net thread. If a rule is recognized by the recipient slave translator or rule net, the From.sub.-- Node variable of type Store.sub.-- Output, is stored in the Other Node Array (global address array). As described before, the hashing key for the entry in the Other Node Buffer is determined by the data of the recipient instance From.sub.-- Node variable containing the thread and source region number. Basic arithmetic operations (+ -*/), string manipulations (strcpy, strcmp), memory and buffer calls (copybuff) and all basic I/O calls to the screen, disk, keyboard, mouse are from a pre-defined source region number (typically 0), and do not get broadcast outside of the chassis over the network as discussed earlier.

Some calls are made to all nodes in the network cluster in order to find one node-processor-thread which will process all of the high level API (application programmer interface) calls belonging to a particular source region. As described above, the source region is defined by APIs groups in source code headers during compilation. These source region groups typically belong to the same application or call related APIs such as database calls, or graphics calls. If the broadcast rule call is sent to a specific address destination, then the broadcast rule call also has an input variable To.sub.-- Node with data which identifies the thread and source region. This data is used to get an offset into the Other Node Buffer (global address array). The thread is ORed with the chassis-processor from the Other Node Buffer array.

Broadcast calls also have output variables which are stored as parameters. Storage of the parameters occurs in order, so the compiler must enforce parameter type checking. In the called rule net or slave translator the entire variable is stored, including the variable number, value, and type, and the variables are stored in the local bindery parameter storage. These values are used in slave translators as buffer handle descriptors or constants.

When a rule call is made to a slave translator, data will be needed for that slave rule to execute. The global controller checks to see if the destination of the rule is outside the physical chassis, and if it is then data will be transferred with the broadcast rule. The preferred implementation of this transfer is done using a bus that is separate from the main bus. Transfer of data may be done by requiring that the rule net and local controller pre-load the data in temporary buffers connected to a secondary bus which connects all the nodes. This relieves the main bus from having to handle data which goes back to the global data memory, because once the slave object finishes processing, it returns the data to the global data memory. Using a duplicate bus helps reduce the data I/O bottleneck on the main bus. An example of a duplicate bus is shown in the case of clusters, such as those that use ServerNet, VIA or SCI cluster fabric for the alternate bus between cluster nodes, and not the Local Area Network. In the preferred embodiment, rule calls would also be made over the same high speed data path that the data is using. In an alternate embodiment of the invention where the entire invention is contained on one physical chassis, the bus could be PCI, ISA, EISA, or VME.

When a broadcast instance is accepted by a slave translator or another rule net, an Accept instance is echoed back. The data of the Accept rule includes the From.sub.-- Net number, and its type is Input. If more than one acceptance message is sent, the first one received is saved, along with the return address of the rule net or slave translator which sent the acceptance. In specific destination calls, or calls to source region 0, the Accept may be omitted because the Complete will return fast enough that an Accept is not needed.

If the Accept is from a broadcast call to all nodes, there is a From.sub.-- This.sub.-- Node output variable with data from the This.sub.-- Node buffer ORed with the thread from the data portion of the From.sub.-- This.sub.-- Node variable. This data identifies the accepting node to the broadcasting rule net. The Accept also has an output variable Destination.sub.-- Addr with data from the Other Node Buffer array. This information identifies the thread and source region for any other calls for the same thread (i.e. similar database or graphics calls.) The thread and source region information are determined through the following process. The offset key into the array is determined by the thread and source region number data of Destination.sub.-- Addr. The array value is ORed with the thread. The recognizing rule generated by this Accept has a From.sub.-- Node variable of Store.sub.-- Output type which will store the From.sub.-- This.sub.-- Node data into the Other Node Buffer array.

If the Accept is from a specific node call then the Accept instance already knew the destination and so it includes a To.sub.-- This.sub.-- Node input variable with data from the This.sub.-- Node buffer. There is also a Destination.sub.-- Addr output variable with data from the Other Node Buffer array, where the offset key is determined by the data of the Destination.sub.-- Addr variable as a thread and source region array. The rule net receiving the Accept rule has a To.sub.-- Node input variable which compares the data to its Other Node Buffer array ORed with the thread number from the data portion of the variable (the array offset is determined by the data portion of the TO.sub.-- NODE variable). This insures that the correct thread is being called.

Once a broadcast rule call has been made, and while the rule net which sent the call is waiting for the Accept rule instance to come back, a timer is reset in the calling rule net with a Reset.sub.-- Counter output variable. The calling rule net has another variable waiting at the same Function State as the Accept rule instance. It has a Counter.sub.-- Overflow input variable. The variable contains the number of thousands of a second to wait before it times out. If that number of seconds is exceeded without an Accept call, then the Function State is reset to issue the call again. If the overflow is triggered a set number of times without getting past this Function State, the Function State is set to the Error.sub.-- Exit level. At the Error.sub.-- Exit Function State there may be an error handling routine for the rule net. Even if there is no error handling routine this error handling rule will be called in the rule net if the Counter.sub.-- Overflow times out the set number of times. This means that error handling is an inherent characteristic of this fault tolerant, highly available computing system.

Once the slave translator has finished processing, it returns a Complete status, or Failed status. In the case of the parallel call, no finished status is expected, because the caller continues processing once the Accept status is received. Every completed call, whether it succeeds or fails, has a To.sub.-- This.sub.-- Node input variable using the data from This.sub.-- Node global buffer ORed with the thread number. There is also a Destination.sub.-- Addr output variable using an offset from the Other Node Buffer array, ORed with the thread number. The offset into the array comes from the data portion of the variable, consisting of the thread and include source region number. The recognizing rule for the Complete or Failure has a To.sub.-- Node input variable which is compared with the data from an offset in the Other Node Buffer array, ORed with the thread number. The offset into the array comes from the thread and include group number in the data portion of the To.sub.-- Node variable. These steps guarantee that the Complete matches up with the correct calling thread.

As part of the Call-Accept-Complete cycle, there is a Counter Reset output variable in the Accept call which is recognized. This starts the counter waiting for a Complete or Fail rule broadcast. If the counter overflow is recognized before the Complete or Fail message, the Function State is reset to the calling instance. This is done a pre-set number of times, after which the net control logic sets the Function State to Error.sub.-- Exit.

In summary, there are two potential points of timeout failure in most call broadcast sequences: 1) while waiting for the Accept instance, and 2) while waiting for the Complete or Failure instance. In the case of predefined calls to a local slave von Neumann functions, the Accept part of the sequence can be omitted, and the calling sequence would simply wait for a Complete or Fail. Omitting the Accept on a network call is possible but would require a lengthy timeout for the Complete, which creates a problem for retrying broadcasts because they would take an inordinately long time period.

The following tables summarize the Call-Accept-Complete cycle. This first table summarizes the variables sent in a general broadcast cycle and is not addressed to a specific rule net or slave translator on a specific rule net.

The next table shows a rule broadcast cycle which is directed to a specific node and address.

Each rule net has an error exit state arbitrarily defined near the highest Function State (or rule number). Rule net exit and return normally occurs before this state is reached or called. The Error.sub.-- Exit can be used to clean up temporary data, set error condition flags in global buffers, closing files, releasing memory, and other error handing routines.

Every rule net and slave translator has a beginning rule which is checked to allow it to recognize that it has been called. All the parameters from the call are stored in the local bindery of the rule net via a group of Store.sub.-- Output variables in the rule. In the preferred embodiment, a maximum of 5 parameters is passed, but the number of parameters which could be passed is theoretically unbounded. Using a limited number of parameters forces some calls to combine parameters into structures (records), which records get passed as pointers. The return variable of a rule call is also a buffer on the temporary buffer space of a rule net.

It should be mentioned at this point that the preferred embodiment of the Function State numbering system is designed for a 16 bit Function State scheme, starting at 0 for the first rule in every rule net. Larger Function State numbering values could be used by providing more bits, but a rule net larger than the number of rule instances which could fill a 16 bit numbering scheme would reduce the speed of the system. Keep in mind that many of the instances generated have the same Function State (such as the Counter Overflow, Complete and Failure rules, or the Accept and Overflow rules), so that a count of 65535 Function States will have at least 120000 instances in the rule net.

Now, the network addressing between the global controller, global bindery, rule nets and slave translators will be discussed. There are 2 ways that rules are broadcast. The first is a generally broadcast rule which does not know where it will be received until it has received an Accept back from a rule net or slave translator. The second is the broadcast of a rule with a specific node address is intended for a particular chassis (same chassis for predefined objects), a processor, and a thread. These directly addressed rule types are broadcast after a thread of execution has been previously set up in a rule net or slave translator. The global bindery which stores the network addressing can be extended to encompass a Local Area Network (LAN) or a Wide Area Network (WAN) with larger timeout tolerances. As the size and delay of the network grows, a longer Counter Overflow must be provided to allow for the propagation delay.

Unlike von Neumann processors, which are serial devices, the inherent nature of the distributed digital rule processor is a parallel network. In order to not broadcast all instances onto the LAN, a filter can be used to select only those instances which are known to need a larger sphere of communication. The use of the TO.sub.-- NODE and DESTINATION variables accomplishes this.

It is also important to note that a broadcast rule is truly a network message, not a serial emulation of network Remote Procedure Calls. A rule net does not care who accepts a rule broadcast, only that the response happens in a timely fashion. This networking permits large Wide Area Networks using satellites and terrestrial links. To the global controllers which may be organized in a hierarchical scheme, this networking appears as one large net, not as separate nets.

The distributed digital rule processor is a different, yet complementary approach to von Neumann processing. A global controller and global bindery enhances the high availability of a cluster of nodes by broadcasting messages to rule nets and slave translators on nodes, and von Neumann processors in each node, so that multiple instantiations of slave translators can provide responses to rule net calls. The distributed digital rule processor is scalable because a hierarchy of global controllers and global binderies (each with at least one von Neumann processor) can be assembled, allowing very large clusters with hundreds of slave von Neumann processors in hundreds of network nodes to interact and be centrally controlled.

The following listing is a pseudo code listing in table form of a chain of rule calls between an originator and a recipient node such as a rule call cycle between 2 rule nets or a rule net and a slave translator. The first set is a broadcast type of rule, and the second set is the specific addressing, where a specific address is used to communicate with a rule net or slave translator. The Funct.sub.-- State values P and P2 represent arbitrary starting points in the rule nets or slave translators which could be 1 or any other relative starting point.

1. Broadcast Type

During the function call recognition Output.sub.-- to.sub.-- Input types become input type variables. Then the originator node goes into a wait state as shown next. The wait state will wait for a certain amount of time and then time out. It will do that a preselected number of times before it fail (preferrably 3 times).

Hopefully the Originator will receive an acknowledge broadcast at the same Function State level before the wait rule times out.

If Parallel then stop and start another call

If there is a destination buffer

else if it is a complete with no data transfer

else if the call failed and the recipient has not crashed or terminated then

2. Specific Address Type of Broadcasts

Again on recognition all the Output.sub.-- to.sub.-- Input type variables become Input variables. The originator node will then wait for its timeout cycles (at least 3).

Hopefully the Originator will receive an Accept broadcast for the node it was specifically addressing at the same Function State level before the wait rule times out.

If Parallel then stop and start another call

If there is a destination buffer

else if it is a complete with no data transfer

else if the call failed and the recipient has not crashed or terminated then

The following examples are of rules in a rule net or broadcast rules. In the example rule instances, all of the numbers are in hexadecimal and use the data structure code directly following the example rules. Each instance has a byte stating total number of variables and the number of input variables. It should be noted that although the variable name and type are shown in the tables, the variable values and variable type are only stored as numerical representations in an order which is known to the local and global controllers and which are pre-defined in the compiler. Some of the defined variables and function calls are shown in the data structures following the example rules.

This instance shows 5 total variables and 2 input variables. This call does not have a TO.sub.-- NODE input variable, so it is a broadcast style to find a recipient slave translator or rule net to use for all subsequent calls.

This is an example of a rule stored in the rule net. It has 6 total variables and 1 input variable. When the function state is recognized and the instance is put on the bindery, the Output.sub.-- to.sub.-- Input variables become Input variables for recognition in the other rule net's set of rule instances.

This example rule instance is recognized in the set of rule net instances, and is broadcast to the local chassis only. This digital rule has 7 total variables and 1 input variable. Again note that the Output.sub.-- to.sub.-- Input Variable instances become Input variable instances when broadcast.

An example of a rule instance with 3 total variables and 2 input variables.

An example of a rule instance with 6 total variables and 4 input variables.

An example of a rule instance with 5 total variables and 3 input variables.

An example of a rule instance with 3 total variables and 2 input variables.

Now a list of defined variables and data structures is included to help explain the structure of the rules in the distributed digital processor shown immediately above.

This section is a pseudo code walk through of the execution of the current invention.

1. Once the digital rules for one or more rule nets are loaded into the DRAM of the rule nets, the global controller generates a main net number or the first function call broadcast onto the global bindery. This will be recognized by the main rule net and then other rule broadcasts will follow.

2. Each rule net receives the broadcast rules and checks to see if it has the rule and should accept the rule. All the rules for a single rule net are stored in Function State order. Thus the local controller of the engine only has to look at a small sequence of rules which match the Function State variable in the local net storage. Other rules in the rule net can be ignored until the Function State matches their first input variable, and the first variable is always the Function State. This economy of comparison allows one local controller to monitor several rule nets' digital rule instances very quickly, if desired.

3. If the local controller of a rule net finds a matching Function State, it proceeds to the next input variable, if available. If it is another local variable, such as the Overflow timer, it is compared also. A rule only matches if the Function State and all the input variables match.

4. If the input variable is a global bindery variable, then the local controller goes out to the global bindery to see if any of the variables of the first rule in the global queue match the input variable. If so, the local controller proceeds. If not, then the compared instance is ignored and the next instance is then compared using the Function State and input variables.

5. If a rule is found in the rule net where the Function State and all the input variables match, the rule is recorded as a possible best match.

6. The local controller must continue to try the next rules until the Function State is unequal. If another rule is found matching, then the local controller makes a decision. The rule with the higher Function State is remembered as the best match.

7. Once all matching rules have been compared, the best matching rule, if there is one, is selected. If a rule matches, then the rule net will broadcast output variables and update the Function State. If the rule net does not find a match it will not broadcast anything.

8. When a rule is selected it is "executed". Execution consists of the following steps. First, the Function State of the rule to be called and the output variables of the selected instance are queued up to be inserted at the tail end of the FIFO global bindery queue. Second, any local variables are processed immediately and affect the local bindery. This includes the Function State and the Counter OverFlow.

9. Part of the broadcast phase is the translation of the Output.sub.-- to.sub.-- Input type variable to Input.sub.-- Type when they are stored in the global queue. Then the data portion of the To.sub.-- Node, From.sub.-- Node, and Destination.sub.-- Addr is stored in the This.sub.-- Node variable or stored in the Other.sub.-- Node.sub.-- Buffer as described above.

10. If the output variable is a Store.sub.-- Output type which indicates that the variable should be stored, then the local controller checks the variable, if it is either:

A. a From Function or Temporary Buffer handle storage variable, the value is just stored directly in the local bindery. This is usually the first output variable in the global bindery instance.

B. a Parameter Store variable, then each variable's name, value and type from the global bindery are copied into the local bindery. The parameters must be in the same order and of the same number as on the global bindery. Typically they follow the From Function output variable.

11. Many rules have no global bindery output or data which should be returned to the global data memory. These rules in the rule net are processed by the local controller internally, causing the Function State to change without a global broadcast. This allows many rule nets to process without having to contend for the global bindery queue. The queue should be long enough to allow rule nets to broadcast and proceed without blocking. Even in a large system, the global queue is preferably not longer than twenty instances. Longer queues are possible but slow down overall processing and broadcasting cycle, because the global controller is handling too many rule nets.

12. Rules from rule nets may also generate a call to a slave translator. That call is stored on the global bindery along with any parameters that will be operated on.

13. Slave translators recognize rules in the same way that rule net does. First the Function State is matched and then the input variables are matched. Once a slave translator finds a match then the slave rule has a corresponding API and output variables (data from the global data memory) which are sent to the correct API where the data is manipulated by the appropriate conventional slave application (i.e. database, web server, e-mail, etc.).

14. Once the data has been operated on, it is returned through the API interface to the slave translator who broadcasts the data as part of a complete instance broadcast. The data is then returned to the calling net, if the calling net and the slave are not on the same node. The calling net may then either return that data to the global data memory, pass it along in a return call to a rule net, pass the data along in another net call or send it to a slave translator. The action taken by the net depends on the action stored in the next rule which was waiting for the complete message.

The following section is pseudo code for the compiler. It also explains how the rule structures are built and how the rules and slave translator rules are divided.

1. For each Net do a while loop getting tokens of source code.

2. Parse the tokens in a recursive descent parser. An intermediate language can be used to assist in optimization.

3. Every function call, assignment operation, comparison operation, arithmetic calculation, or branching decision generates a series of rules as described above.

4. A call to a slave translator or rule net is created, along with what this rule net number is (the From Function), and any parameters needed for data. Each parameter is usually a buffer handle, or an indirection to a buffer handle via a pointer in another buffer handle. If the rule is to be sent to a single location versus a general broadcast, the use of the To.sub.-- Node, From.sub.-- Node, Destination.sub.-- Addr, From.sub.-- This.sub.-- Node, To.sub.-- This.sub.-- Node variables are used as described above.

5. Immediate values can be used as parameters. Each of the 16 bit values used as a pair in variables can contain numbers from 0-65535. Anything larger must be stored in a 4 byte integer, or a floating point buffer.

6. Result values are returned in a buffer(s) which are one or more of the parameters.

7. In order to use fewer numbers of parameters, record structures are should be used. Each 256 byte or 4K buffer can contain many variables in a precise configuration. Each buffer number is the first 16 bit half of a parameter variable. Offsets into the buffer for record field locations are in the second 16 bit value. The variable type byte describes whether indirection is used, and the size of the field. For simple buffers with no record structure, the second value is 0.

8. Each function call is followed by two parallel instances. One sets the error Function State for timeout failure, and the other recognizes the Accept response from the receiving rule net or slave translator. In an optimized version, the Accept can be eliminated for address specific instances, which reduces bus and/or network traffic.

9. If the Accepted instance is recognized, then three more parallel instances may follow. If the source code function call is for a parallel execution thread, then there is no pausing for a completion status. Serial execution requires a pause for the Complete status. It will either be a Success, Failure, or Timeout. A failure automatically forces the Function State to the Error Exit location for cleanup. A successful completion sets the Function State to where the next function call starts. A timeout changes the Function State back to the rule which was broadcast from the rule net. After 3 timeouts (the number could change as needed), the timeout goes to Function State Error.sub.-- Exit. It is permissible to not wait for the Accept, but proceed directly to the Complete, Success, Failure, Timeout set of parallel instances. This is possible for predefined functions which go to a local von Neumann processor for immediate processing, or if the low latency network interconnect is very reliable and packets will not get lost on the network.

10. In the case of a branch decision, the Success or Failure of the comparison causes different Function States to be set, but not to the Error Exit. These function states are derived from a table which saves where the function state and instances are at the start and end of the IF, ELSE, GOTO, DO, FOR and WHILE statements. At the end of parsing each rule net, the table values are loaded into the respective rule variables. This permits forward and backward branching.

11. Each rule net which is compiled also has a sequence of rules generated before and after the body of source function statement calls (i.e. the opening and closing brackets imply some setup and termination activity). For object oriented source languages, such as Java or C++, constructors and destructors or finalizers could formalize the allocation of temporary buffers, initialization of variables and so forth.

12. Each slave translator or rule net sends an Accept broadcast back to the caller Net.

13. Then the rule net calls Get Buffer to ask for the 256 byte temporary work space. It waits for an Accept and a Successful completion instance.

14. The body of the net is executed.

15. At the end of the rules execution the temporary buffer is freed. This also has an Accept and Successful completion wait. A Failure instance is broadcast after the Error.sub.-- Exit to handle internal rule or rule net call sequence failures. This provides for cascading the timeout or other failures to higher rule net calls for quicker error recovery.

16. The Function State is returned to 0, where the rule net can wait for another call. If another rule net calls this net while in the middle of executing the same thread, the call will only be Accepted if there are multiple local binderies which allow for multiple threads. If a slave translator gets the call, it will queue it up and has a better chance of responding to the call, although the timing may be so long that the caller net may time out.

17. The use of multiple threads in the rule generation by the compiler allows for To.sub.-- Node, From.sub.-- Node to handle multiple threads of execution of the same rules.

18. With each set of local bindery registers is the thread number in a register. It is ORed with the middle bits of the To.sub.-- Node, From.sub.-- Node, From.sub.-- This.sub.-- Node, To.sub.-- This.sub.-- Node, Destination.sub.-- Addr variables to access the correct offset in the Other.sub.-- Node.sub.-- Buffer array of Node-Processor-Thread values. At load time the number of threads would have to be set for all rules. This hardware requirement forces the compiler, software developer and global controller to be in synch as to how many threads could be handled.

19. An alternative threading scheme which is more memory intensive is to have the compiler hard code the thread number into multiple rule generations. This would require that complete copies of the rule sets be used in every rule net with local registries. This would require more rule net storage space, but would save in having to automatically OR the thread number in the To.sub.-- Node, From.sub.-- Node, From.sub.-- This.sub.-- Node, To.sub.-- This.sub.-- Node, Destination.sub.-- Addr variables to access the correct offset in the Other.sub.-- Node.sub.-- Buffer array of Node-Processor-Thread values.

20. Because each Function State has a 16 bit limit, and F800 or higher is the error exit area, the number of rules in a single rule net are limited to however many will fit in function states less than F800 (63488). Because many of the instances are parallel and have the same state number, up to 120000 instances may be in a single net. This limits the number of source lines. If 7 to 50 rules are needed per source line, then 2400 to 17000 source lines could exist in one rule net. These estimates demonstrate that large functions as commonly found in commercial software will fit in this parallel machine.

21. The Function State number range of 3000-6FFF gives us 16,383 nets. This is a large enough set of function numbers for a large commercial software package. If larger number ranges were used, such as 32 bits instead of 16 bits, then much larger ranges of predefined functions, net functions, records, etc. could be used.

22. After compilation, the instance sets are stored in binary form on a secondary storage device (hard disk). At execution time they are instantiated from the secondary storage to the DRAM storage on the rule nets.

23. The main net number is specially recorded so that it can be used as the initial global function broadcast to start off the parallel nets.