JP2010224897A - Buffer control system, method, partition sharing interface using the same, and system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To detect a stagnation state before detection of time-out by a usual transaction, and to increase speed of buffer read control on occurrence of a failure, for reducing a multi-partition failure caused when a sound partition is involved in a time-out failure in one partition. <P>SOLUTION: A system for controlling a buffer of an interface shared by a plurality of partitions includes: a stagnation monitoring part autonomously monitoring a data stagnation state in the buffer to detect occurrence of the time-out failure due to the data stagnation state before detection of time-out by a request and a reply of a transaction; and an address skip control part skipping an address for storing data of a partition that causes the time-out failure in reading data from the buffer when occurrence of the time-out failure due to the data stagnation state of a certain partition is detected. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a buffer control system, a method, and a partition sharing interface and system using the same, and more particularly to a shared buffer control of a system having an interface shared by a plurality of partitions.

  FIG. 4 shows the configuration of a system having an interface shared by a plurality of partitions of the related art. A node S01 consisting of processors S21, S22, S23, S24, memories S31, S32, S33, S34 and a disk S41, a node S02 consisting of processors S25, S26, S27, S28, memories S35, S36, S37, S38 and a disk S42. Is a system connected via crossbars S11 and S12.

  As a partition dividing method, for example, a partition 1 is constituted by processors S21, S22, S25, S26, memories S31, S32, S35, S36, and a disk S41, and processors S23, S24, S27, S28, memories S33, S34, S37, When the partition 2 is configured by S38 and the disk S42, the crossbars S11 and S12 have a configuration in which a buffer shared by the data of the partition 1 and the data of the partition 2 exists.

  In the shared buffer, the data of the partition 1 and the partition 2 transferred from the node S01 to the node S02 are mixed in the shared buffer S11T on the transmission side of the crossbar S11 and the shared buffer S12R on the reception side of the crossbar S12. Similarly, in the shared buffer S12T on the transmission side of the crossbar S12 and the shared buffer S11R on the reception side of the crossbar S11, data of partition 1 and partition 2 transferred from the node S02 to the node S01 are mixed.

  In such a system configuration where a buffer shared by multiple partitions exists, if a timeout failure occurs in a certain partition, the transfer data of the failed partition stagnate in the shared buffer. The stored transfer data of other partitions may also be stagnant, resulting in a multi-partition failure. This situation will be described with reference to FIG.

  FIG. 5 illustrates in detail the crossbar portion (crossbar N01 in the illustrated example) of one side node in the system shown in FIG.

  In the shared buffer N50 of the partition sharing unit where data of a plurality of partitions shown in FIG. 5 is mixed, for example, when the top entry is transfer data addressed to the processor N40 of the partition 2, the read control unit N51 of the shared buffer N50 A transmission request is issued to the arbitration unit N42. Normally, the permission response is returned from the arbitration unit N42 to the read control unit N51, the transfer data is read from the shared buffer N50, and is transmitted to the processor N40 via the destination control unit N52, the selector unit N41, and the buffer N43. .

  At that time, there is also transfer data addressed to the processor N40 from the processor N30 of the partition 2, the read control unit N32 of the buffer N31 issues a transmission request to the arbitration unit N42, and the arbitration unit N42 performs arbitration. When the permission response is returned only to the read control unit N32, the selector unit N41 selects only the request from the processor N30, and the state in which the request from the shared buffer N50 is not selected is continued.

  When read permission is not issued from the arbitration unit N42 to the read control unit N51 in this way, for example, if transfer data addressed to the processor N20 in the partition 1 is stored as data of the subsequent entry in the shared buffer N50, read control is performed. There is a problem that a transmission request cannot be issued from the unit N51 to the arbitrating unit N23, and the transfer data of the healthy partition 1 causes a timeout failure due to a stagnation of the transfer data of the partition 2.

  In relation to this, in Patent Document 1, regarding a buffer control shared by a plurality of nodes, there is a shared buffer for reducing the influence on the entire system when a failure occurs in one node and re-installation after the node is restored. It is disclosed. This shared buffer is provided corresponding to each node which is a computer, a register which holds information indicating whether or not the corresponding node is incorporated in the system and can operate normally, and data output from the node Corresponding to a register including a plurality of data fields to be stored sequentially and a memory including a configuration field holding information indicating whether the data is invalid, and the register being in a state in which the register can be normally operated Control means for invalidating a config field corresponding to a data field in which data is stored when a failure occurs in a node.

Japanese Patent Laid-Open No. 11-096035

  In order to avoid a multi-partition failure in a system having an interface shared by a plurality of partitions according to the related technology shown in FIGS. 4 and 5, a configuration without an interface shared by a plurality of partitions or a small number of partitions is used. In some cases, it was necessary to minimize partition sharing as much as possible, such as by owning a buffer for each partition.

  Also, even in the case of buffer control shared by multiple partitions, in the case where a failure occurs in a partition and a timeout is detected in particular, it remains in the shared buffer during failure processing of the failed partition. There was a case where a multi-partition failure occurred due to a time-out failure caused by data of a healthy partition that did not have a subsequent failure.

  Furthermore, even in the case of performing buffer control shared by a plurality of nodes as in Patent Document 1, in the case where there is a failure notification from the outside of the shared buffer, failure notification means must be provided for the shared buffer from the number of all nodes. There is a case where there is no failure notification or delay in detecting a timeout failure, and there is a possibility that data of a healthy partition is involved. In addition to such a background, in recent years there is a trend of server integration, processor multi-core, VM (Virtual Machine), IO (Input Output) virtualization, LSI (Large Scale Integration) integration into one chip. A situation has arisen in which it is necessary to have a part shared by a plurality of partitions.

  The object of the present invention is to solve the above problems, detect a stagnation state before time-out detection by a normal transaction, speed up buffer read control when a failure occurs, and trigger another time-out failure in one partition. It is an object to provide a buffer control system, method, and partition sharing interface and system using the same, which can reduce a multi-partition failure caused by a healthy partition being involved.

  To achieve the above object, a buffer control system for a partition sharing interface according to the present invention is a system for controlling a buffer of an interface shared by a plurality of partitions, and independently monitors a stagnation state of data in the buffer, A stagnation monitoring unit that detects the occurrence of a timeout failure due to the stagnation state of the data before the timeout detection by the transaction request and reply, and the stagnation monitoring unit caused a timeout failure due to the stagnation state of data in a certain partition And an address skip control unit for skipping an address in which data of the partition causing the time-out failure is stored when data is read from the buffer.

  The buffer control method for a partition shared interface according to the present invention is a method for controlling a buffer of an interface shared by a plurality of partitions, wherein a stagnation monitoring unit independently monitors a stagnation state of data in the buffer, and the data When the occurrence of a timeout failure due to a stagnation state is detected before timeout detection by a transaction request and reply, and the address skip control unit detects that a timeout failure has occurred due to a stagnation of data in a partition When reading data from the buffer, an address at which data of a partition causing the timeout failure is skipped.

  According to the present invention, in shared buffer control of an interface shared by a plurality of partitions, it is possible to detect a stagnation state before time-out detection by a normal transaction by independently monitoring stagnation data in the shared buffer by timer monitoring. . In addition, read control from the shared buffer speeds up buffer read control in the event of a failure by enabling only the data of a healthy partition to be read and skipped, and triggered by a timeout failure in one partition. Multi-partition failures that occur when other healthy partitions are involved can be reduced.

It is a block diagram which shows the structure of the buffer control system of the partition shared interface which concerns on embodiment of this invention. It is a figure which shows an example of the transfer data stored in the shared buffer of FIG. 3 is a timing chart for explaining the operation of the buffer control system of the partition sharing interface shown in FIG. 1. It is a block diagram which shows the structure of the system which has the interface which multiple partitions of related technology share. FIG. 5 is a block diagram showing a detailed configuration of a crossbar portion of one side node in the system shown in FIG. 4.

  Next, embodiments of a buffer control system and method, a partition sharing interface using the same, and a system according to the present invention will be described in detail with reference to the drawings.

  FIG. 1 shows a configuration of a buffer control system according to the present embodiment. The buffer control system shown in the figure controls a shared buffer B08 of an interface (shared part) shared by a plurality of partitions as shown in FIGS. 4 and 5, for example, and includes a packet analysis part B01, a register, etc. First storage means B02 to B05, second storage means B06, B07, read pointer B13, stagnation monitoring timer (stagnation monitoring part) B14, and address skip control part B15.

  The packet analysis unit B01 analyzes the packet B00 input to the shared buffer B08 of the interface (shared unit) shared by a plurality of partitions, and passes through the first storage unit B02 to B05 and the second storage unit B06, B07. And stored in the shared buffer B08. Here, the input packet B00 is composed of a plurality of flits in which one packet is not a fixed length, and in addition to flit data, one flit includes Valid information indicating the validity of the flit, packet partition as its control information. Information indicating a number and information indicating the number of flits of a packet are stored directly or information capable of calculating them. Among these, information indicating the partition number of the packet and information indicating the number of flits of the packet are stored in any one of a plurality of flits constituting one packet, for example, the first flit or other flits. ing. Which flit is stored differs depending on the format of the input packet B00.

  In the present embodiment, the packet analysis unit B01 stores the flit data itself in the first storage means B05 among the input packets B00 in which one packet is composed of a plurality of flits, and also shows Valid information indicating the validity of the flits. Are stored in the first storage unit B02, information indicating the packet partition number is stored in the first storage unit B03, and information indicating the number of packet flits is stored in the first storage unit B04, either directly or calculated.

  Here, when the information indicating the packet partition number and the information indicating the number of flits of the packet are in the first flit of a plurality of flits constituting one packet, the validity of the flits as shown in FIG. Is stored in the second storage unit B06 from the first storage unit B02, and the flit data is stored in the second storage unit B07 from the first storage unit B05, and indicates the partition number of the packet. In addition, information indicating the number of packet flits and a time difference of a predetermined time (1T) are provided and stored in the shared buffer B08.

  On the other hand, depending on the format of the packet, when the information indicating the partition number of the packet and the information indicating the number of flits of the packet are not in the first flit of a plurality of flits constituting one packet, A time difference corresponding to the difference from the frit for storing information is added and stored in the shared buffer B08.

  In the present embodiment, when the input packet B00 is stored in the shared buffer B08 via the first storage means B02 to B05 and the second storage means B06 and B07, when read from the shared buffer B08, Valid information B09 indicating the validity of the flit, information B10 indicating the partition number of the next packet, information B11 indicating the number of flits of the next packet, and flit data B12 of the current first entry are stored in the shared buffer B08. A buffer configuration for reading from the same entry at the same address can be realized.

  The read pointer B13 designates the address of the entry storing the flit read from the shared buffer B08 as the output packet B16.

  The stagnation monitoring timer B14 operates before the time-out is detected by a normal transaction request / reply by operating the timer when the read flit is valid and the read pointer B13 does not change when reading from the shared buffer B08. Detect stagnation timeout. Here, it is assumed that the timer value can be variably set to an appropriate value shorter than a timeout due to a normal transaction request reply.

  The address skip control unit B15 detects information B10 indicating the partition number of the next packet, information B11 indicating the number of flits of the next packet, and flit data of the current first entry when the stagnation monitoring timer B14 detects a stagnation timeout. Based on B12, an address skip value (+ x) is determined and notified to the read pointer B13.

  Next, a method for controlling the shared buffer according to the present embodiment will be described with reference to FIG. This figure shows the buffer control when the input packet B00 is stored in the shared buffer B08 and the output packet B16 is transmitted from the shared buffer B08.

  First, when an interface (shared unit) shared by a plurality of partitions receives the input packet B00, the packet analysis unit B01 analyzes the packet contents and sends Valid information indicating the validity of the flit to the first storage unit B02, the packet Are stored in the first storage means B03, the number of flits of the packet is stored in each storage means of the first storage means B04, and the flit data is simultaneously stored in the first storage means B05.

  Further, before storing in the shared buffer B08, information indicating the validity of the flit and flit data are stored in the second storage means B06 and B07, respectively, with a time difference from the partition number and the number of flits. By storing in the shared buffer B08, the current flit validity information B09, the current flit data B12, the next packet partition number B10, and the next packet flit number B11 are read from the same entry in the shared buffer B08. It becomes possible.

  Furthermore, in the present embodiment, the current flit validity information B09 indicates the flit valid state, and the stagnation for monitoring the flit that is not read from the shared buffer B08 but is stagnant even though the read pointer B13 is readable. A monitoring timer B14 is provided.

  As a result, when the stagnation monitoring timer B14 detects a timeout, the address skip control unit B15 generates a stagnation timeout from the information of the partition number B10 of the next packet, the number of flits B11 of the next packet, and the current flit data B12. Reading from the shared buffer B08 is performed by skipping the address where the partition data is stored and calculating the address where other healthy partition data is stored and notifying the read pointer B13.

  By doing so, it is possible to invalidate the partition data stagnating before the time-out detection by the transaction request / reply, and to read only the subsequent healthy partition data, thereby increasing the speed.

  FIG. 2 shows an example of transfer data stored in the shared buffer B08 of FIG.

  The transfer data is stored in order from entry 0 of shared buffer B08. The example of FIG. 2 shows that the data is stored in the order of flits 1 and 2 of partition 2, flits 1 and 2 of partition 1, flits 1, 2 and 3 of partition 2, and flit 1 of partition 1.

  Each entry includes a Valid field indicating the valid state of a flit, a field indicating the partition number of the next packet, a field indicating the number of flits of the next packet, and a field storing flit data. Therefore, for example, in entry 1, together with flit 2 of partition 2, Valid information indicating the validity of the flit, partition number 1 which is the partition number of the next packet, and flit number 2 which is the number of flits of the next packet are the same. It is the structure of the buffer stored in the entry.

  Here, when data is stored in the shared buffer F00 in the flit order as shown in FIG. 2, the flit 1 of the partition 2 is stored in the first entry 0 of the shared buffer F00, and the flit detects a stagnation timeout. Will be described.

  The stagnation monitoring timer B14 monitors the state where the flit 1 of the partition 2 of the entry 0 is valid and the address skip control unit B15 does not issue a pointer update instruction to the read pointer B13, and operates the timer. It is assumed that the timer value can be variably set to an appropriate value shorter than a timeout due to a normal transaction request reply.

  In this state, when a stagnation timeout is detected by the timer operation by the stagnation monitoring timer B14, the detection pulse is notified to the address skip control unit B15. Then, immediately after the stagnation timeout is detected, the address skip control unit B15, based on the information from the current flit data B12 and the partition number of the current packet and the number of flits, detects flit 1 of partition 2 and flit 2 of partition 2 that have been timed out. The address is skipped by the number of stored entries (+2 entries) and notified to the read pointer B13.

  The address skip control unit B15 increments by 1 so that the address indicates entry 3 as usual because the next partition number B10 is the partition number 1 which is not the timeout detection partition in the subsequent read of entry 2.

  Further, in the read of entry 3, since the partition number B10 of the next packet is the partition number 2 which is the partition in which the stagnation timeout is detected, not immediately after the timeout is detected, the address skip control unit B15 sets the number of flits in the next partition to 3. A value incremented by 1 is calculated, the address is skipped by +4 so that the address indicates entry 7, and the read pointer B13 is notified.

  By continuing such an operation, the stored data of the entries 0, 1, 4, 5, and 6 in which the partition data in which the stagnation timeout is detected are invalidated as fast as possible, and the healthy partition data is stored. The stored data of the entries 2, 3, and 7 can be read out without affecting.

  In FIG. 3, when a stagnation monitoring timeout is detected in the shared buffer, the address of the shared buffer in which the partition data is stored is skipped, and the address in which the healthy partition data is stored is calculated and read from the shared buffer. The timing chart of the address skip at the time is shown.

  When the flit data T03 of the flit of the partition 2 stored in the entry 0, that is, Dat0 is stagnated without being read during the time t0 to t4 in FIG. 3, the buffer read address T06 indicates Adr0. In this state, the partition number T04 of the next packet is 2, and the flit number T05 of the next packet is 2.

  The flit in this state continues to stagnate, and at time t4, the timeout error T01 of the partition 2 is detected by the operation of the stagnation monitoring timer B14, and the timeout error detection pulse T02 is notified to the address skip control unit B15. Accordingly, when the address skip control unit B15 detects the timeout error detection pulse T02, the address skip control unit B15 calculates how many flits the packet including the flit is composed of from the current flit data T03. In this case, since the partition number T04 of the next packet indicates 2 and the number of flits T05 of the next packet is 2, the address skip control unit B15 increases the address skip amount +2 (address skip from Adr0 to Adr2). At time t5, Adr2 with the address Adr0 set to +2 is designated as the read pointer B13, thereby reading the flit flit data T03 stored in the entry of the address Adr2 of the shared buffer B08, that is, Dat2.

  Subsequently, since the partition number T04 of the next packet is 1, the address skip control unit B15 designates Adr3 with the address +1 as usual at the time t6 as the read pointer B13, thereby the address Adr2 of the shared buffer B08. The flit data T03 of the flit stored in the entry, that is, Dat3 is read.

  At this time, while the timeout error T01 of the partition 2 is being detected, the partition number T04 of the next packet indicates 2 and the number of flits T05 of the next packet is 3, so the address skip control unit B15 +3 (address skip from Adr3 to Adr7) is determined, and at time t7, Adr7 with the address Adr3 set to 3 + 1 is designated as the read pointer B13, thereby the flit stored in the entry of the address Adr7 of the shared buffer B08 The flit data T03, that is, Dat7 is read.

  Thereafter, similarly, when the timeout error T01 of the partition 2 is being detected and the partition number T04 of the next packet indicates 2, the address skip control unit B15 calculates an address with the number of flits T05 + 1 of the next packet. By doing so, it is possible to skip to the address of the entry in which a healthy partition flit is stored.

  As described above, in this embodiment, in order to reduce the multi-partition failure, a time-out failure can be predicted in advance by monitoring the stagnation state of the data in the shared buffer by using a timer. The invalidation process has started. Further, by providing an independent stagnation timeout detection means in the sharing unit without providing a failure notification means for each partition, it becomes easy to deal with a case where there are many shared partitions.

  Furthermore, in the present embodiment, in the read control from the shared buffer, the speed is increased by skipping and reading the address where the data of the partition in which the stagnation timeout failure has occurred in the shared buffer is stored. This can reduce the case where transfer data of a partition is affected and causes a timeout failure.

  That is, in this embodiment, when a timeout failure occurs in a certain partition in the buffer control of an interface shared by multiple partitions, the data in the non-failed partition is affected among the data remaining in the buffer shared by the partition. The data in the failed partition can be invalidated at high speed without giving as much as possible, and the case of a multi-partition failure can be reduced.

  As described above, according to the present embodiment, in the shared buffer control of the interface shared by a plurality of partitions, the stagnation data in the shared buffer is independently monitored to detect the timeout before the normal transaction timeout detection. A stagnation state can be detected. In addition, read control from the shared buffer speeds up buffer read control in the event of a failure by enabling only the data of a healthy partition to be read and skipped, and triggered by a timeout failure in one partition. Multi-partition failures that occur when other healthy partitions are involved can be reduced.

  The information used as the address skip condition is not limited to the partition number and the number of flits, but may be any information that can identify them. Even if the information is a packet of a protocol that is not located in the first flit, the same operation can be handled by adjusting the time difference from the flit data. Further, the number of partitions to be shared is not limited to the two partitions in the above embodiment, and information can be obtained from time-out detection and the partition number, so that it is possible to deal with a case where a larger number of partitions are shared.

  The above-described buffer control system is not particularly limited in hardware and software configuration, and any buffer control system can be used as long as it can realize the functions (means) of each unit described above. For example, the circuit may be configured independently for each function of each unit, or may be configured by integrating a plurality of functions into a single circuit.

  Further, when at least a part of the functions of each unit constituting the buffer control system is realized using a program code, the program code and a recording medium for recording the program code are included in the scope of the present invention. In this case, when the above functions are realized in cooperation with other software such as an operating system (OS), those program codes are also included.

  Although the present invention has been described with reference to the embodiments, the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

  INDUSTRIAL APPLICABILITY The present invention can be used for a buffer control system, a method, a partition sharing interface using the buffer control system, and a system.

B01 Packet analysis unit B02 to B05 First storage unit B06, B07 Second storage unit B08 Shared buffer B13 Read pointer B14 Stagnation monitoring timer (stagnation monitoring unit)
B15 Address skip controller

Claims (12)

In a system that controls an interface buffer shared by multiple partitions,
A stagnation monitoring unit that independently monitors the stagnation state of the data in the buffer, and detects the occurrence of a timeout failure due to the stagnation state of the data before the timeout detection by a transaction request and reply; and
When the stagnation monitoring unit detects that a timeout failure has occurred due to the stagnation state of data in a partition, when reading data from the buffer, the address storing the partition data causing the timeout failure is skipped. A buffer control system comprising: an address skip control unit that performs The data is composed of a plurality of flits constituting a packet,
When storing the data in the buffer, the flit data, information indicating the validity of the flit, the partition number of the packet, and the number of packet flits are stored in a first storage means, and the first storage The flit data stored in the means and the information indicating the validity of the frit are stored in the second storage means, the partition number of the packet stored in the first storage means, the number of packet frits, and the second storage 2. The buffer control system according to claim 1, wherein flit data stored in the means and information indicating the validity of the frit are stored in the buffer.   The address skip control unit reads information indicating validity of the current flit, current flit data, the partition number of the next packet, and the number of flits of the next packet from the same entry of the buffer. The buffer control system according to claim 2.   The stagnation monitoring unit includes a stagnation monitoring timer in which a timer value is variably set to a value shorter than a timeout due to a transaction request reply, and the stagnation monitoring timer indicates that the frit validity information indicates a frit valid state. 4. The buffer control system according to claim 3, wherein a timeout of a flit that is stagnant without being read from the buffer although the read pointer of the buffer is readable is detected.   When the occurrence of the timeout is detected by the stagnation monitoring timer, the address skip control unit determines the address of the address from the partition number of the next packet, the number of flits of the next packet, and information on the current flit data. 5. The buffer control system according to claim 4, wherein a skip value is determined and notified to a read pointer of the buffer. In a method of controlling an interface buffer shared by multiple partitions,
The stagnation monitoring unit independently monitors the stagnation state of the data in the buffer, detects the occurrence of a timeout failure due to the stagnation state of the data before the timeout detection by the transaction request and reply,
When the address skip control unit detects that a timeout failure has occurred due to a stagnation of data in a partition, when reading data from the buffer, it skips the address where the partition data causing the timeout failure is stored. And a buffer control method. The data is composed of a plurality of flits constituting a packet,
When storing the data in the buffer, the flit data, information indicating the validity of the flit, the partition number of the packet, and the number of packet flits are stored in a first storage means, and the first storage The flit data stored in the means and the information indicating the validity of the frit are stored in the second storage means, the partition number of the packet stored in the first storage means, the number of packet frits, and the second storage 7. The buffer control method according to claim 6, wherein flit data stored in the means and information indicating validity of the frit are stored in the buffer.   The address skip control unit reads information indicating validity of the current flit, current flit data, the partition number of the next packet, and the number of flits of the next packet from the same entry of the buffer. The buffer control method according to claim 7.   The stagnation monitoring unit uses a stagnation monitoring timer in which a timer value is variably set to a value shorter than a timeout due to a transaction request / reply, information indicating the validity of the frit indicates a frit valid state, and the read pointer of the buffer is The buffer control method according to claim 8, wherein a timeout of a flit that is stagnant without being read from the buffer despite being readable is detected.   When the occurrence of the timeout is detected by the stagnation monitoring timer, the address skip control unit determines the address of the address from the partition number of the next packet, the number of flits of the next packet, and information on the current flit data. The buffer control method according to claim 9, wherein a skip value is determined and notified to a read pointer of the buffer. An interface shared by multiple partitions,
6. A partition sharing interface using the buffer control system according to claim 1.   A system using the partition sharing interface according to claim 11.

Images