Patent Grant
7716630
The invention relates to the control of computations in data processing systems and, more particularly, to managing parameters for graph-based computations.
Complex business systems typically process data in multiple stages, with the results produced by one stage being fed into the next stage. The overall flow of information through such systems may be described in terms of a directed data flow graph, with vertices in the graph representing components (either data files or processes), and the links or “edges” in the graph indicating flows of data between components.
The same type of graphic representation may be used to describe parallel processing systems. For purposes of this discussion, parallel processing systems include any configuration of computer systems using multiple central processing units (CPUs), either local (e.g., multiprocessor systems such as SMP computers), or locally distributed (e.g., multiple processors coupled as clusters or MPPs), or remotely, or remotely distributed (e.g., multiple processors coupled via LAN or WAN networks), or any combination thereof. Again, the graphs will be composed of components (data files or processes) and flows (graph edges or links). By explicitly or implicitly replicating elements of the graph (components and flows), it is possible to represent parallelism in a system.
Graphs also can be used to invoke computations directly. The “CO>OPERATING SYSTEM®” with Graphical Development Environment (GDE) from Ab Initio Software Corporation, Lexington, Mass. embodies such a system. Graphs made in accordance with this system provide methods for getting information into and out of individual processes represented by graph components, for moving information between the processes, and for defining a running order for the processes. This system includes algorithms that choose interprocess communication methods and algorithms that schedule process execution, and also provides for monitoring of the execution of the graph.
Developers quite often build graphs that are controlled in one way or another through the use of environment variables or command-line arguments which enable generation of instructions (e.g., shell scripts) that are translated into executable instructions by a graph compiler at “runtime” (i.e., when the graph is executed). Environment variables and command-line arguments thus become ad hoc parameters for specifying information such as file names, data select expressions, and keys (e.g., sort keys), making the applications more flexible. However, a user may have to read a generated shell script and search it for references to environment variables and command-line arguments to find the set of parameters that control the execution of a particular graph.
In a general aspect, the invention features a method, and corresponding software and system, for ordering parameters in a graph-based computation. The method includes determining a desired first ordering of a set of parameters associated with graph elements in a computation graph; determining an ordering constraint for the set of parameters; and determining a second ordering of the set of parameters that satisfies the ordering constraint according to the desired first ordering.
This aspect can include one or more of the following features:
The method further includes accepting a specification of the computation graph wherein the desired first ordering is based on an order in which the graph elements appear in the specification.
The graph imposes the ordering constraint.
The method further includes determining values of the parameters according to the second ordering.
Determining values of the parameters includes prompting for user input.
The method further includes determining a parameter value for a first of the parameters based on a user response to the prompting, and executing the graph using the determined parameter value as the value for the first parameter.
Determining the ordering constraint includes determining dependencies between values of the parameters.
The ordering constraint includes ordering a first parameter after a second parameter if the first parameter depends on the second parameter.
The ordering constraint correspond to one or more directed acyclic graphs, wherein each vertex of the one or more directed acyclic graphs corresponds to a parameter in the set.
Determining the second ordering includes performing a topological sort of the parameters in the set based on the ordering constraint.
Satisfying the ordering constraint according to the desired first ordering includes preserving a similarity to the first ordering for at least some parameters that are not required to be reordered to satisfy the ordering constraint.
Aspects of the invention can include one or more of the following advantages:
The interface of a graph in terms of runtime parameters has been formalized. The interface for a graph has been defined well enough for the system to know what parameters need to be supplied and how they should be prompted for.
The metadata that controls components can be specified or computed, directly or indirectly, by runtime parameters.
The structure of a graph can be modified based on the values of runtime parameters controlling conditional components, so that components are present or absent based on user choices.
A benefit of runtime parameterization of graphs is that an application can be parameterized richly enough to enable end users, such as business analysts and statistical modelers, to request data that meets their needs. The complexity of modern corporate data environments has led to a state of affairs in which a significant amount of direct human involvement is usually needed in the process of data collection and pre-analysis transformation. The invention provides powerful tools to end users that enables them to define and retrieve the data they want without requiring expert data analyzers in the critical path for each query type.
Metadata that is propagated within a graph can include metadata that is functionally transformed, such as metadata that is defined as a function of other metadata. The propagation can occur, for example, at edit time before the graph is run. Enabling propagation of transformed metadata can enhance a user's ability to view and/or manipulate metadata even before the graph is run.
There can be a library of reusable (inter-connectable) components with runtime parameters. A graph can be built from these components with an automatically determined prompting order for all of the runtime parameters in the graph. In some cases parameters may need to be reordered to satisfy certain constraints. Reordering parameters to satisfy those constraints according to a desired ordering (e.g., an ordering specified by a developer) reduces the chance of prompting a user for parameters in an order that deviates significantly from the desired ordering.
Other features and advantages of the invention will become apparent from the following description, and from the claims.
Like reference symbols in the various drawings indicate like elements.
The repository 104 preferably is a scalable object-oriented database system designed to support the development and execution of graph-based applications and the interchange of metadata between the graph-based applications and other systems (e.g., other operating systems). The repository 104 is a storage system for all kinds of metadata, including (but not limited to) documentation, record formats, transform functions, graphs, jobs, and monitoring information. Repositories are known in the art; see, for example, U.S. Pat. Nos. 5,930,794; 6,032,158; 6,038,558; and 6,044,374.
The parallel operating system 106 accepts the representation of a data flow graph generated in the GDE 102 and generates computer instructions that correspond to the processing logic and resources defined by the graph. The parallel operating system 106 then typically executes those instructions on a plurality of processors (which need not be homogeneous). A suitable parallel operating system is the CO>OPERATING SYSTEM® available from the assignee of the present invention.
The Web Interface 108 provides a web-browser-based view of the contents of the repository 104. Using the Web Interface 108, a user may browse objects, create new objects, alter existing objects, specify application parameters, schedule jobs, etc. The Web Interface 108 automatically creates a forms-based user interface for a parameterized graph based on information stored in the repository 104 for the graph's runtime parameters.
The executive 110 is an optional repository-based job scheduling system accessed through the Web Interface 108. The executive 110 maintains jobs and job queues as objects within the repository 104, and the Web Interface 108 provides a view of and facilities to manipulate jobs and job queues.
The flow of data represented by a “flow” in a data flow graph can be organized into discrete data elements. For example, the elements can include records from a dataset that is organized into records (or rows) and fields (or columns). Metadata describing the sequence of fields and data types corresponding to values in a record is called a “record format.”
Components and datasets in a graph have input and/or output ports for connecting to flows. The “source ends” of the flows 124 and 128 interface with an output port of the input dataset 122 and with an output port of the filter component 126, respectively. The “sink ends” of the flows 124 and 128 interface with an input port of the filter component 126 and with an input port of the output dataset 130, respectively. An input or output port of a dataset or component is associated with metadata, such as a record format for the data flowing into or out of the port.
A parameter including a record format for a port or other metadata associated with a component is bound to a value according to rules for parameter scoping. A parameter can be bound to a value at design time or at runtime (i.e., a “runtime parameter,” as described below). The value of a parameter can be defined, for example, by a user over a user interface (e.g., in response to a prompt), defined from a file, or defined in terms of another parameter in the same context or a in different context. For example, a parameter can be exported from a different context (e.g., a parameter evaluated in the context of a different component) by designating the parameter to have a “same as” relationship to another parameter.
A component used in a graph can be implemented using other components that are interconnected with flows forming a “sub-graph.” Before a sub-graph is used as a component in another graph, various characteristics of the component are defined such as the input and/or output ports of the component. In some cases, characteristics of a component having to do with relationships among sub-graph components should be specified before the component is used in a graph. For example, a prompting order for runtime parameters of sub-graph components may need to be selected. An approach for selecting a prompting order for runtime parameters of components in a graph is described in more detail below.
Metadata Propagation
The value of metadata associated with a port, such as a record format parameter, can be obtained by “propagation.” Metadata propagation can occur “externally” or “internally.” For external metadata propagation, the value of a record format parameter for a port of a first component can obtain a value by propagating a record format value for a port of a second component that is connected to the first component by a flow. The value is able to propagate either downstream from the source end to the sink end of a flow or upstream from the sink end to the source end of a flow. Metadata propagates from a port that has defined metadata to a port that does not have defined metadata.
For internal metadata propagation, metadata defined for one port of a component propagates to another port of that component based on a sub-graph that implements the component. In some cases, internal metadata propagation occurs over “non-transforming” internal data paths. For example, a user may provide metadata for the input port of a sort component that specifies the data type of records flowing into the sort component. Since the sort component re-orders but does not transform the records, the data type is not changed by the sort component and the data type propagates unchanged to the output port of the sort component accurately describing the data type of the records flowing out of the sort component.
Some components do transform (or optionally transform) data flowing through them. For example, a user may provide metadata for the input port of a filter component that specifies the fields of records flowing into the filter component. The filter component may remove values of a given field from each record. A metadata definition can be used to specify that the metadata for the output port of the filter component is related to the metadata of the input port according to the filter action of the component. For example, the filtered field may be removed from the metadata specifying the record fields. Such a metadata definition can be supplied even before the input port metadata is known. Therefore, metadata can propagate even over transforming internal data paths by allowing metadata associated with a port to be specified as a function of one or more parameters, including metadata for another port, as described in more detail below.
This internal and external metadata propagation can optionally be configured to occur at design time while a graph is being constructed and a user supplies metadata for some ports of some components in the graph. Alternatively, metadata propagation can occur after a graph is constructed, including at or just before runtime.
Runtime Parameters
A runtime parameter allows an application builder to defer the value of a parameter setting (e.g., the key parameter of a sort function, file names, record formats, transform functions, etc.) to runtime (e.g., the time a program is executed or soon to be executed on a computer system). The values of runtime parameters may be supplied by the end user or be derived from a combination of other runtime parameters or objects stored in an object repository.
Runtime parameters add a certain amount of flexibility to an application. Additional flexibility is achieved by using those parameters to compute metadata (data formats or types, and program logic or transforms) on demand. Types and transforms may be synthesized from other types and transforms, user-supplied parameter values, and stored objects (e.g., from a repository). This makes it possible to build “generic” applications that work on input data of any type, or that produce data through a series of transforms whose construction is controlled, directly or indirectly, through runtime parameter values.
In some implementations, when creating or editing a runtime parameter, a developer may specify a prompt for each parameter and the conditions for displaying the prompt. The system interprets the prompting directives to present, if conditions are met, a graphical user interface (GUI) control for receiving the parameter value.
Designation of Runtime Parameters
Runtime parameters provide a mechanism for a developer to modify the behavior of a graph based on external input at graph execution time (i.e., runtime). In the preferred embodiment, these external values are provided by direct user input. However, these external values also may come from a number of different sources, including environment variables and command line parameters. The GDE 102 generates the correct code to handle all of these situations as well as prompting the developer for test values when the graph is executed directly from the GDE. Using runtime parameters, a developer can, for example, explicitly declare that the path of an input file will be provided by an environment variable with a particular name; that environment variable then becomes a known part of the graph's interface. Thus, there is a well-defined interface to such parameters. There is no need, for example, to read a generated shell script and search it for references to environment variables and command-line arguments to find the set of parameters that control the execution of a particular graph.
A runtime parameter may be designated or defined in a number of ways. One way is by use of a runtime parameters grid displayed in the GDE 102.
When the graph is run, the parameters are processed to obtain values for each parameter from user input or from external programmatic sources (e.g., command line parameters or environmental variables). In the illustrated embodiment, the runtime parameters grid 300 includes the following fields:
Name 302—This field contains the name of the runtime parameter. “Score_threshold” is the example shown for a name.
Type 304—This field contains the type of value to be allowed in the runtime parameter. “Integer” is the example shown for a type. Supported types in the illustrated embodiment are:
Location (Loc) 306—This field is used with record format and transform types. It specifies whether the type field 304 describes a file location or whether it contains an embedded description. Supported locations are:
Default Value 308—This field contains either (1) the default value for the runtime parameter which will be used if no other value is provided from an external programmatic source, or (2) a rule or expression describing how to derive the runtime value from user input or how to obtain that information interactively from the user executing the graph. In the latter case, a second default value field (not shown) may be used to provide a value for the runtime parameter if the user does not provide an input value. For types of “boolean” and “choice”, this field limits the user to the valid choices. For “layout” types, this field is read-only and displays the currently defined layout definition. For all other types, this field preferably is a simple text editor into which the user may type a valid string.
Edit 310—Clicking on the edit space 310 (or an icon; for example, a pencil icon) in a parameter row will bring up a more advanced edit window, which walks a user through the various options for editing the default value field 308. In the illustrated embodiment, the following editors are available for their associated types:
The above editors are launched unless the Kind field value (see below) is “PL” (for Parameter Language). In this case the user is presented with an editor with which to define the rules for deriving or prompting for the parameter value at graph execution time.
Description 312—This is a free format field in which a developer describes the expected values of the runtime parameter. It is used as a prompt at runtime if the default value contains a rule for asking the user for an input value.
Kind 314—This field defines where a graph is to obtain the value for the associated parameter at graph execution time. Supported kind field 314 values are:
In the illustrated embodiment, a useful convention for referencing a runtime parameter that has been created directly on the runtime parameters grid 300 is to simply enter the parameter name preceded by the dollar sign “$”. For example, $key references a runtime variable named key. In the illustrated embodiment, new runtime parameters default to a type of “string” and a default kind based on the value in the advanced options dialog for the default runtime kind (the default runtime kind is “Environment”).
Because runtime parameter values can are determined at runtime, and PL scripts can provide conditional testing, “conditional” runtime parameters can be created. A conditional runtime parameter causes a prompt to be generated for user input only if all of the conditions for the parameter—determined at runtime—are enabling. Thus, for example, if a user responds to a first prompt requesting whether a data set is to be sorted with “NO”, a second, conditional prompt that requests a sort key need not be displayed.
Thus, during a design phase (“design time”), a developer designates a particular parameter of a graph component as a “runtime” parameter. An object associated with that graph component is then stored with the relevant parameter data (e.g., the types of information from the parameters grid 300 of
If the user does not input a value for a particular parameter (STEP 406), the default value for the parameter may be selected (STEP 408). Alternatively, an error condition may be raised to indicate the lack of user input. In any event (assuming no error condition because of a lack of user input), a determination is made of the final value for the parameter, taking into account transformations of the input and dependencies and conditions based on other parameters (STEP 410).
If a determination is made that user input is not indicated for a particular parameter (STEP 402), a determination is then made as to whether the parameter value is to be externally supplied programmatically, such as by an environment variable or a command line parameter (STEP 412). If not, the default value for the parameter is selected (STEP 414). Alternatively, an error condition may be raised to indicate the lack of available input of the specified type. In any event (assuming no error condition because of a lack of external input), a determination is made of the final value for the parameter, taking into account transformations of the input and dependencies and conditions based on other parameters (STEP 410).
Once the final parameter values are determined, as an optional step all conditional components (discussed below) can be removed either completely or replaced by flows (i.e., a graph link or edge), according to the specified conditions and the rules outlined above (STEP 416). Once the operational graph structure is finalized and the final parameter values are determined, the graph is executed in conventional fashion (STEP 418).
Test Values
In order to support a developer during the creation and testing of graphs with runtime parameters, the preferred embodiment of the GDE 102 also supports test values for runtime parameters. When a developer runs a graph with runtime parameters or wants to view the underlying code affecting a graph component, the GDE 102 displays an associated test parameters grid where the user can enter new test values for one or more runtime parameters. Preferably, the last set of test values used is remembered and saved with the graph.
For each runtime parameter, the developer enters a desired test value in a test value column. An edit field may be associated with each test value column. The test value field and edit field behave the same as the default value field and edit field in the runtime parameters grid 200 except when the parameter kind is PL.
If a PL expression indicates that the user is to be prompted for a value for a particular runtime parameter, then the test value field and the edit behavior are based on the interpretation of the associated PL expression. If the PL expression simply derives a value based on other input, then in normal mode the runtime parameter is not visible in the test values grid.
Specifying How Runtime Parameters Get Their Values
After a parameter has been designated as a runtime parameter, a corresponding object is created in the repository 104. If the runtime parameter has a kind field 214 value of “PL”, the default value field 308 for the parameter includes a prompt_for pseudo-function with the following preferred form:
As indicated above, the prompt_for pseudo-function may be part of a conditional expression that determines whether a prompt is to be displayed based on prior input.
For such objects, a user interface is used to present direct entry runtime parameters to a user. In the preferred embodiment, the Web Interface 108 provides this function. In particular, during runtime, each prompt_for pseudo-function of each—runtime parameter object is parsed by the Web Interface 108 to generate a web page (e.g., in HTML) having a corresponding user prompt. (Alternatively, such web pages can be generated before runtime and simply presented at runtime. However, runtime generation of such web pages provides greater flexibility. In particular, the contents of a page can depend on prior user input.) The Web Interface 108 is used in conjunction with a conventional web browser that can display such web pages and receive user input.
The prompt_for pseudo-function indicates to the Web Interface 108 how to prompt for a parameter value. In particular, the prompt-kind parameter, a string constant, indicates what kind of user interface (UI) element to present (text box, dropdown list, etc.). The modifiers part of the string, a comma-separated list of keywords, provides some options common for various kinds of prompts. In the illustrated embodiment, space is not significant within the modifiers string. Modifier keywords are interpreted as follows:
Following are some examples of prompt_for calls with different kinds of modifiers:
The remainder of this section lists a variety of prompt-kinds and their corresponding options and explains how each would appear in a web page generated by the Web Interface 108.
text [size]—Presents a conventional single-line text box size characters wide (if size is not supplied it defaults to the browser's default size for text boxes).
radio choice-list [description-list]—Presents a conventional “choose one” prompt in the form of a set of radio buttons, one button for each element of the choice-list. If description-list is supplied, each choice is labeled with the corresponding description; otherwise, the choices are labeled with the string form of the corresponding item from the choice-list.
radioplus choice-list [description-list]—Like radio, but presents an additional button next to a text box, to allow a user to choose a “write-in” value not in the choice-list.
checkbox choice-list [description-list]—Presents a conventional “choose zero or more” prompt in the form of a set of check boxes, one button for each element of the choice-list. If description-list is supplied, each choice is labeled with the corresponding description; otherwise, the choices are labeled with the string form of the corresponding item from the choice-list.
dropdown choice-list [description-list, size]—Presents a conventional “choose one” prompt in the form of a dropdown list for the elements of the choice-list. If description-list is supplied, each choice is labeled with the corresponding description; otherwise, the choices are labeled with the string form of the corresponding item from the choice-list. If size is supplied, that many choices will be visible at once; otherwise, only one will be visible.
multidropdown choice-list [description-list, size]—Presents a conventional “choose zero or more” prompt in the form of a dropdown list for the elements of the choice-list. If description-list is supplied, each choice is labeled with the corresponding description; otherwise, the choices are labeled with the string form of the corresponding item from the choice-list. If size is supplied, that many choices will be visible at once; otherwise, the browser's default number of items is shown.
key type-obj [size]—Presents a prompt for a key (also known as a collator) made up of fields from the given type-obj. The key can have as many as size parts, which defaults to the number of fields in type-obj.
filter type-obj—Presents a prompt for a filter expression made up of conditions on each field of the given type-obj. The blank ok modifier has no effect for filters; a blank filter yields a “True” expression.
flexifilter type-obj—Similar to the filter prompt, but presents a prompt for a filter expression made up of conditions on each field of the given type-obj where the field name on each line is selectable from a dropdown list. This permits using the same field for multiple conditions (e.g., field STATE=MA OR field STATE=CA).
rollup type-obj key [size]—Presents a prompt for a rollup computation based on the fields of the given type-obj being rolled up by the given key. The rollup can have as many as size rules, which defaults to the number of fields in type-obj. The blank ok modifier has no effect for rollups; a blank rollup yields a package that provides just the key value for each group.
reformat type-obj [size]—Presents a prompt for a reformat computation based on the fields of the given type-obj. The reformat can have as many as size rules, which defaults to the number of fields in type-obj.
outputspec—Presents a prompt for an output dataset specification. The displayed control includes a dropdown control for presenting available format options, and a text edit box for entering the name of a specific instance of the output dataset. The blank ok modifier has no effect for output dataset specifications.
fpath starting-point—Presents a prompt for a file path. The prompt is essentially a text box, but has a “Browse” button next to it that will cause a popup window to appear for browsing for a file path. If the text box is non-blank, then it will be used as the starting point for the browsing operation; if it is blank, the starting-point argument is used.
rpath starting-point—Presents a prompt for a repository path. The prompt is essentially a text box, but has a “Browse” button next to it that will cause a popup window to appear for browsing. If the text box is non-blank, then it will be used as the starting point for the browsing operation; if it is blank, the starting-point argument is used.
radio fpath choice-list [description-list]—Like radioplus, but presents an fpath-style box-plus-browse-button in the “write-in” slot.
radiorpath choice-list [description-list]—Like radioplus, but presents an rpath-style box-plus-browse-button in the “write-in” slot.
Conditional Components
Some implementations include a conditional components mechanism that permits changes to the structure of the components and flows of a graph based on parameter values and computed metadata. Each component of a graph has a condition which controls whether or not that component will appear in the graph at runtime. The condition can be computed directly or indirectly through runtime parameters. Conditional components can be used for a variety of purposes, such as to optimize or specialize graphs. For optimization, an application might omit processing of certain datasets if values from them will not be used, thus allowing the graph to run more efficiently. For specialization, an application might condition the production of several different output datasets based on the level of detail desired, or allow execution of one of several optional portions of a graph.
In the illustrated embodiment, a conditional component can be any graph component that defines a vertex (i.e., a dataset component such as an input/output file, a processing component such as a reformat or sort component, or other graphs, known as subgraphs). In the preferred embodiment, a conditional component is controlled by two special parameters: a Condition and a Condition-interpretation. A Condition is a boolean expression or value whose evaluation is deferred until runtime. In the illustrated embodiment, the values “false” and “0” specify a false condition, all other values (including empty) indicate a true condition. A Condition-interpretation parameter has two allowed mutually exclusive values: Remove Completely and Replace With Flow.
Remove Completely: With this interpretation, if the Condition is met, the component and all of its connected flows (i.e., graph links or edges) are to be removed from the graph. An active Remove Completely condition functionally removes the component and all its directly connected flows from a graph. Remove Completely conditions can be used on any component.
A conditional component that is removed from a graph can “poison” other connected components that depend on the presence of the conditional component, causing their removal.
In one implementation, the detailed semantics of poisoning (also known as “implied conditions”) are as follows:
There are some surprising consequences of these rules. For example, a component with only optional ports can never be removed because of poisoning. Therefore, it must be explicitly removed if desired.
Replace With Flow: With this interpretation, if the Condition is met, the component is to be replaced with a flow (i.e., a graph edge). A Replace With Flow condition-interpretation needs additional information. Referring to
Metadata for a graph can be supplied, for example, by a graph developer, by a graph user, or by propagation from another portion of the graph. Various kinds of metadata can be propagated, including metadata associated with the data or computations on the data such as: a record format for a port (e.g., sequence of fields and data types of records flowing into or out of a port), sortedness, compression method, character set, binary representation (big-endian, little-endian), partitioning, what computing resources (e.g., processor, temporary disk space) the component may use, data transformations, and amounts of memory the component may use. Various aspects of graph construction can affect the propagation of metadata. Two of these aspects are described below.
Propagation After Component Removal
In some implementations, when a flow is generated after the removal of a graph component, a choice must be made as to how metadata defining the data in such flow should propagate in the revised graph. Metadata may be available from either end of the flow. In some implementations, the metadata from the upstream end of the flow is preferred.
If the upstream end of the flow is a removed component (or a component that has been replaced by a gather component), then the GDE 102 finds metadata for the flow by “walking” upstream in the graph until it finds a component that has not been removed. The metadata exposed by that upstream component is used to define the characteristics of the data for the generated flow.
Propagation of Transformed Metadata
As described above, metadata can propagate even over transforming internal data paths by allowing metadata associated with a port to be specified as a function of one or more parameters, including metadata for another port. For example,
To enable propagation of transformed metadata through this “smart join” component, a graph developer defines the metadata (e.g., metadata for describing the fields) for the output port 2316 of the “smart join” component 2306 as a function of metadata for the first input port 2310 input0.metadata, metadata for the second input port 2314 input1.metadata, and the key field parameter key_field:
The output port metadata is determined by binding the function arguments to values (with respect to the appropriate context) and performing the function metadata_join on the results. In this example, since metadata for the ports 2310 and 2314 are undefined, propagated metadata are bound to the metadata parameters input0.metadata and input1.metadata. A user supplies metadata for the output port 2308 that specifies fields “A” and “B” for records flowing from port 2308 to input port 2310 of the “smart join” component 2306. The user also supplies metadata for the output port 2312 that specifies fields “A” and “C” for records flowing from port 2312 to input port 2314 of the “smart join” component 2306. This user-supplied metadata propagates to the ports 2310 and 2314. The key field for the join operation is field A, so the “formal parameter” key_field is bound to the value “A.”
The function metadata_join determines the output metadata by first determining whether the value of the key_field parameter is a member of both sets of fields specified by input0.metadata and input1.metadata. If so, the output metadata is the union of the two sets of fields. If not, the output metadata indicates an empty set of fields.
After the metadata propagates to the input ports of the “smart join” component 2306 (or is otherwise supplied, for example, by a user), the transformed metadata for the output port of the “smart join” component 2306 includes fields A, B and C. This transformed metadata can then be propagated to other components. In this example, the transformed metadata propagates to the filter component 2318.
Metadata, whether supplied by a user or propagated between ports, can be displayed to the user. For example, the user can use an input device (e.g., a mouse) to select a portion of a component for which to view metadata values. The metadata propagation can also be triggered in response to such a user selection.
Exemplary Metadata Propagation Process
The process 2400 starts at the beginning of the worklist and, for each component in the worklist, the process 2400 propagates metadata internally 2404 within the component (e.g., from an input port to an output port, or from an output port to an input port) based on a specification of the sub-graph implementing the component (e.g., an data flow in the sub-graph). This internal metadata propagation includes transferring metadata untransformed between ports on either end of an non-transforming data path. Internal metadata propagation also includes deriving metadata for a port that has a metadata definition that refers to parameters of the graph and/or metadata for other port(s), as described above. When the process 2400 encounters such a metadata definition, the process 2400 evaluates any parameters whose values are needed to derive the metadata.
After performing internal metadata propagation for a component on the worklist, the process 2400 propagates metadata externally 2406 from each port of the component that has metadata to a port of a related component that does not have metadata. Any component that acquires metadata by this external propagation is moved 2408 to the end of the worklist. The process 2400 terminates 2410 after the last component on the worklist is processed.
One type of relationship between components that supports this type of external metadata propagation is a data flow link between ports of two components (e.g., from an input port to an output port, or from an output port to an input port).
Another type of relationship between components that supports this type of external metadata propagation is a link indicating that metadata for one port may also be used for another port. This type of “metadata link” does not necessarily correspond to a data flow link. For example, a port can have a metadata link to metadata in a graph that is not associated with any port in particular.
Runtime Parameters in Componentized Sub-Graphs
Before a sub-graph is “componentized” to be used as a component in another graph, various characteristics of the component are defined, such as the input and/or output ports of the component. For a sub-graph that includes components with runtime parameters, a prompting order for the runtime parameters should be selected. Since components in a graph are not necessarily sequentially ordered, there can be multiple possible global orderings of the runtime parameters for prompting a user. Some of the global orderings are not as consistent with the original orderings associated with each component. It is useful to generate a global ordering for prompting that preserves as much as possible the orderings of the parameters in each component, while reordering when appropriate to take dependencies into account. For example, a component may order a prompt asking “what data would you to process?” before a prompt asking “where would you like to store the processed data?” Even though it may be possible to provide the prompts in either order, it may be desirable to provide the prompts in this order.
Since it may be necessary to evaluate non-prompted runtime parameters in the process of evaluating prompted runtime parameters, the prompting order is obtained from an evaluation order for all of the runtime parameters. One approach for determining an evaluation order for the runtime parameters of a graph (including parameters for the graph that are not associated with any component) includes performing a topological sort based on one or more directed acyclic graphs representing dependencies among the parameters. However, some topological sort algorithms may reorder parameters unnecessarily, resulting in an undesirable prompting order for runtime parameters.
In a first example, a parameter sorting process provides an initial list of parameters for parameters of two graph components: Component I, and Component II connected to Component I. In this example, the parameters have only “intra-component” dependencies. That is, parameters of a component depend only on other parameters in the same component. The parameters are defined as follows.
Component I includes the following parameters:
Component II includes the following parameters:
The order in which the parameters are listed define a desired order in which to prompt a user for values. The initial list of parameters maintains this “initial ordering” for each component. An “ordinal” is assigned to each parameter to indicate that parameter's place in the initial ordering. The following table lists the parameters in this initial ordering.
The “dependencies” column indicates other parameters on which the listed parameter depends. The dependencies impose an ordering constraint on the evaluation of the parameters: a parameter needs to be defined before it is used (e.g., referenced) by another parameter.
A “common topological sort” algorithm passes through the list transferring parameters with zero dependencies into an ordered output list on each pass. After each pass, any transferred parameters are removed from the dependencies column. This process repeats until all parameters have been transferred. The order of parameters in the output list represents the “final ordering” such that parameters that depend on other parameters are evaluated after those other parameters have been evaluated.
In this example, on the first pass, the parameters x, q, a and c are transferred into the output list. On the second pass, the parameters y and b are transferred into the output list. On the third and final pass, parameter z is transferred into the output list. Thus, the final ordering for the parameters is: x, q, a, c, y, b, z. While this ordering does satisfy the ordering constraint imposed by the parameter dependencies, it unnecessarily reorders the parameters. In this example, the initial ordering also satisfies the ordering constraint imposed by the parameter dependencies.
Other approaches for determining an evaluation order for the parameters of a graph that satisfies the ordering constraint do respect the initial ordering. For example, some approaches order the parameters to satisfy the ordering constraint, choosing the ordering according to a criterion based on the initial ordering. The criterion can include any of a variety of criteria that give preference to keeping the order close to the initial ordering (e.g., minimize a metric based on changes to the initial ordering). In some cases, there may not be a unique “best” ordering, since multiple orderings may satisfy a given criterion equally well according to the criterion.
An example of an approach that respects the initial ordering is a “modified topological sort” approach. In this approach, the criterion based on the initial ordering is to minimize the number of parameters that are transferred from the initial list before a preceding parameter that does not depend on any untransferred parameter is transferred. In other words, the “modified topological sort” removes a transferred parameter from the dependencies column before transferring the next parameter with zero dependencies. For the example above, the “modified topological sort” approach generates a final ordering that is the same as the initial ordering: x, y, z, q, a, b, c.
Modified Topological Sort Process Respecting Initial Ordering
Pseudocode is given below for two exemplary “modified topological sort” processes that both respect initial ordering as determined by an assigned ordinal for each parameter. The second process includes an optimization to improve time efficiency for some cases. The processes manipulate data structures generated from input data for the parameters.
Assuming there are N parameters to be ordered, the input data includes a list of N triples consisting of a unique parameter name, a set of parameters upon which the named parameter depends (called a “dependency set”) and an optional attribute data object storing information related to the named parameter.
Associated with this input data are one or more directed acyclic graphs that represent the dependencies among the parameters, called “dependency graphs.” Each unique parameter name corresponds to a node in a dependency graph, and the associated dependency set corresponds to a set of links from other nodes to that node. So a link points from a first node for a first parameter to a second node for a second parameter that depends on the first parameter. Alternatively, the correspondence between the link direction and parameter dependency could be reversed.
An output data structure result_list includes a list of the N parameters from the input data reordered (if necessary) so that a parameter is evaluated before it is used for evaluating another parameter while giving preference to keeping the order close to the initial ordering. To generate the output data structure result_list, the processes “eliminate” parameters by transferring parameters one at a time from a working data structure param_list to the output data structure result_list. The output data structure is complete after all parameters have been eliminated.
A first “modified topological sort” process includes two phases. In the first phase, the process builds working data structures based on the input data for use in generating the sorted output data structure. In the second phase, the process iteratively sorts and eliminates parameters according to the dependency constraint represented by these working data structures.
Some of the working data structures that the process builds in the first phase are dictionaries, which are data structures based on hashing. Items in dictionaries can be accessed effectively in O(log N) time. The following exemplary data structures are built in the first phase:
parm_list[index]: an ordered list of non-eliminated parameter names, indexed by a number index (where index=0 corresponds to the first item in the list). This data structure is “dynamic” (i.e., changes during the execution of the process). The list is indexed by position, such that if an item is removed from the middle of the list, then the index of items after the removed item are shifted accordingly.
n_dependencies_dict[name]: a dictionary keyed by a parameter name (name), whose entries contain the number of parameters on which the keyed parameter depends. This dictionary is dynamic.
dependers_dict[name]: a dictionary keyed by a parameter name (name), whose entries are dictionaries (also keyed by parameter name), representing the set of parameters that depend on the keyed parameter. This dictionary is “static” (i.e., does not change during execution of the process).
order_dict[name]: a dictionary keyed by a parameter name (name), storing the ordinal position, an integer ranging from 0 to N-1, of the parameter in the initial ordering. This dictionary is static.
attribute_dict[name]: a dictionary keyed by a parameter name (name), storing the optional attribute data object for the keyed parameter. This dictionary is static.
result_list[index]: an ordered list of parameter names and attributes representing the output of the process, indexed by a number index (where index=0 corresponds to the first item in the list). This data structure is initially empty. This data structure is dynamic.
For the purposes of analyzing the time efficiency of the processes, the average “degree” (or number of links from a node) of the dependency graphs is assumed to be z. Building these data structures take O(N) time, except for n_dependencies_dict and dependers_dict, which take O(N*z) time.
In the second phase, the process sorts the parameters in the param_list data structure according to a sort criterion by_n_deps_and_order that orders parameters first by the number of non-eliminated parameters on which they depend (i.e., by their value of n_dependencies_dict), from lowest to highest, and then by their ordinal (i.e., by their value of order_dict), from lowest to highest. The process then eliminates the first parameter in the sorted param_list. The value of n_dependencies_dict for this parameter should be zero. (If the value of n_dependencies_dict for the first parameter in the sorted param_list is not zero, then an error is flagged.)
To eliminate a parameter, the process appends it to result_list (along with any corresponding attributes) and decrements the dependency count (i.e., the value of n_dependencies_dict) of all of its dependers (i.e., parameters in dependers_dict) by one. Finally, the parameter is deleted from parm_list. This sorting and eliminating of the resulting first parameter is repeated until all parameters have been eliminated.
The following is a pseudocode definition for an eliminate procedure:
The arguments of the eliminate procedure are list (whose value is, e.g., param_list) and index. The function result_list.append appends the indicated list item at position index along with its associated attribute to result_list. Then, the procedure decrements the value of n_dependencies_dict for each parameter depender that is a member of the dependers_dict data structure, keyed on the parameter being eliminated. Then, the procedure deletes the parameter from list. The run time for the eliminate procedure is O(z log N).
The following is pseudocode for a sort/eliminate loop for the first “modified topological sort” process:
The process first performs an initial sorting of param_list using the function parm_list.sort(by_n_deps_and_order) that orders parameters of param_list according to the sort criterion by_n_deps_and_order described above. The process then performs the eliminate procedure followed by another sorting of param_list until param_list is empty. The process checks to make sure that the number of dependencies for the first parameter (with index=0) in param_list is zero. If not, the process removes the parameter, records a circularity error, and continues. The sort takes O(N log N) and the loop range is N, so the estimate for the overall run time for the loop is O(N2 log N).
A second “modified topological sort” process takes advantage of the cases in which the dependency graphs are sparse, such that z<<N. After one initial sort, the process can maintain the sortedness of a list candidates of parameters that do not depend on any other parameters. This reduces this expected run time as described below.
The following is pseudocode for the second “modified topological sort” process:
The process first performs an initial sorting of param_list using the function parm_list.sort(by_n_deps_and_order) that orders parameters of param_list according to the sort criterion by_n_deps_and_order described above. The process then performs a loop having three sections (labeled “# section 1,” “# section 2,” and “# section 3”).
In section 1, the process builds a candidates list that contains only parameters with zero dependencies. The process scans all of the parameters in parm_list and appends them to candidates, preserving their relative ordering.
In section 2, the process performs a loop in which parameters from candidates are eliminated and new parameters are merged into candidates. The first parameter in candidates, saved as this_parm, is eliminated from candidates and deleted from param_list. A function get_new(this_parm) returns a list of names of parameters that are members of dependers_dict for the newly eliminated this_parm and have zero dependencies left. These parameters, representing parameters that have had their last dependency removed, are then sorted according to by_n_deps_and_order (to ensure they ordered according to their respective ordinals) and merged into candidates. Thus, the candidates list remains a list of zero-dependency parameters sorted by ordinal.
Section 3 is only entered if there is a “circularity error” caused, for example, when two parameters are defined in terms of each other. In this case, the process sorts parm_list again, and if the first parameter in parm_list has nonzero dependencies it is deleted and the loop repeats with section 1.
Assuming there are no circularity errors, the N-parameter list parm_list is sorted only at the beginning, resulting in a sorting time of O(N log N). Thereafter, sorting only occurs on the much smaller list of newly generated zero-dependency parameters resulting from eliminating the parameter at the head of the candidates list. The size of this list is less than z (on average), resulting in a sorting time of O(z log z) and a merging time of O(z). Thus, one iteration of the loop is O(z log z) and the overall time is O(Nz log z+N log N). For the cases in which z does not grow with increasing N, this time is effectively O(N log N).
In another example, a parameter sorting process (e.g., the first or second “modified topological sort” process) determines an initial list of runtime parameters for a graph 2500 having graph components 2502, 2504 and 2506, as shown in
Dependencies are indicated in
The parameter sorting process assigns an ordinal to each of ten parameters p0, p2, . . . , p9 for various graph elements according to the order of placement of the elements in the graph 2500. In
The following listings of the parameters in param_list and result_list at various stages of processing correspond to the first “modified topological sort” process described above. The param_list is shown sorted according to the sort criterion by_n_deps_and_order at each stage.
p9
The following listings of the parameters in candidates and result_list at various stages of processing correspond to the second “modified topological sort” process described above. It is not necessary to sort candidates between stages since the parameters remain in the same order at each stage.
Thus, referring to
Typical Usage
Typically, a user sits in front of the Web Interface 108 and finds in the repository 104 the graph of an application the user would like to run. By scanning all of the objects associated with the application graph, the Web Interface 108 generates web page forms that allow the user to specify values for the runtime parameters of the application. Once all runtime parameters have been specified, the combination of the application and the parameter settings are brought together as a job, which is scheduled for execution by the executive 110. When it comes time to run the job, the executive 110 queues the application for execution under the parallel operating system 106, in known fashion. The parallel operating system 106 collects tracking information and job status and stores this information in the repository 104 so that users and administrators can track the progress and performance of jobs.
Script Implementation
While the GDE 102 facilitates construction of parameterized graphs, sometimes there are non-graph programs for which one would like to provide a forms-based interface. Using application-level PL and the repository 104, one can parameterize arbitrary shell scripts. For example, the description of an application can be written to a file with a structure similar to the following:
General Computer Implementation
The invention may be implemented in hardware or software, or a combination of both (e.g., programmable logic arrays). Unless otherwise specified, the algorithms included as part of the invention are not inherently related to any particular computer or other apparatus. In particular, various general purpose machines may be used with programs written in accordance with the teachings herein, or it may be more convenient to construct more specialized apparatus to perform the required method steps. However, preferably, the invention is implemented in one or more computer programs executing on one or more programmable computer systems each comprising at least one processor, at least one data storage system (including volatile and non-volatile memory and/or storage elements), at least one input device or port, and at least one output device or port. The program code is executed on the processors to perform the functions described herein.
Each such program may be implemented in any desired computer language (including machine, assembly, or high level procedural, logical, or object oriented programming languages) to communicate with a computer system. In any case, the language may be a compiled or interpreted language.
Each such computer program is preferably stored on a computer-readable medium such as a storage media or device (e.g., solid state, magnetic, or optical media) readable by a general or special purpose programmable computer, for configuring and operating the computer when the storage media or device is read by the computer system to perform the procedures described herein. The inventive system may also be considered to be implemented as a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer system to operate in a specific and predefined manner to perform the functions described herein.
A number of embodiments of the invention have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the invention. For example, a number of the function steps described above may be performed in a different order without substantially affecting overall processing. For example, STEPS 402 and 412 in
| Number | Name | Date | Kind |
|---|---|---|---|
| 5701400 | Amado | Dec 1997 | A |
| 5966072 | Stanfill et al. | Oct 1999 | A |
| 6006242 | Poole et al. | Dec 1999 | A |
| 6088716 | Stanfill et al. | Jul 2000 | A |
| 6173276 | Kant et al. | Jan 2001 | B1 |
| 6208345 | Sheard et al. | Mar 2001 | B1 |
| 6400996 | Hoffberg et al. | Jun 2002 | B1 |
| 6538651 | Hayman et al. | Mar 2003 | B1 |
| 6816825 | Ashar et al. | Nov 2004 | B1 |
| 7417645 | Beda et al. | Aug 2008 | B2 |
| 20040041838 | Adusumilli et al. | Mar 2004 | A1 |
| 20040111469 | Manion et al. | Jun 2004 | A1 |
| 20040205726 | Chedgey et al. | Oct 2004 | A1 |
| 20050034112 | Stanfill | Feb 2005 | A1 |
| 20050039176 | Fournie | Feb 2005 | A1 |
| 20050059046 | Labrenz et al. | Mar 2005 | A1 |
| 20050262470 | Gavrilov | Nov 2005 | A1 |
| 20050289527 | Illowsky et al. | Dec 2005 | A1 |
| 20060206872 | Krishnaswamy | Sep 2006 | A1 |
| 20060294459 | Davis et al. | Dec 2006 | A1 |
| 20070035543 | David et al. | Feb 2007 | A1 |
| 20070139441 | Lucas et al. | Jun 2007 | A1 |
| 20070150429 | Huelsman et al. | Jun 2007 | A1 |
| 20070285440 | MacInnis et al. | Dec 2007 | A1 |
| Number | Date | Country |
|---|---|---|
| WO 0211344 | Feb 2002 | WO |
| Number | Date | Country | |
|---|---|---|---|
| 20070011668 A1 | Jan 2007 | US |
