KR19990030724A - Transaction control method and apparatus of multiprocessor system - Google Patents

Abstract

The present invention relates to a method and apparatus for controlling a transaction in a multiprocessor system. The present invention relates to two or more nodes including a processor and a memory and sharing a local bus, and interconnecting each other using another local bus. Transactions driven on local buses are accessed in local memory by the type of transaction, current local directory state, and the address range in which transactions are executed, or by generating the transactions required for local buses, it is possible to efficiently process the driven transactions. You can implement multiple processors at a low cost. The transaction controller of such a multiprocessor system includes a plurality of processors and memory, and performs a transaction driven by a local bus in a multiprocessor system configured by connecting two or more nodes sharing a local bus with an interconnect bus. First transaction processing means for performing; Second transaction processing means for storing a transaction driven on the interconnection bus and driving the transaction on a local bus; This is accomplished by including a control unit that controls the state transition of the local directory for cache coherency of the local bus and the interconnect bus.

Description

Transaction control method and apparatus of multiprocessor system

The present invention relates to transaction control of a multiprocessor system, and more particularly, in a multiprocessor system in which another local bus is used to interconnect two local buses sharing a plurality of processors. The present invention relates to a transaction control method and apparatus of a multiprocessor system capable of accessing a transaction to a minimum.

In general, in a system of distributed shared memory architecture, since the shared memory is distributed to each processor node, the access time to the memory of its own node and the memory of another node is not constant.

More specifically, referring to the drawings, as shown in FIG. 1, which is a schematic block diagram of a multi-processing system having a distributed shared memory structure, a multi-processing system having a distributed shared memory structure includes a plurality of CPUs, cache memories, and the like. A plurality of nodes 10-14 having distributed shared memories 10a-14a, and an interconnection bus that interconnects each node 10-14 and requests and provides addresses and data to the corresponding node according to remote access. It consists of.

In the multi-processing system configured as described above, the memory access of the processor generated at a particular node is shorter than the time of accessing the memory of the local node.

Since memory access to the remote node is performed over the interconnect bus, it is relatively slower than the access time of the local memory. As a result, the memory access time of each processor is inconsistent, and thus data access by two or more nodes connected to the local bus maintains data consistency on the interconnect bus, and then performs data access between the nodes.

That is, in order to maintain data consistency between nodes in a system in which a plurality of nodes are connected and configured, each node is connected as one connection bus and data consistency is maintained by hardware on the connection bus. This structure is called a "cache coherent NUMA" structure.

However, the distributed shared memory system can interconnect a plurality of nodes through a connection bus, but when processing a transaction from one node to another, the data processing speed is slower than the bus speed of the actual processor, and the processor bus There was also a problem in that a lot of time was spent in converting a protocol to connect a bus with another bus.

Accordingly, the present invention interconnects two or more nodes including a processor and a memory, and shares a local bus, using another local bus, and transfers transactions generated from multiple processors to a local directory state, a local bus. An object of the present invention is to provide a method and apparatus for controlling a transaction in a multiprocessor system that efficiently processes the snoop result, the address range of the transaction, and the transaction type.

1 is a schematic structural diagram of a multiple processing system of a general distributed shared memory structure.

2 is a schematic structural diagram of a multiprocessor system for implementing the present invention.

Figure 3 is a block diagram showing a transaction control device of a multiprocessor system according to the present invention.

<Explanation of symbols for main parts of the drawings>

40 to 43: first through fourth processors 44: bus transaction control unit

46, 52: first and second queue processor 47: memory controller

48: memory module 49, 53: transponder

50: delay queue controller 51, 54: first and second requestor

55, 56: first and second data queue 57: local directory

58: local directory controller

In order to achieve the above object, a transaction control method of a multi-processor system according to an aspect of the present invention comprises a plurality of processors and memory, etc., connecting two or more nodes sharing a local bus, interconnection bus In the multi-processor system configured to perform the first step of detecting a transaction driven in the interconnection bus; When the driven transaction is a simple read operation that refers to its own memory area, when the local directory is clean and shared, an operation of reading data from its memory is performed, and when referring to another memory area, it is related to the local directory. A second process of driving the light back without; When the transaction refers to its own memory area and other memory areas, and a specific processor reads the address information, each node's processor stores the most recent address information in its memory regardless of the local directory. Performing a transaction; If the transaction is simply a write operation referring to its memory area, if the local directory is dirty, the transaction may include a fourth process of performing a write operation on the corresponding memory.

Here, in the second step, when a simple read operation refers to its own memory area and the local directory is dirty and a specific processor of the local node has changed information, another processor of the remote node reads the same address. When performing the write operation, the write back is performed by the specific processor, and a transaction is generated in which the local directory states of the processors of the remote node are transferred to the shared state.

In addition, in order to achieve the object as described above, a transaction control apparatus of a multiprocessor system according to another aspect of the present invention, the two or more nodes including a plurality of processors and memory, and share a local bus, A multiprocessor system configured to connect by a connection bus, comprising: first transaction processing means for executing a transaction driven by the local bus; Second transaction processing means for storing a transaction driven on the interconnection bus and driving the transaction on the local bus; And a controller for controlling a state transition of a local directory for cache coherency of the local bus and the interconnect bus.

Here, when the transaction generated in the interconnection bus accesses address information in the memory of the local node, the second transaction processing means requests a transaction from the local bus, and accesses address information in the memory of the remote node. The address information is deleted after the transaction ends.

The second transaction processing means may process a transaction according to a transaction type, a memory area, and a local directory state driven on the interconnection bus.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

2 is a schematic structural diagram of a multiprocessor system for implementing the present invention. As shown in the drawing, the first to fourth processors 21 to 24 and the first bus transaction control unit 25 that share the first local bus are received from each processor 21 to 24 to receive input / output requests. It is transmitted through an I / O bus connected to the 1 bus transaction control unit 25, and is connected to the PCI bridges 27 and 28 for interfacing the I / O bus and the PCI bus, and the first bus transaction control unit 25 to store data. And a first memory 26 to which a plurality of memories are mounted for reading. The first node 20 of the multi-process system configured as described above is symmetrically connected to the second node 30 through the third local bus. In this case, the second node 30 is configured in the same manner as the first node 20, and only reference numerals are changed.

That is, up to four Pentium Pro processors are usually mounted on a CPU local bus, and a multi-process implementation using more than one Pentium Pro processor is connected to another node using the interconnection bus as described above.

At this time, when a transaction occurs in the processor of one node 20, 30, the first and second bus transaction controllers 25 and 35 determine which memory corresponds to address information of the transaction, and in case of local memory access. Reads or writes data in the first and second memories 26 and 36 which are local memories.

In addition, even in the case of a transaction to a remote node, the first and second bus transaction controllers 25 and 35 may determine this, and may include a transaction type, a memory area, a local directory state, and a snoop result of the local bus. ), Each performs different transactional access.

The control device for transaction processing in such a multi-processing system is shown in FIG. A first queue processor 46 for storing transactions occurring in a plurality of processors 40 to 42 sharing a local bus and determining whether an address of the generated transaction is a local memory or a remote memory; A memory controller 47 connected to the first queue processor 46 to patch data from the memory module 48 and to drive a response signal on the local bus when the transaction is local access; Delay queue controller 50 for temporarily storing a transaction when the address transmitted from the first queue processor 46 is a remote memory transaction and is a delayable transaction: a transaction from the delay queue controller 50 or the first queue processor 46. Receiving the first requestor 51 for requesting a transaction through the interconnection bus; A second queue processor 52 for storing and managing transactions originating from other nodes via an interconnect bus; A second responder (53) for driving a response signal for the transaction delivered to the second queue processor (52) to the interconnect bus; A second requester 54 which receives the transaction from the second queue processor 52 and requests the bus to use on the local bus and drives the transaction; A local directory controller 58 that controls the state of the local directory 57 and sends the memory state of the other node to the first queuer 46 and the second queuer 52 for cache coherency; And a first data queue 55 and a second data queue 56 connected to the local bus and the interconnect bus and storing data driven or driven from the local bus and interconnect bus.

Here, reference numeral 44 is a bus transaction controller (BTC).

In the multiprocessor system configured as described above, first, when a plurality of processors 40 to 43 generate a transaction, the first queue processor 46 stores the transaction and then determines local access or remote access. In this case, in the case of local access, the memory controller 47 patches the corresponding data for the transaction directly from the memory module 48 and provides it to the processor.

However, in the case of remote access, a request is made to the remote node through the first requestor 51. At this time, a transaction which can be delayed among the transactions continuously generated is temporarily stored in the delay queue controller 50 and then processed.

On the other hand, a transaction generated in a remote node may refer to a memory area of a local node through an interconnection bus. Thus, transaction processing driven in the interconnection bus is performed by transaction type, current local directory state, and transaction execution. The address range is used to access the local memory or generate a transaction required for the local bus. Such a transaction operation is performed in the second queue processor 52 according to the following Table 1a and Table 1b, which will be described in more detail with reference to the following.

transaction Access division Local directory status Action Memory lead (MR) Local memory CleanSharedDirty MR Request for local memory (Ready 2) Remote memory N.A MR Request for Remote Address (2nd requester) with expecting Write Back Memory Lead Invalid (MRI) Local memory CleanSharedDirty MRI Request for localmemory (Second Requester) MRI Request for localmemory (Second Requester) N.A Remote memory N.A Request for Remote memory (2nd request)

transaction Access division Local directory status Action Memory leads (MW) Local memory CleanSharedDirty N.AN.AWrite to local memory (memory controller) Remote memory N.A N.A Implement Memory Writeback (IMW) Local memory CleanSharedDirty N.AN.AWrite to local memory (memory controller) Remote memory N.A N.A

Here, "N.A" represents Not Allowed.

In addition, the local directory state indicates a case where another node has its own memory in a system composed of a plurality of nodes, and performs cache coherency between nodes, and has three states as follows.

In the clean state, another node does not have the corresponding address information. In the shared state, another node has the corresponding address information, and this information indicates the same state as the contents of local memory. ) Indicates that another node has the address information and has more recent address information than the contents of local memory.

In this state, a case of processing a memory read MR and a memory read invalidate (hereinafter, abbreviated as "MRI") transaction generated in a predetermined processor as described in Table 1a will be described. As follows.

First, in the case of a transaction (MR) that simply reads the memory area of the local node, it is a case of simply reading its memory when the local directory is clean. At this time, since the local directory is in a clean state, but the processor of the local node is caching, the corresponding address value is read from the local bus through the second requester 54.

If the local directory indicating that both the local node and the remote node are caching to the shared state is shared, the desired address value is immediately read from the memory module 48 via the memory controller 47.

If the local directory is dirty, writeback is performed by the processor having the address information even if another processor accesses the same address in the remote node because the remote node has the corresponding address information. do.

At this time, since each processor of the remote node transitions to the shared state, the processor of the local node should be informed that it is stored with the latest address information. This transactional operation is called Implement Write Back (IMW).

In the case of remote access, an IMW is generated regardless of the local directory. That is, the local node has dirty address information from the remote node, and the remote node performs the write back operation before reading the address information.

On the other hand, in an MRI transaction for reading the corresponding address information about a transaction originating from a specific processor, if the local directory is clean and shared in the case of local access, the cache line is stored to store the remote node on the local bus of the local node. Request and invalidate the cache line. At this time, the local directory state transitions to a dirty state.

If the local directory is dirty, a specific processor is caching dirty at the remote node. If another processor in the remote node starts MRI, the caching processor performs a write back, and then stores the stored address information to another processor. Inform.

Next, as shown in Table 1b, a transaction in which a specific processor writes corresponding address information about a transaction generated is described in detail as follows.

First, in the case of local access in a memory write transaction, if the local directory state is a clean state, the transaction is not generated because the remote node does not cache the address.

Also, if the local directory is shared, the remote node is also caching the address shared, so no transaction occurs.

If the local directory is dirty, the write operation is performed to the local memory. In this case, the local directory is changed to the clean state because the remote node invalidates the caching contents.

On the other hand, if an MRI transaction occurs for a local address at a remote node to read address information, and a memory read request occurs at another processor in the remote node for the same address, a write back is performed in the remote node and each processor is shared. As it transitions. At this time, it is necessary to change the local directory to a shared state. That is, when an IMW transaction occurs when the local directory state is dirty, the most recent information is stored in the local memory. At this time, since the remote node keeps caching in the shared state even after the write operation is completed, the local directory state transitions to the shared state.

However, if the local directory is clean, no writeback transaction occurs because the remote node is not caching the address, and even if the local directory is shared, the remote node is caching the address shared. Therefore, no transaction occurs.

In the case of remote access, since a write operation for the address of the remote node is stored in its memory area, such a transaction is not transferred to the local node.

As described in detail above, the present invention includes a processor, a memory, and the like, wherein two or more nodes sharing a local bus are interconnected using another local bus, and are driven on the interconnected local bus. Transactions can be processed efficiently by accessing the local memory or generating the necessary transactions on the local bus by the transaction type, the current local directory state, and the address range in which the transaction is performed.

Claims (5)

1. A multiprocessor system comprising a plurality of processors and memory and sharing two or more nodes sharing a local bus, comprising: a first step of detecting a transaction driven on the interconnect bus; When the driven transaction is a simple read operation that refers to its own memory area, when the local directory is clean and shared, an operation of reading data from its memory is performed, and when referring to another memory area, it is related to the local directory. A second process of driving the light back without; When the transaction refers to its own memory area and other memory areas, and a specific processor reads the address information, each node's processor stores the most recent address information in its memory regardless of the local directory. Performing a transaction; And a fourth process of performing a write operation on a corresponding memory when the local directory is dirty when the transaction is simply a write operation by referring to a memory area of the transaction. 2. The method of claim 1, wherein in the second process, when a simple read operation refers to its own memory area and the local directory is dirty and a specific processor of the local node has changed information, another processor of the remote node has the same address. And performing a read operation with respect to the processor, write back is performed by the specific processor, and a transaction is generated in which local directory states of the processors of the remote node transition to a shared state. 1. A multi-processor system comprising a plurality of processors and a memory and sharing two or more nodes sharing a local bus, the interconnection bus comprising: first transaction processing means for executing a transaction driven on the local bus; Second transaction processing means for storing a transaction driven on the interconnection bus and driving the transaction on the local bus; And a controller for controlling a state transition of a local directory for cache coherency of the local bus and the interconnect bus. 4. The method of claim 3, wherein the second transaction processing means requests a transaction from the local bus when the transaction generated on the interconnection bus accesses address information in the memory of the local node, and sends address information to the memory of the remote node. In case of access, the transaction control apparatus of the multi-processor system, characterized in that to delete the address information after the end of the transaction. 5. The apparatus of claim 4, wherein said second transaction processing means processes a transaction based on a transaction type, a memory area, and a local directory state driven on an interconnect bus.