JPH09218861A - Scheduler - Google Patents

Abstract

PROBLEM TO BE SOLVED: To shorten the idle time of waiting for another job while keeping the intermediate generation data of a processing in performing the scheduling of parallel processings. SOLUTION: This scheduler is provided with a task division means 1 for dividing a task to be division-processed into plural sub tasks provided with task groups provided with mutual dependency, a load prediction means 2 for predicting the processing time of the respective divided sub tasks, a dependency storage means 3 for holding the relation of processing dependency among the respective sub tasks, a scheduling means 4 for deciding the processing schedule of a sub task group and an idle time adjustment means 5 for optimizing the decided schedule. For the sub task scheduled so as to generate an idle state after execution by an arithmetic processing unit 6, the idle time adjustment means 5 changes the start time to the time for which the processing time predicted by the load prediction means 2 is subtracted from the dissolving time of the idle state and exchanges post-processing idle time with the pre- processing idle time of a non-load state.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a scheduler for performing job scheduling of network distributed processing and parallel processing in a multiprocessor system, and more particularly to executing a drawing instruction string such as a page description language in a plurality of processing devices to achieve high speed processing. The present invention relates to a scheduler suitable for application to an image forming system.

[0002]

2. Description of the Related Art The conventional job scheduling policy in distributed processing or parallel processing is to keep the deadline for executing each job, or to minimize the time required to process all jobs. Based on such criteria, it is known that there is no efficient algorithm for finding an optimal solution in the so-called general job shop problem, which processes M jobs by N machines, but various optimal jobs are known. A heuristic method using the optimization method has been proposed. There, by operations such as job replacement,
It is intended to move the job execution start time in each processing machine as early as possible.

On the other hand, one typical processing system that requires job scheduling as described above is a print information processing system. In a printing device that has become a server on a computer network and accepts print requests from many users, in order to process a large number of print jobs at high speed, multiple processors are connected by a bus,
A print processing device that performs interpretation and execution of print information in parallel (for example, see Japanese Patent Laid-Open No. 4-128068), or
A technique such as a print processing system (see, for example, PCT / JP91 / 00456) that connects a plurality of processors via a network and interprets and executes print information in parallel has been proposed.

[0004]

However, in the above-described parallel processing scheduling, there may be a case where not only the idle time waiting for processing but also the idle time after processing becomes a problem. Even if a certain first job is completed, the amount of intermediate products resulting from the processing of the first job is enormous, and the second job that requires it is started and the second job of the second job is started. This is the case when it is considered to be a cost to hold the work until it is handed over to the execution machine.

This situation is particularly remarkable in the print processing system. For example, in the technique disclosed in Japanese Patent Application Laid-Open No. 7-104987, it is necessary to execute jobs having processing order dependencies with each other. Therefore, a job waiting for the start of another job in a certain processor is Only during the waiting time for synchronization, it is necessary to hold the partial image that is the processing result of the job or write it out on the disk. That is, more generally speaking, in a job with a large intermediate product of the process, the large intermediate data file is left in the storage area for data retention, so that the resource starvation state continues for a long time, Even if the processor has completed the job, it makes other jobs unavailable or difficult to use. Such a wasteful or inutilizable idle period of the processor becomes noticeable when there is a job that is executed too early even though the result of the job schedule is not a prerequisite for other jobs.

In principle, in order to reduce the idle period after such a job is completed, the first job is moved backward in time, or implemented in the processor in which the first job is processed. It is necessary to move the second job, which is the job to be performed and which is located temporally behind the first job, to the temporal front. However, such a move operation may delay the progress of the entire schedule or reverse the context of the jobs that make up the schedule, so it is dangerous to perform it unconditionally. There was a problem.

The present invention has been made in view of the above circumstances, and provides a scheduler for effectively utilizing processing resources by shortening idle time such as waiting for another job while holding intermediately generated data of processing. The purpose is to do.

[0008]

FIG. 1 is a block diagram showing the principle configuration of a scheduler according to the present invention. In the figure, the scheduler according to the present invention divides one task and schedules the divided tasks to be processed in parallel by a plurality of arithmetic processing units 6, and the tasks to be divided have a mutual dependency. Task dividing means 1 for dividing into a plurality of subtasks including a task group,
Load predicting means 2 for predicting the processing time of each subtask,
The dependency storage unit 3 that holds the processing dependency relationship between each subtask, the scheduling unit 4 that determines the processing schedule of the subtask group, and the schedule generated by this scheduling unit 4 are executed by each arithmetic processing unit 6. When the idle time occurs after the execution of the subtask, the idle time adjusting means for changing the subtask to start the processing at a time obtained by subtracting the processing time predicted by the load predicting means 2 from the end time of the idle time after the execution of the subtask. It is composed of 5 and.

In the above scheduler, the scheduling means 4 is provided with a task information table indicating whether or not each subtask to be scheduled is an execution start condition of another subtask, and the idle time adjusting means 5 is executed in the task information table. The start time of the first subtask, which is the start condition, is predicted by the load prediction means from the predicted start time of the corresponding second subtask.
Processing time of the first subtask that is scheduled by the processing time of the second subtask is subtracted, and the start time of the first subtask that is not the execution start condition is predicted by the load prediction means 2 from the predicted completion time of the final subtask. Is configured to be scheduled at a reduced time.

According to the scheduler having the above structure, first, the input task is divided into a plurality of subtasks by the task dividing means 1. The task dividing unit 1 analyzes the dependency between subtasks at the time of division, and stores the detected processing dependency in the dependency storage unit 3. The processing time of the divided subtasks is estimated and evaluated by the load predicting means 2, and the subtasks are held as additional information in the subtasks or held in separate storage means such as the task information table described above. Further, the scheduling unit 4 schedules a set of all subtasks as a processing plan on the plurality of arithmetic processing devices 6. At this time,
The dependency relationship information in the dependency storage means 3 and the predicted time value evaluated by the additional prediction means 2 are referred to for planning. Upon receipt of the schedule created here, the idle time adjusting means 5 schedules the processing start time of the subtask having the idle time after processing to be delayed. As a result, the processing device 6 that processes a subtask that has idle time after processing does not perform any processing during the idle time that processes the subtask first and waits for the next subtask. It is exchanged for idle time before processing without using resources. The subtasks scheduled by the scheduling means 4 and the schedule optimized by the idle time adjusting means 5 are distributedly processed by the respective arithmetic processing means 6,
The master arithmetic processing unit collects the processed subtask.

Accordingly, in a scheduler that divides a task and schedules it to be processed in parallel by a plurality of arithmetic processing devices, when idle time occurs after processing subtasks executed in the arithmetic processing device,
By shifting the processing start time of that subtask backward in time, the idle time waiting for the processing of the next subtask while completing the processing of that subtask and holding the intermediately generated data is It will be exchanged for idle time with no load, and the processing resources will be effectively utilized.

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION A case where the embodiment of the scheduler of the present invention is applied to a print processing system in which a drawing instruction sequence such as a page description language is executed by a plurality of processing devices to form an image at high speed will be described below. And explain.

FIG. 2 is a diagram showing a configuration example of a print processing system. The illustrated print processing system includes a plurality of client computers 1 that create print data and output print requests.
, 12, the printer 20, and the Ethernet 30, which is one of the local area networks. In addition, the printing device 20 includes the data receiving unit 2
1, a job management unit 22, a plurality of rasterization processing units 23a, 23b, 23c, 23d, a page memory 24,
And a job management unit 22,
Rasterization processing units 23a, 23b, 23c, 23d,
The page memory 24 and the printing device 25 are connected by a shared bus 26.

A printing device 20 arranged on the network
Receives print request data from the client computers 11, 12, ... Via the Ethernet 30. The print data is spooled by the data receiving unit 21 and sent to the job management unit 22. Job management unit 22
, Divides the print job issued from the client computer 11, 12, ... into a plurality of jobs and distributes them to the plurality of rasterization processing units 23a, 23b, 23c, 23d. Each rasterization processing unit 23a, 23b, 23c, 2
The job processed in 3d is transferred to the page memory 24,
When the final job is completed, it is transferred to the printer 25. The printing device 25 is a color using a laser scanning electrophotographic method capable of outputting a full-color image by repeating exposure, development, and transfer for each color of CMYK (cyan, magenta, yellow, black), for example. It is a page printer. Its performance is, for example, recording size A
3, resolution 400 dpi (dot per inc)
h), each color is 8 bits. Therefore, the page memory 24 has 128 MB (Mega Byte).

Next, a detailed configuration of the job management unit 22 to which the scheduler according to the present invention is applied will be described. FIG. 3 is a block diagram showing a configuration example of the job management unit.

The job management unit 22 divides a single print job into a plurality of job groups, and a load prediction unit 222 that evaluates the processing time of each divided job.
And a dependency storage table 223 for storing information regarding processing order constraints between the divided jobs, and the rasterized processing units 23a, 23b, 23 for the divided job groups.
a scheduling unit 224 that creates a schedule for distribution to c and 23d, and a scheduling unit 224
The schedule improving unit 225 performs an optimization process mainly for adjusting the idle time with respect to the schedule output from the above.

The job dividing section 221 includes a data receiving section 21.
A print job, that is, a drawing instruction sequence described in the page description language is received from the plurality of rasterizing processing units 23.
It is a processing module that divides into a plurality of job (drawing instruction sequence) groups for processing by a, 23b, 23c, and 23d, and analyzes the processing dependency between the divided instruction sequences at the same time. This partitioning is performed by analyzing the data dependencies using the general approach of parallel compilers.
This division method does not affect the configuration of the present invention as long as it generates a plurality of processing elements that include order dependency with each other.

Next, Po which is one of page description languages
stScript (Postscript, Adobe S
An example of simple division of the graphics drawing command sequence described by the Y. Systems Inc.) is shown below.

FIG. 4 is a diagram showing an example of division of a print job. In the figure, reference numeral 40 denotes an original print job, and here, six drawing elements are included. Six drawing elements included in the print job 40
Is divided into four jobs 41 to 44, and a setgray instruction which is a drawing attribute instruction is copied to each.
As can be seen from the overlapping situation described below, the execution of the job 41 must precede the execution of the jobs 42 and 43. In the case of a graphics command, job division is performed by dividing the command in charge of actual drawing in this way and copying the drawing attribute command to each. In addition, although the example of the graphics command was given here,
Needless to say, the jobs to be divided may be bitmap operations or character drawings as long as they can detect the overlap.

FIG. 5 is a diagram showing an image formed when a divided instruction sequence is executed. As can be seen from this formed image, the element, the element, and the element are drawn on the element. So the element is the element,
Must be formed before drawing elements and elements. The job dividing unit 221 simultaneously analyzes such a dependency while dividing the instruction sequence.

The dependency analysis can be carried out most simply by comparing the minimum circumscribed rectangles of the drawing elements.
The minimum circumscribed rectangle of a drawing element is a list of its constituent points for each kind of drawing element {(x1, y1), ..., (xn, y
n)} to obtain (X coordinate minimum value, Y coordinate minimum value) and (X coordinate maximum value, Y coordinate maximum value). These minimum circumscribed rectangles are rectangles indicated by broken lines in the example of FIG.

To define the constituent point list for each type of drawing element, for example, a point sequence may be extracted according to the following rules. Straight line: 2 endpoints Curve: 2 endpoints and control points (1 point or 2 points) Circle: (center point X value-radius, center point Y value-radius), (center point X value + radius, center point Y value + radius) ) Rectangle: lower left point and upper right point Here, needless to say, the circle of element ,,, and the rectangle of element are the minimum circumscribed rectangles.

The minimum circumscribing rectangles BBi and B thus obtained
To determine whether Bi's overlap each other, (Bx1, By1) is BB.
i is the lower left coordinate point, and (Bx2, By2) is the upper right coordinate point of BBi,

[0024]

BBi: {(Bx1, By1), (Bx2, By2)} (1), the following conditional expression

[0025]

## EQU2 ## {(Bx1> Bx2 ') ∧ (Bx2>Bx2') ∧ (By1> By2 ') ∧ (By2>By2')} ∨ {(Bx1 '> Bx2) ∧ (Bx2'> Bx2) ∧ (By 1 ′> By 2) ∧ (By 2 ′> By 2)} (2), and if this is satisfied, it is concluded that the two are independent, otherwise they are overlapping. The inter-job dependencies analyzed in this way are summarized in a table format.

The load predicting unit 222 immediately estimates the processing time of the job group divided by the job dividing unit 221. This predicted value can be easily obtained by referring to the standard execution time table of each instruction that constitutes the process.
When the job is composed of a group of subroutines, it is possible to more efficiently predict the time by providing a standard execution time table for each subroutine. This prediction is calculated according to the content of the processing such as the area of the figure to be drawn, the length of the line segment, etc. from the drawing command sequence, but it is even possible to predict the processing time of the job or an amount corresponding to it. If possible, any other method may be used. An example of a table used for load calculation in the load prediction unit 222 is shown below.

FIG. 6 is a diagram showing an example of a load calculation coefficient table of the load predicting section. According to the illustrated load calculation coefficient table, the time coefficient Qt and the space coefficient Qs are defined according to the drawing type of the drawing command, and the load predicting unit 222 rasterizes the jobs divided by the job dividing unit 221. The predicted time required for execution in the processing unit is calculated by referring to the load calculation coefficient table.

In the case of the stroke drawing time Ts, the processing time T is estimated by multiplying the size L of the drawing element by a time coefficient Qti (i: is the drawing type number) predetermined for each drawing type. = Qi × L. The size L of the drawing element is the maximum side of the minimum circumscribed rectangle of the drawing element, that is, m
ax {bx, by} (bx, by: side of minimum circumscribed rectangle)
And

In the case of the filled drawing, the drawing time Ts of the contour drawing is first calculated, and the value obtained by multiplying the area S of the minimum circumscribed rectangle by a predetermined space coefficient Qsi (i: the drawing type number) is used. Calculated, add this to Ts, and add Tf
= Ts + S × Qsi. However, only in the case of straight line drawing, S = L × line width.

The load predicting unit 222 writes these values in the dependency storage table 223 as the processing prediction time. The load predicting unit 222 predicts the amount of data that will be held after the processing, in addition to the prediction of the processing time. This is an index showing how much the cost is in the sense that it occupies the memory space and disturbs other processes when it becomes idle after processing. This data amount can be replaced by S × Qsi. Run length, JPE
When a compression format such as G is used, it is multiplied by a predetermined coefficient according to the compression format. The load predictor is
The data amount calculated in this way is entered in the dependency storage table 223 as the cost amount.

At the stage when the job division and load prediction are completed, the analyzed data dependency is stored in the dependency storage table 22.
3 is stored. An example of the dependency storage table 223 is shown below.

FIG. 7 is a diagram showing an example of the dependency storage table. The illustrated dependency storage table 223 is “JOB”.
/ Mark ”,“ idlecost ”,“ TIME ”,
"PRED JOB / mark" and "POST"
It is composed of columns. It should be noted that the numerical values in the dependency storage table 223 are different from the above-described print job 40, and are values for a print job divided into, for example, seven jobs.

According to this dependency storage table 223,
The "JOB / mark" column is an identification number of each job / a work flag indicating whether or not it has been executed. "Idle cos
The “t” column is a number indicating how inconvenient each job is if it waits for a unit time, and is, for example, an index indicating a storage amount, and is a cost amount calculated by the load prediction unit 222. Therefore, those having a small value have little relation to the schedule improvement processing described later. The “TIME” column is the predicted execution time of each job evaluated by the load prediction unit 222. "PRED JO
The “B / mark” column is an execution start condition of each job, that is, a job identification number to be completed before the execution of each job / a work flag indicating whether the job has been executed. First
And the second job "PRED JOB / mark"
Although the column is negative, this indicates that the job may be started independently of the progress of other jobs, that is, there is no dependency with other jobs. The "POST" column shows "PRED"
Contrary to the “JOB / mark” column, it is the identification number of the job (group) whose start condition is each job. That is, in order to start execution of each job in the “POST” column of the job Ji, the processing of the job Ji must be completed.

Next, the scheduling unit 224 for determining the order and time for arranging the mutually dependent job groups in the plurality of processor groups will be described. FIG. 8 is a flowchart illustrating the operation of the scheduling unit.

In the scheduling section 224, first, as an initialization, the dependency storage table 223 is copied to the work buffer of the scheduling section 224 (hereinafter referred to as a job table), and the startable time T () of each rasterization processing section PEj. All j) are set to 0 (S1).

Next, a job Ji that is not marked among the jobs in the job table and has a negative "PRED JOB / mark" column or all the identification numbers in the "PRED JOB / mark" column, that is, start Set a set of possible jobs in the candidate list (S
2). The marked job is a job that has already been placed in the execution schedule, and the mark is a flag indicating this. A flag “1” indicates that the job has already been placed, and a flag “0” indicates that the job has not been placed.

The candidate list is a linear array [Jk, Jm, Jn, ...] Of the job Ji. However, the linear order is expedient here and has no actual meaning. The meaning of the mark is, as will be described later, the "PRED JO
The flag "B" is "1" when the processing has already been completed, and "0" otherwise.

If the candidate list is empty, skip the following and go to step S8 (S3). If the candidate list is not empty, then one job Ji having the shortest job processing time is selected from the candidate list and assigned to the rasterization processing unit PEj (S4). That is, the scheduling here is based on the SPT rule (shortest processing time rule). The present invention aims to reduce the waiting time until the processing result is received after processing a certain job. Therefore, how the initial scheduling itself is configured is not a problem related to the essence of the present invention, and the EDD rule (earliest delivery rule) may be used as another allocation heuristic rule, Various optimization methods may be applied as the solution.

The job Ji is added to the schedule list SL as a set data of <assignee PEj, identification number of job Ji, start time of Ji>. The selection of PEj designates the rasterization processing unit that can start the processing earliest. This startable time T (j) is determined as follows. That is, the time at which the job Ji can start at PEj is the maximum of the startable time T (j) of the rasterization processing unit and the end time TJ (k) of the job Jk designated by the “PRED” column identification number of the job Ji. Value (latest time). Therefore, the rasterization processing unit PEj
The startable time T (j) of

[0040]

## EQU00003 ## T (j) = max [T (j), TJ (k)] + PT (i) ... (3) is calculated. Here, PT (i) is the processing time of the job Ji. When the processor PEj is selected, T (j)
Is also updated (S5).

When the allocation of the job Ji is completed, all the "PRED JOB" columns in the job table are inspected,
If Ji is found, mark it, and also mark job Ji itself. That is, the flag is set to "1" (S
6).

The procedure for selecting one job from the above candidate list and assigning it to the rasterization processing section is repeated until the candidate list becomes empty (determination in step S3). If the candidate list becomes empty, it is determined whether all jobs in the job table have been marked (S
8) If there are unmarked jobs, a candidate list is set again from unmarked jobs, and the above cycle is repeated until all the rows in the job table are marked. When all the lines are marked, the scheduling unit's initial schedule creation process is completed.

An example in which the job group represented in the dependency storage table 223 of FIG. 7 in this way is scheduled by the above procedure is shown below. FIG. 9 is a Gantt chart showing an example of the schedule.

This Gantt chart shows how the jobs J1 to J7 are arranged with respect to the three rasterizing processing units PE1 to PE3. According to the illustrated example, the rasterizing processing unit PE1 has the job J2. , J3 and J6 are assigned to the rasterization processing unit PE2,
First, the job J1 and the job J4 are allocated from the end time of the job J3, and the job J5 starting from the end time of the job J2 and the job J7 starting from the end time of the job J6 are allocated to the rasterization processing unit PE3. . Here, the shaded portion means an idle time while waiting for the start of the delivery destination job after the job is completed. The schedule shown in this Gantt chart is internally represented by the following schedule list SL.

[0045]

SL: <1,2,0><2,1,0><1,3,6><3,5,6><1,6,15><2,4,15><3 , 7, 21> (4) Next, the schedule improving unit 225, which is the center of the present invention, will be described. The schedule improving unit 225 is a processing unit that improves idle time in the schedule generated by the scheduling unit 224. The concept of improving the idle time in the scheduling improving unit 225 will be described. First, assume that the scheduling unit 224 generates a schedule as shown in the Gantt chart of FIG.
Let's improve this schedule. In the job group of the Gantt chart shown in FIG. 9, for example, job J
Paying attention to No. 5, the job J5 must start processing after waiting for the processing of the job J2, and must finish before the processing of the job J7 starts. Further, in order to shorten the processing time of the entire print job, the rasterization processing unit PE3 sets the processing start time to "6" immediately after the end of the job J2 so that the processing is completed as quickly as possible, and the processing result is the job. I keep it until I hand it over to J7.

FIG. 10 is a schematic explanatory diagram of job movement in the schedule improvement unit. According to the Gantt chart shown in the figure, for the jobs J1 to J7 included in the schedule generated by the scheduling unit 224, the schedule improvement unit 225 does not hold this until the job J5 passes the processing result to the job J7. Paying attention to the fact that it does not happen, we aim to reduce this idle time I1. In order to reduce the idle time I1, the processing start time of the job J5 is delayed by the time t0. As a result, the idle time I1 in which the processing result must be held even after the rasterization processing unit PE3 finishes processing is the rasterization processing unit PE3.
Will be exchanged for the unloaded idle time t0. Similarly, the job J1 can also be moved backward in time, and the idle time I2 until the start of the processing of the job J4 can be exchanged with the idle time when the rasterization processing unit PE2 has no load.

Next, the operation of the schedule improving unit 225 having such a purpose will be described. First, the job table DT obtained by copying the dependency storage table 223 is expanded using the schedule list SL. The procedure will be described.

FIG. 11 is a flow chart showing the procedure for expanding the job table. When expanding the job table DT, first, i = 1 is initially set (S11), and then,
i is incremented (S12), and it is determined whether the following processing has been repeated for the number of jobs (S13). Next, the i-th schedule item <PEj, Jk, T> is extracted from the schedule list SL (S14). Here, PEj is the rasterization processing unit number, Jk is the job identifier, and T is the processing start time of Jk. Next, k rows of the expanded job table DT2, that is, job Jk
The processing start time T of Jk is entered in the "START" item column of (S15), and k of the expanded job table DT2 is entered.
Line, enter T + TJ (k) in the "END" item column (S
16). TJ (k) is the processing time of Jk and is obtained by referring to the "TIME" column of the k row of the job table DT. This extended procedure ends after repeating the number of jobs.

An example of the job table DT2 created as above is shown below. FIG. 12 is a diagram showing the structure of the extended job table. The extended job table DT2 shown in the drawing has the "JOB" of the identification number of each job.
Columns, an "idle cost" column, which is an index showing how inconvenient each job is if it waits for a unit time, a "TIME" column, a "PE" column of an identification number of the rasterization processing unit, and each job Execution condition “PRED” column, estimated execution time of each job “POST” column, processing start time “START” column, and processing end time “EN”
D ”column. "START" column and "E
The “ND” field is a field newly expanded by the job table expansion procedure.

Next, as a second preparation, the schedule list SL is rearranged in the order of the PE number which is the identification number of the rasterization processing section. Here, for example, the previous schedule list SL becomes the next schedule list SL2.

[0051]

SL2: <1,2,0><1,3,6><1,6,15><2,1,0><2,4,15><3,5,11><3 , 7, 21> (5) The schedule improving unit 22 while using the schedule list SL2 and the extended job table DT2.
In No. 5, the improvement process is performed so that the result as shown in FIG. 10 is obtained. The improvement procedure will be described below.

FIG. 13 is a flow chart showing the flow of the operation of the schedule improving unit. Schedule list SL
The improvement procedure of the first is the extended job table DT
The maximum value in the “END” column of No. 2 is searched for, and preparations are made to set the schedule completion time Tend (S21). next,
It is checked whether the number of repetitions of the following processing is less than or equal to a predetermined number of repetitions N (S22). This N
Is a number that represents the accuracy of schedule improvement, an appropriate number,
For example, it may be 1. Next, the processes of the following steps are applied to each job Ji. First, the job number is set to i = 1 (S23), and it is checked whether the job number i is the number of jobs or less (S24). Next, i is incremented (S25), and the completion time T1 of the job Ji (the “END” column of the k row of the expanded job table DT2) is compared with the start time T2 of the succeeding job in the rasterization processing unit of the job Ji. Yes (S26). The start time T2 of the succeeding job refers to the start time of the next job of Ji in the schedule list SL2 if it is the same PE number as Ji. If the PE number is different, that is, if Ji is the last job. For example, Tend-TJ (Ji). Here, TJ (Ji) is the processing time of Ji, that is, “TIM
It is the value in the "E" column.

If T1> = T2, the next job Ji +
1, the process returns to step S24 to check the time, and if T1 <T2, the start time of the job group having the job Ji as the execution start condition is inspected (S27). The "POST" column in the i-th row of the expanded job table DT2 is referenced, and the start time {Tpm} of each job group {Jm} therein is referenced. {Tpm} is the “ST of each Jm
This is the value in the "ART" column. If the job J is in Tpm
If there is something equal to the completion time Tcomp of i,
In order to perform the above processing for the next job Ji + 1,
It returns to step S24.

In the determination of step S27, if Tpm> Tcomp for all Tpms, there is an increase in the idle time of the job that is the execution start condition of the job Ji and a decrease in the idle time of the job Ji itself. The comparison is made (S28). "PRED" of job Ji
Referring to the job group {Jm} in the column, the "idlec" of each Jm
The value Cd in the “ost” column is totaled. This value ΣCd (m) is compared with “idle cost” of job Ji, and if
If ΣCd (m) <Cd (i) −ε, the process proceeds to the next step S29. ε is a predetermined threshold time, and may be 0. ΣCd (m)> = Cd
If (i), the process returns to step S24 to perform the above processing for the next job Ji + 1.

If the "PRED" column of the job Ji is empty (-1), the process goes to the next step S29. The “START” column of the job Ji is replaced with a value obtained by subtracting the processing time TJ (i) of Ji from the minimum value of the start time of the job group {Jm} whose start condition is job Ji. That is, START (Ji) = min {START (J
m)}-TJ (i).

According to the above procedure, the idle time after processing is exchanged for the idle time after processing, and the load is reduced without delaying the overall processing progress. Further, by performing rescheduling (optimization), it is possible to accept a new job in the increased idle time before processing.

According to the operation of the schedule improving unit 225 described above, as shown in FIG. 10, the expected processing time of the job J5 is subtracted from the scheduled processing start time of the job J7 whose execution start condition is the completion of the processing of the job J5. The time when the job J5 is started is defined as the start time. However, since the estimated processing time of each job is different from the actual processing time, the job J5
If the processing time of is longer than the expected processing time, a negative idle time will result, and thus the processing time of the entire print job may become longer. Further, since the job J5 depends on the job J2, if the start time becomes too late after the end time of the job J2, the rasterization processing unit PE1
The load on is increased. In order to avoid this, it is necessary to set the start time taking these into consideration.

FIG. 14 is an explanatory diagram of job movement by another schedule improvement. According to the illustrated Gantt chart, the start time of the job J5 is set before the time obtained by subtracting the estimated processing time of the job J5 from the start time of the job J7. At this time, the start time of the job J5 is "idl" of the job J2 having a dependency relationship with the job J5.
e cost ”and“ idle cost ”of job J5
It is decided in consideration of. For example, the idle time I1 of the job J5 is set to “idle c of the job J2 and the job J5.
The idle time I3 is divided by the reverse ratio of “ost”, and one of the divided idle times is the idle time I3, and the start time of the job J5 is estimated from the start time of the job J7 and the idle time I3. I am trying to set the time to be less.

Thereafter, the job management unit 22 distributes the drawing command sequence to each rasterizing processing device after performing the scheduling and the schedule improvement. In the distribution plan with the improved schedule, the cost of holding the intermediately generated images during processing for a long period of time is kept low.

In this embodiment, the parallel processing device is used in which the rasterizing processing devices are connected by the shared bus. The post-processing idle time is replaced with the pre-processing idle time so that another job is accepted. It was meant to make it easier.
On the other hand, the present invention can be applied to a distributed processing system by network connection in the same manner. In this case, exchanging the idle time has the effect of reducing the load not only for the print job but also for other general user jobs as a whole.

[0061]

As described above, according to the scheduler of the present invention, the start time of a divided subtask is set to the start time of the divided subtask in the scheduler for scheduling parallel processing of mutually dependent job groups. By configuring the system to include the idle time adjustment unit that changes the time when the idle state is resolved to the time obtained by subtracting the predicted processing time by the load prediction unit, the idle time that waits for another job while holding the intermediately generated data of the processing is It is exchanged for idle time before processing, and the period of holding intermediately generated data for a long time is reduced, so the processing unit can reduce the load on general user processes other than the processing of subtasks. , The processing resources are effectively used.

[Brief description of drawings]

FIG. 1 is a block diagram showing a principle configuration of a scheduler according to the present invention.

FIG. 2 is a diagram illustrating a configuration example of a print processing system.

FIG. 3 is a block diagram illustrating a configuration example of a job management unit.

FIG. 4 is a diagram illustrating an example of division of a print job.

FIG. 5 is a diagram showing an image formed when a divided instruction sequence is executed.

FIG. 6 is a diagram showing an example of a load calculation coefficient table of a load prediction unit.

FIG. 7 is a diagram showing an example of a dependency storage table.

FIG. 8 is a flowchart illustrating an operation of a scheduling unit.

FIG. 9 is a Gantt chart showing an example of a schedule.

FIG. 10 is a schematic explanatory diagram of job movement in a schedule improvement unit.

FIG. 11 is a flowchart showing a procedure for expanding a job table.

FIG. 12 is a diagram showing a configuration of an extended job table.

FIG. 13 is a flowchart showing an operation flow of a schedule improving unit.

FIG. 14 is an explanatory diagram of job movement due to another schedule improvement.

[Explanation of symbols]

 1 task dividing means 2 load predicting means 3 dependency storing means 4 scheduling means 5 idle time adjusting means 6 arithmetic processing devices 11, 12, ... Client computer 20 printing device 21 data receiving unit 22 job management unit 23a, 23b, 23c , 23d Rasterization processing unit 24 Page memory 25 Printing device 26 Shared bus 30 Ethernet

Claims (5)

[Claims] 1. A scheduler for scheduling a divided task so that one task is divided and processed in parallel by a plurality of arithmetic processing units, a plurality of tasks including a group of tasks having mutual dependence. Task dividing means for dividing into subtasks, load predicting means for predicting the processing time of each subtask, dependency storing means for holding the processing dependency relationship between each subtask, and scheduling for determining the processing schedule of the subtask group Means and idle time adjusting means for changing the start time of the subtask scheduled so that the idle state occurs after execution by the arithmetic processing unit to a time obtained by subtracting the processing time predicted by the load predicting means from the idle state resolution time. A scheduler characterized by including. 2. The scheduling means includes a task information table indicating whether or not each subtask to be scheduled is an execution start condition of another subtask, and the idle time adjusting means includes an execution start condition in the task information table. The start time of the first subtask is scheduled to be a time obtained by subtracting the processing time of the first subtask predicted by the load predicting means from the estimated start time of the corresponding second subtask, and becomes the execution start condition. The start time of the first subtask that has not been completed is the first time predicted by the load prediction means from the estimated completion time of the last subtask.
2. The scheduler according to claim 1, wherein the scheduler is configured to perform scheduling at a time when the processing time of the subtask is reduced. 3. The idle means for adjusting the idle time includes a cost storage means for storing an idle cost coefficient which evaluates a resource cost in an idle time after the processing of each subtask until the processing of a subsequent subtask is started. Is the cost of the idle state of the second subtask with the first subtask as the start condition, which is calculated by referring to the dependency storage means and the cost storage means when the start time of the first subtask is changed. 2. When the value of increase exceeds the cost decrease of the idle state of the first subtask, the change of the start time of the first subtask is stopped. Scheduler. 4. The idle time adjusting means is configured to change the start time of the subtask to a time obtained by subtracting the time obtained by adding the coefficient time based on the idle cost coefficient to the estimated processing time from the idle state cancellation time. The scheduler according to claim 3, characterized in that 5. A job management apparatus for a printing system, wherein print information is distributed to a plurality of arithmetic processing units in order to perform rasterization processing for obtaining pixel information required for printing from a source file describing print information in parallel. The job division means that divides the source file that describes the command sequence into multiple command sequences, the dependency analysis method that analyzes the spatial coherence between the image elements formed by each command sequence, and the rasterization processing time for each command sequence. A load predicting means for predicting, determining a processing schedule of a command sequence group for the plurality of arithmetic processing devices based on each load prediction result of each command sequence and a spatial coherence analysis result of an image. And the first command sequence executed in the plurality of arithmetic processing units is Only if you have not become a condition for starting the connection of the second command sequence,
The allocation of the first command sequence is scheduled to start at the time when the rasterizing process time of the first command sequence analyzed by the load predicting means is subtracted from the time when the subsequent third command sequence starts processing. A job management apparatus for a printing system, comprising: a scheduling unit.