JPH1040224A - System for monitoring action of network inside parallel computer - Google Patents

Abstract

PROBLEM TO BE SOLVED: To minutely analyze the communication state of a communication message in the network of a parallel computer by monitoring a hardware without changing a program. SOLUTION: A result obtained by monitoring the hardware concerning whole communication packets is preserved in a storage device, the monitored result is analyzed, rectangulars 10 and 11 corresponding to the communication message which is color-divided by communication kind and communication data length(8) with the route node number 30 and the time 31 of the network as axes are arranged in places corresponding to the route node and the passage time of the network and a stay time in the route node is expressed by rectangular width. Moreover, the passage route of the communication packets is tracked and the rectangulars 10 and 11 expressing the communication packets are connected by a straight line 21.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a system for monitoring the internal network behavior of a parallel computer.

[0002]

2. Description of the Related Art Computer Performance Evaluation and Modeling Techniques and Tools
g Techniques and Tools-, 7th International Confere
nce Proceedings, pp.76-88, May.3-6, 1994 (Springer-V
According to erlag), the performance monitor ParAide mounted on Intel Paragon (two-dimensional mesh structure parallel computer) is a system performance visualization (System Performance Visualizati).
on: SPV) and ParaGraph (MTHeath, JAEtheridge:
“Visualizing the Performance of Parallel Program
s ”, IEEE Software, 8 (5): 29-39, September.1991)
Has a network monitor and drawing tools between processors.

[0003] The SPV is used in each processor node.
It is a tool that collects the CPU usage rate, bus usage rate, and network traffic, partly with a hardware performance monitor counter, partly with the Mach kernel, and drawing it on a two-dimensional mesh. Operates independently of application code. This may be due to the accumulation of the integrated value or sampling data.

ParaGraph collects an event trace by linking a source library with a performance library, converting the code, and executing the program. Event trace results can be saved in ParaGraph
Then, use ParaGraph's drawing tool to draw the CPU usage rate and communication status between processors using the drawing tool of ParaGraph. The communication network analysis diagram is a diagram based on time and node numbers. It is not known whether the communication message is being tracked or the message stay time in the node is taken into account, but the source code must be converted to create the trace data.

[0005] Also, Computer Performance in the above publication.
Evaluation pp.25-51 and May.1994 include Pablo Communic
There is a tool called ation Matrix. The vertical axis is the sender (Send
er), the number of bytes of the communication data amount between the Sender / Receiver pairs or the number of communication messages is represented in a corresponding grid by a color tone on a grid-like graph with the horizontal axis representing a receiver. In this tool, to obtain execution result data, an object code in which the performance acquisition code is incorporated into the application source code is generated, and the execution result is collected as a trace result by linking the data capture library.

[0006]

According to the prior art, it is necessary to embed a monitoring code for measuring the execution performance of a program on a parallel computer in a source code. It was necessary to run through a dedicated compiler and get the event trace results. The object code after the insertion of the monitoring code may have a different behavior from the execution operation of the object code before the insertion, and the execution time is extended for monitoring.

Conventionally, an OS or a hardware performance monitor counter has been used to obtain a part of information such as a CPU operating rate and an inter-processor communication amount. Only information can be obtained, and in particular, it is not possible to obtain detailed information on passage of an interprocessor communication message.

Further, according to the expression form in which the Sender / Receiver of the inter-processor communication is arranged in a grid pattern in the execution result and the method of expressing the communication amount between the Sender / Receiver, the inter-processor communication message includes a plurality of processor nodes,
Or, if the network structure is such that it passes through multiple path nodes and arrives at the Receiver node, communication messages compete and pass on that path, causing a communication performance bottleneck. In this case, the state on the route cannot be recognized. Furthermore, it is not possible to obtain a communication result for each communication message.

Similarly, even in a method of displaying only the amount of data communication with an adjacent node in a two-dimensional mesh network, communication is performed at any place on a route that reaches a destination node through a plurality of nodes. It is difficult to find out if a performance bottleneck is occurring.

According to the present invention, the program code is not modified at all and the program behavior is not affected.
The purpose of the present invention is to hardware monitor all inter-processor communication messages passing through a network and to trace the behavior of the inter-processor communication messages in detail.

Further, even when an identification number is added to a communication message, even when the communication message does not have an identification number, the type and data length of the communication message can be determined to determine the passage of the communication message. The purpose is to guess and track and to grasp the behavior of communication messages.

[0012]

In order to solve the above-mentioned problems, a monitoring mechanism for an inter-processor communication message is added to each path node existing on a network between processors of a parallel computer and relaying the inter-processor communication message. It monitors the type and length of the communication message, the name of the processor node and the destination node that issued the communication message, and the time when the head and tail of the communication message have passed. The monitor of the path node generates a monitoring record every time a communication message passes, sends the monitoring record through a dedicated line for monitoring results, and stores the monitoring record in a storage device independent of the program operation of the parallel computer. deep. After executing the monitoring target program, analyze the trace result of the communication message.

In the pre-analysis processing, it is determined whether the path node number and the communication message are monitoring records when the communication message has passed through the sending node, the relay node, or the receiving node, and a record queue is created in time order. Keep it.

Next, the path of the communication message is traced.
From the head of the monitoring record queue corresponding to the sending node, the relay node of the route node to be passed next,
Or from the record queue of the receiving node's record:
The communication message identification number, or the type and length of the communication message, and the processor node name that issued the communication message and the destination node name all match, and among the records with the largest time, the monitoring record of the sending node and the most recent time Are linked as the next passing data of the communication message.

The result of tracing the route of the communication message is arranged in a time chart on the axis of the route node number and the time as a rectangle of a color predetermined by the width of the route node residence time, the communication message type and the length. Then, based on the result of tracing the rectangle representing the communication message, a connection is made.

[0016]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described with reference to FIGS.

FIG. 1 is a tracing diagram of an inter-processor communication packet path.

It is an object of the present invention to monitor inter-processor communication packets and express the behavior in a time chart graph as shown in FIG. Reference numeral 1 denotes a window representing a communication packet path trace diagram between processors. The vertical axis 30 is an axis representing the path node number on the inter-processor communication path, and the horizontal axis 31 is an axis representing the passage of time. Vertical axis 30
Upper 3 and 4 represent path node numbers. 10,11,1
2,13,14,15,16,17,18,19,20
Represents a communication packet observed at each path node. Reference numeral 8 denotes a legend for determining the type and length of a communication packet by color or hatching pattern. The types of packets are classified into an urgent packet 80 with a high priority and a normal packet 81. The urgent packet is indicated by diagonal lines, and the normal packet is indicated by solid color. Further, the packet length, for example, 82 is a short packet of 0 to 127 bytes, 83 is about 128 to 511 bytes, and 84 is 512-
4095 Bytes and 85 are classified as long packets of 4096 Bytes or more, and are classified into light and dark colors or arbitrary colors such as blue, green, yellow, and red. In the communication packet rectangle, the time when the communication packet enters the route node is set to the left end, and the time spent in the route node is set to the width in the time axis direction.

In the figure, reference numeral 2 denotes a scroll bar for scrolling the path node axis 30, and reference numeral 7 denotes a time slider for scrolling the time axis 31. The absolute time on the time axis 31 is set by moving the scale 70 left and right. Reference numeral 6 denotes a time scale slider for defining the width of the time axis 31 to be displayed. By moving the scale 60 of the time scale slider, the display time width of the communication packet path tracking diagram 1 is changed. 5 represents a traffic indicator. It indicates how much inter-processor communication is occurring as a whole in all the parallel computer systems. For example, a section 50 indicates a medium level, a section 51 indicates frequent communication, and a section 53 indicates that communication hardly occurs. The traffic indicator may set an option and express it as statistics of communication packets passing through the path node for a specific path node or a specific set of path nodes.

The communication packet 10 has a packet length of 0-1.
This is a normal packet of 27 Bytes, and the route node R (0,
0). Similarly, the communication packet 11 indicates that the communication packet 11 has passed the path node R (1, 0). The width of the communication packet rectangles 10 and 11 in the time axis direction is
Communication packets 10 and 11 are route nodes R (0,
0) or R (1,0).

In each of the inter-processor communication path nodes,
Since observation of communication packets is performed, 10, 11, 12, 1
As shown in 3, 14, 15, 16, 17, 18, 19, and 20, rectangles representing communication packets are distributed throughout the inter-processor communication packet path tracking diagram.

Further, in order to track the passage of a communication packet, if the packet is regarded as the same packet, 2
A rectangle representing a communication packet such as 1, 22, 23 is connected. 10, 21, and 11 have a packet length of 0 to 127 bytes.
s indicates that the normal packet has passed in the order of the route node R (0,0) → the route node R (1,0). Similarly, the communication packets 14, 15, and 16 have a packet length of 128-5.
11 Bytes normal packet, route node R (4, 0)
→ R (0,0) → R (0,1).
The reason why the widths of the rectangles 14, 15, and 16 are different even for the same communication packet is that the time spent in the route node is different. Even if packets of the same length are waiting for communication due to congestion on the route or the like, the time spent in the route node may take a long time, and thus the route node stay time may be different. In the performance analysis, it is possible to analyze in detail the communication packets in which the route node stay time is unnecessarily large with respect to the packet length and the communication packets that occur during the same time.

FIG. 2 is a diagram showing the overall configuration of the parallel computer internal network behavior monitoring system.

Reference numeral 200 denotes a parallel computer. 201,2
03 denotes a PU (processor unit), and 205 and 206 denote communication networks between processors. 202, 204
Is a router of a communication network between processors, and is hereinafter treated as a network communication path node. A clock is generated by the clock width adjuster 216 and the clock oscillator 215 and distributed to each network communication path node via the clock distribution line 209. As a result, a clock indicating the same time is realized for all network communication path nodes in the parallel computer 200.

Reference numeral 210 denotes a monitoring control device that monitors the inter-processor communication packet and analyzes and calculates the monitoring result. Packets communicating between the processors are monitored by the network communication path nodes 202 and 204, and the monitoring result is notified to the monitoring control device 210 via the dedicated monitoring lines 207 and 208. The monitoring control device 210 has a monitoring record analysis unit 211, which analyzes a record of a monitoring result and stores it in the communication packet analysis table 212. Further, the detailed result is analyzed by the part 213, and is displayed as the one-processor communication packet path diagram of FIG.
14 is displayed.

FIG. 3 is a diagram showing a router and a processor unit on the interprocessor communication network.

Reference numeral 201 denotes a processor unit, 202 denotes a router, and 205 and 206 denote inter-processor networks. The communication packet between processors is a processor unit.
The communication packet input / output unit 303 of the router 202 and the routing determination unit 304 issue an inter-processor network 205,
The transmission packet input / output unit 305 determines which of the transmission destinations is to be transmitted, and transmits it to the inter-processor network from either the transmission unit 314 or the transmission unit 312 via the communication packet input / output unit 305.

A communication packet from the outside to the processor unit is transmitted to the inter-processor network 205,
From any of 206, 315 or 313 from router
The information is input to the network interface unit 302 via the communication packet input / output unit 305, the routing determination unit 304, and the communication packet input / output unit 303, and is decoded.

The routing determination unit 304 determines whether to output or input to the X-direction inter-processor network 205,
An output to or an input to the directional inter-processor network 206, or a relay from the X-directional inter-processor network 205 to the Y-directional inter-processor network 206 or a It is determined whether the packet is relayed to the directional processor network 205, and the communication packet passes through each router route.

In this embodiment, the path through which the communication packet passes from the source processor unit to the destination processor unit is a plurality of arbitrary path nodes, and the order is uniquely determined in advance. Is assumed.

Reference numeral 306 denotes a communication packet monitor unit.
In step 307, every time a communication packet passes, a monitoring result of the communication packet such as the transmission / reception processor unit number of the communication packet and the passage time of the communication packet is generated as a monitoring record 500, and the monitoring record 500 in FIG. To the monitoring controller. Reference numeral 308 denotes a time stamp generation unit for imprinting a time stamp on the monitoring record, based on a signal oscillated from a single clock oscillator 215 for all routers.
Generate a time stamp using the counter.

FIG. 4 shows an example of an inter-processor communication packet format.

(A) is a communication packet 400 between processors without a communication packet ID, and (b) is a communication packet ID4.
12 shows an inter-processor communication packet 410 with twelve. The inter-processor communication packet 400 includes a communication start processor ID 401, a communication end processor ID 402, a communication type designation unit 403 indicating the priority and type of inter-processor communication,
A packet information tag including a communication data length 404 and the like;
It comprises a communication data section 406. The inter-processor communication packet 410 includes the communication packet I in the packet information tag section.
D412 is added.

FIG. 5 shows a monitoring record format of an inter-processor communication packet.

(A) The monitoring record 500 corresponds to the inter-processor communication packet 400, and (b) the monitoring record 510 corresponds to the inter-processor communication packet 41.
This is a format corresponding to 0.

The monitoring record 500 indicates that the inter-processor network router 202 has transmitted the communication packet 400
Each time the monitoring record is passed, the monitoring record generation unit 307
Is generated. The monitoring record 500 includes a router ID 501 for which the monitoring record has been generated, a 502 communication start processor ID, a 503 communication end processor ID, a 504 communication type, and a 505 communication created based on the information of the packet information tag portion of the communication packet 400. Data length,
It is composed of the route ID in the router when passing through the routing determination unit 304, the passage time 507 at the head of the communication packet when the communication packet passes through the router, and the passage time 508 at the end of the communication packet.

The monitoring record 510 has a configuration in which a communication packet ID 512 is added.

Since the path through which the communication packet passes is uniquely determined in advance, the communication packet passes through the next information from the four pieces of information of the router ID 501, the communication start processor number 502, the communication end processor number 503, and the route ID 506 in the router. The router to be determined can be determined.

FIG. 6 shows a packet queue record format.

The packet queue record 600 is a record obtained by sorting the monitoring records of the inter-processor communication packets by time and router ID in order to draw the inter-processor communication packet path tracing diagram 1 of FIG. . The packet queue will be described with reference to FIG.

The packet queue record 600 stores a monitoring record of the monitor of the inter-processor communication packet passing through the network router as packet information 601,
Router passage information 602, pointer 60 to next record
3. The pointer comprises a pointer 604 to a record of the communication packet in a router through which the communication packet passes next. The packet information 601 includes a communication start processor ID 605, a communication end processor ID 606, a communication type 607, and a communication data length 608. The router passage information 602 includes a router ID 609 and a route ID 6 within the router.
10, the communication packet head passage time 611, the router residence time 612 of the communication packet calculated from the communication packet head passage time and the communication packet tail passage time.

One use of the packet queue record is drawing information for drawing the communication packets 10 and 11 in FIG. For example, the drawing position of the communication packet 10 in FIG. 1 is the position corresponding to the router ID 609 in the vertical direction, and the communication packet head passage time 6 in the horizontal time axis direction.
At a position corresponding to 11 on the left end, a rectangle having the router residence time 612 of the communication packet as the width in the time axis direction is drawn. The color of the rectangle is communication type 607, communication data length 608
Draws a predetermined color from. The straight line 21 connecting the communication packet rectangle 10 and the communication packet rectangle 11 is obtained after the communication packet passing route search described with reference to FIG.

Another use of the packet queue record is convenient when searching for a passage route by paying attention to a communication packet, such as a straight line 21 connecting the communication packet rectangle 10 and the communication packet rectangle 11.

(B) The packet queue record 620 is
This is a packet queue record corresponding to the monitoring record 510. A communication packet ID 625 is added.

FIG. 7 is a diagram showing a packet queue and a time index table.

The packet queue 700 is a set of queues obtained by sorting the packet queue records 600 in time order and by router ID. Reference numerals 703, 704, and 705 indicate router IDs. 701 is a time axis. OUT706,
The PASS 707 and the IN 708 respectively link the monitoring record of the communication packet output from the router, the monitoring record of the communication packet relayed through the router, and the monitoring record of the communication packet input to the router in chronological order in the packet queue record format 600 of FIG. It is a queue. These are referred to as an OUT queue, a PASS queue, and an IN queue, respectively.

For example, when the router ID is Router [0] and a communication packet starting from this router is observed,
The links are linked to the IN queue 709 in the order of time, such as 600, 710, and 712. 613, 711, and 713 are pointers to the next record. Router [0]
The records 715 and 717 are linked to the PASS queue 714 for the communication packet having the relay point as a relay point. Similarly, for a communication packet having Router [0] as the end point, I
Link N queues. Further, when a packet departed from Router [1] relays Router [0], assuming that records regarding the packet are 722 and 715, the packet link 723 is linked to link information regarding the same packet. This link corresponds to a straight line 21 connecting the communication packet rectangles 10 and 11 in FIG.

Reference numeral 730 is a time index table. A drawing target time in the inter-processor communication packet path trace diagram of FIG. 1 and a start time of a communication packet passing route search are set, and a time index table for time search is provided so that drawing can be started from an arbitrary time. When the unit time is TimeSlice 702, the time scale is T
[0] 732, T [1] 735, and T [j] 736. This table is linked to the record of the packet queue 700. For example, 740 indicates the first record 600 among the records of T [0] or more and less than T [1] with respect to Router [0]. Similarly, T for Router [1]
It points to the first record 722 of records that are not less than [0] and less than T [1].

By linking the packet queue 700 and the time index table 730, a record at an arbitrary time can be efficiently extracted.

FIG. 8 is a diagram showing a flow of the entire processing up to drawing FIG. 1 of the inter-processor communication packet path tracing.

In order to create data for the inter-processor communication packet path trace diagram, a time range Ts, Te for which a communication packet path is to be searched is designated 801. If the time range cannot be specified, specify the number of monitoring records. Enter the monitoring records in the specified range, 802
Packet queue 700 and time index table 7
803, 804 for which 30 has been created. Next, a search 805 of the communication packet passage route is performed to search and link the packet links 604 and 723 in FIG.

When the packet queue 700 and the time index table 730 are completed, the time sliders 7, 70 and the time scale sliders 6, 6 in FIG.
From 0, a drawing start time and a time scale are input 806. Assuming that the drawing start time is T [0], this value needs to satisfy the time range Ts ≦ T [0] <Te. Also, a time index value (time slice width) 702
Is the time scale ΔT6 which is the drawing width,
0 is defined as a multiple of Δt.

After the drawing start time and the drawing width are defined, a communication packet and a passage route are drawn 807. The traffic indicator 5 is calculated and drawn 808, and finally the number of search failures and the number of searched communication packets are reported 809, and the whole process ends 810.

FIG. 9 is a diagram showing a flow of the method 803 of creating a packet queue.

While a monitoring record is being input 90
1, Determine router number 902, Record to be registered in OUT queue 903, Record to be registered in PASS queue 905, Record to be registered in IN queue 9
07, the OUT queue 904 and the PAS, respectively.
The information is registered in the S queue 906 and the IN queue 908 in order of time.

FIG. 10 is a flowchart showing a method 804 of creating a time index table.

A time index width (time slice width 702) Δt, a time index start point time T0, and a time index total number MAX 1001 are designated. The routers 703, 704, and 705 of the packet queue, the records registered in the queues of the record types OUT706, PASS707, and IN708 are 1002, and each queue is processed. Curr the address of the first record
1003 is set as ent_p. current_p is a pointer to the record currently being handled.

When Ti = T0 + i * Δt, 100
5. Search for a record corresponding to Ti in the time index table from the head of the queue. If the time of current_p is smaller than Ti, 1006 is a record relating to T [i-1] or earlier, so current_p is moved to the next record 1012. If the time of current_p is equal to or greater than Ti, 1007 is a record relating to T [i + 1] or later, so that the current pointer remains unchanged as I
Is incremented 1013, and the next processing is performed. curr
The time of ent_p is equal to or greater than Ti, and T [i
1001 is a record related to T [i]. The record that satisfies the condition 1008 for the first time since the search was performed from the head of the queue is the record to be pointed to by the T [i] pointer of the router ID and the packet queue record type 1009. Current pointer and i
Is incremented by 1010, and the process proceeds to the next process 1020.

If i exceeds MAX, 1004, the processing for the router ID and packet queue record type ends 1011.

The same processing is performed 1002 for all router IDs and all packet queue record types.

FIG. 11 is a flowchart of a search for a passage of a communication packet.

A pointer 60 to the next passing router record of the communication packet in the packet queue record of FIG.
4 and FIG. 7 are flowcharts for searching for links of the packet links 604 and 723 in the packet queue 700.

Time index Ts of range to be searched
The search is started 805 by specifying (start point) and Te (end point). 1101 to search from the router ID number 0. The time index Ts of the OUT queue of the router ID
1102 to search from the records existing in. For the OUT queue record, a record corresponding to the communication packet is searched from the PASS queue record or the IN queue record of the router ID that the communication packet should pass next, and a packet link is established. Communication start point processor I
D, since the passage route can be uniquely specified from the communication end processor ID, whether the router ID to be passed next becomes a relay point or an end point, that is, an input to the processor unit, depends on the communication start processor ID and the communication end point. Processor I
D, the router ID, and the intra-router ID can be uniquely calculated.

Hereinafter, the flow of the search process will be described with reference to FIG.

The current pointer is set to current.
In addition, a pointer for following a passage route of the communication packet is set to path_p 1104. 110 in search range
3, path_p = current, 1126, the communication start point processor I of the communication packet of the record indicated by the current pointer
D (Sender) 605, communication end point processor ID (Receiv
er) 606, router ID (Router) 609, and router ID (path) 610, from the four pieces of information, the router ID (next_Router) that the communication packet should pass next, and the type of relay point or end point ( next_Type) 1105. out of the packet queue 700 corresponding to next_Router, next_Type, which is larger than the communication packet head passage time 611 of the path_p record, the communication start processor ID (Sender) 605, and the communication end processor ID
(Receiver) 606, communication type 607, communication data length 60
A record in which all four pieces of information 8 match, or 6
In the case of the packet queue record of the 20 type, a record matching the communication packet ID 625 is searched 1106.

If there is a record satisfying the condition of 1106, the record 1122 considers the record with the smallest passage time 611 of the head of the communication packet among the records satisfying the condition as the data on the second route of the communication packet. Then, it connects 1107 to the packet link of path_p.

If the communication packet ID 625 matches, the packet is regarded as data on the second path of the communication packet, and is connected to the packet link of path_p 110.
7. The pointer of path_p is moved 1108, and the processing of 1105, 1106, 1107, 1108 is repeated 1109, 1126 until the communication end point processor ID 606 of path_p matches the router ID 609.

If there is no record that satisfies the condition of 1106, in step 1123, the fail counter is incremented by 1110, and the search for the packet is terminated.
11. At the end of the search, whether it is an end point or not, path
Packet link connected to the end of _p is set to nil 1
111. Set the current pointer current pointing to the record to O
The next record in the UT queue is moved to 1112, and the same processing is performed for the next communication packet 1125.

When the search end point Te of the currently handled OUT queue is reached 1103, and among the time slices at the search start point Ts, if there is an unsearched one among the records of the PASS queue, the PASS record is set at the head. A search is made for a packet link 604 from the record of the communication packet according to a later passing route 1113.

The above-described process of searching for a passage of a communication packet is performed for all router IDs.
5.

FIG. 12 is a diagram showing a statistical graph of the occurrence frequency of communication packets.

If the packet queue and the time index table of FIG. 7 are used, it is easy to draw a statistical graph as shown in FIG. 12 in addition to the tracking diagram of the inter-processor communication packet path shown in FIG.

Reference numeral 1200 denotes a window for drawing a communication packet occurrence frequency statistical graph.

Reference numerals 1201, 1202, and 1203 denote inputs at the router R (0, 0) and router R (1, 0), respectively.
7 is a bar graph showing the frequency of occurrence of communication packets when observing an input at 0) and an output at router R (0,0). The horizontal axis represents the length of the communication packet, and the vertical axis represents the frequency of occurrence. Depending on the type of communication packet shown in the legend 1205, the occurrence frequency is accumulated as indicated by 1211, 1212 and 1213. The communication packet type 1205 includes a preference communication having a high priority, a normal communication ordinary, and a broadcast communication Broa communicating from one processor to all processors.
dcast, Point-to-P from one processor to one
Classify into oint communication. Prefer for communication priority
Broa that indicates the number of communication destinations by color-coding ence and ordinary
The distinction between dcast and point-to-point is to distinguish a rectangular fill pattern by diagonal or full fill. The type of communication is expressed by the combination of the color and the paint pattern. 12
Reference numeral 04 denotes an observation packet drawing menu, which designates input monitoring in, output monitoring out, and crossbar transfer pass drawing. A time slider 1207 sets an observation time. Reference numeral 1206 denotes a window for confirming the snapshot time. 1201,120
In order to draw a communication packet occurrence frequency statistical graph of 2,1203, the packet queue record 6 included in each time slice 702 of the packet queue 700 of FIG.
00 may be calculated.

The packet queue 700 and the time index table 730 shown in FIG. 7 have a data format effective for providing a plurality of pieces of analysis information as shown in FIGS.

[0076]

According to the present invention, information obtained by monitoring the inter-processor communication packet on each path node on the inter-processor communication network is represented by a rectangle colored by the type and length of the communication packet, and the path node and time are represented. By arranging them on the axis time chart, the traffic on the inter-processor communication network can be accurately grasped.

According to the communication packet passing route search method of the present invention, not only when there is a communication packet identifier but also when there is no communication packet identifier,
It is easy to detect the passage status of communication packets on the network, and when drawing a rectangle of a communication packet on a time chart based on a route node and time, information on the same communication packet is connected by a straight line. Thus, the behavior of the communication packet can be accurately grasped.

By using the packet queue and the time index table format of the present invention, an arbitrary time,
For a given path node, the communication state of whether the communication packet has departed from the source processor unit or whether the communication packet has reached the destination processor unit,
Drawing information such as the type of communication packet, the length of the communication packet, and the passage time of the communication packet can be easily derived.

[Brief description of the drawings]

FIG. 1 is an explanatory diagram of inter-processor communication packet path tracking.

FIG. 2 is a block diagram of the overall configuration.

FIG. 3 is a block diagram showing a router and a processor unit.

FIG. 4 is an explanatory diagram of an inter-processor communication packet format.

FIG. 5 is an explanatory diagram of a monitoring record format.

FIG. 6 is an explanatory diagram of a packet queue record format.

FIG. 7 is an explanatory diagram of a packet queue and a time index table.

FIG. 8 is a flowchart of the entire processing in tracking a communication packet path between processors.

FIG. 9 is an explanatory diagram of a method of creating a packet queue.

FIG. 10 is an explanatory diagram of a method of creating a time index table.

FIG. 11 is a flowchart of a communication packet passage route search.

FIG. 12 is a communication packet occurrence frequency statistical graph.

[Explanation of symbols]

1 ... Packet path tracking diagram, 3,4 ... Router number, 5 ... Traffic indicator, 8 ... Legend indicating type and length of packet 10, 11, ... Rectangle indicating packet, 21 ...
A straight line representing a packet passage path, 31... A time axis.

Claims (8)

[Claims] 1. In a parallel computer having a plurality of processors, the inter-processor communication message passes through a uniquely determined route including a plurality of arbitrary route nodes, and a node name of a source and a destination and a message A means for monitoring the inter-processor communication message and a common time stamping means for each path node, having information on message types such as length, message priority, etc., and a common time stamping means; When monitoring the route node number through which the communication message passed, the source and destination node names, the message length, the message type such as the priority of the message, the passage time of the route node, and the resident time at the route node, the monitoring was performed. It has means to accumulate and analyze communication message information and display the analysis result. The analysis result of the communication message is shown in a time chart graph with the path node number and time as axes, and the communication message information is displayed at the start point of the time when the communication message entered the path node, the width of time spent in the path node, and the communication packet. Is represented as a rectangle expressing the type and length of the color, and placed on the time chart graph at the location corresponding to the path node number and the passage time of the communication message,
An internal network behavior monitoring system for a parallel computer, characterized by expressing a passage path of a communication message by connecting rectangles representing communication message information. 2. The representation of the inter-processor communication message information analysis result in which the statistics of the traffic of the entire parallel computer system or an arbitrary selected path node within a unit time are displayed below the time chart graph. Internal network behavior monitoring system of parallel computer with method. 3. The method according to claim 1, wherein said communication packet is transmitted from a source node and a destination to all communication message information obtained by monitoring a communication packet in advance. The communication path node number is determined in the communication packet information queue corresponding to the source node, the destination node, and the relay node, which is created for each communication path node number. A record that describes information such as a message type such as a source node name, a destination node name, a communication message length, a message priority, a time at which the message entered the route node, and a time at which the message resided at the route node. Registered in chronological order, and for the communication packet information queue created in advance,
The record is taken out from the head of the communication packet information queue corresponding to the source node and registered in advance,
According to the route through which the communication packet should pass, from among the unsearched records of the corresponding relay node communication packet information queue and the receiving node communication packet information queue, the source node name, destination node name, communication message length, message All information about the message type, such as priority, matches and the message entry time is new,
An internal network behavior monitoring system for a parallel computer, wherein a closest record is selected, and a history of communication path nodes passed for each communication packet is searched. 4. The communication between processors according to claim 3, wherein statistics of the communication amount of the entire parallel computer system within a unit time or an arbitrary selected path node are displayed at a lower part of a time chart graph. An internal network behavior monitoring system for a parallel computer equipped with a method for expressing message information analysis results. 5. The communication message information according to claim 1, further comprising means for storing and analyzing the monitored communication message information when the inter-processor communication message has a unique communication message identification number, and displaying the analysis result. The analysis result is shown in a time chart graph with the path node number and the time as axes, and the communication message information is provided. The starting point is the time when the communication message enters the path node, the time spent in the path node is the width, the type of communication packet and The length is represented as a rectangle that is represented by color, and on the time chart graph,
It is arranged at a location corresponding to the path node number and the passage time of the communication message, and expresses the passage path of the communication message by connecting rectangles representing the communication message information. An internal network behavior monitoring system for parallel computers equipped with an expression method. 6. An internal network behavior monitoring system for a parallel computer according to claim 5, wherein statistics of the traffic of the entire parallel computer system within a unit time or of an arbitrary selected path node are displayed at the bottom of a time chart graph. 7. The method according to claim 5, wherein said communication packet is transmitted from a source node and a destination to all communication message information obtained by monitoring communication packets in advance. The communication path node number is determined in a communication packet information queue corresponding to a source node, a destination node, and a relay node, which is created for each communication path node number, by judging whether the packet has passed through a node or a relay node. , Communication message identification number, source node name, destination node name, communication message length, message type such as message priority, time when the message entered the route node, time when the message resided at the route node Are registered in chronological order, and a record is created for the communication packet information queue created in advance. Te retrieves a record from the beginning of the communication packet data queue corresponding to the source node,
A record whose communication message identification number matches among the unregistered records of the corresponding relay node communication packet information queue and the reception node communication packet information queue registered in advance according to the path through which the communication packet should pass. And an internal network behavior monitoring system of the parallel computer for searching for a communication path node history that has passed for each communication packet. 8. The processor-to-processor communication according to claim 7, wherein statistics of the communication amount of the entire parallel computer system or an arbitrary selected path node within a unit time are displayed at the bottom of the time chart graph. An internal network behavior monitoring system for a parallel computer equipped with a method for expressing message information analysis results.