WO1999046681A1

WO1999046681A1 - Cache coherence unit for interconnecting multiprocessor nodes having pipilined snoopy protocol

Info

Publication number: WO1999046681A1
Application number: PCT/US1999/005523
Authority: WO
Inventors: Wolf-Dietrich Weber
Original assignee: Fujitsu Limited
Priority date: 1998-03-12
Filing date: 1999-03-12
Publication date: 1999-09-16
Also published as: US20010013089A1; JP2001525095A; US6631448B2

Abstract

The present invention consists of a cache coherence protocol within a cache coherence unit for use in a data processing system. The data processing system is comprised of multiple nodes, each node having a plurality of processors with associated caches, a memory, and input/output. The processors within the node are coupled to a memory bus operating according to a 'snoopy' protocol. This invention includes a cache coherence protocol for a sparse directory in combination with te multiprocessor nodes. In addition, the invention has the following features: the current state and information from the incoming bus request are used to make an immediate decision on actions and next state; the decision mechanism for outgoing coherence is pipelined to follow the bus; and the incoming coherence pipeline acts independently of outgoing coherence pipeline.

Description

1 CACHE COHERENCE UNIT FOR

INTERCONNECTING MULTIPROCESSOR NODES

HAVING PIPELINED SNOOPY PROTOCOL

CROSS-REFERENCE TO RELATED APPLICATIONS

This application is related to co-pending U.S. Patent Application Serial No. 09/003,7771, entitled "Memory Protection Mechanism For A Distributed Shared

Memory Multiprocessor With Integrated Message Passing Support," filed on January

7, 1998; and co-pending U.S. Provisional Patent Application Serial No. 60/084,795, entitled "Cache Coherence Unit For Interconnecting Multiprocessor Nodes Having A

Sparse Directory With Enhanced Replacement Control," filed on May 8, 1998; which

are hereby incorporated by reference.

2 BACKGROUND OF THE INVENTION

1. Field of Invention

The present invention generally relates to cache coherence for multiprocessor data processing systems, and more particularly to cache coherence for a plurality of

multiprocessor nodes, each node having a snoopy bus protocol.

2. Discussion of the Background Art

Multiprocessor architectures are classified according to types of address space

and memory organization. Address space architecture classifications are based upon

the mechanism by which processors communicate. Processors communicate either by

explicit messages sent directly from one processor to another or by access through shared-memory address space. The first classification is called a message passing

architecture while the second is a shared-memory architecture.

Memory organization is classified as centralized or distributed. In a

centralized organization memory system, the entire memory is located concentrically

or symmetrically with respect to each processor in the system. Thus, each processor

has equivalent access to a given memory location. In a distributed organization

system, on the other hand, each processor within the multiprocessor system has an

associated memory that is physically located near the processor; furthermore, every

processor has the capability of directly address its own memory as well as the remote

memories of the other processors. A distributed, shared-memory system is known as

a distributed shared-memory (DSM) or a non-uniform memory access (NUMA)

architecture. DSM architecture provides a single shared address space to the programmer where all memory locations may be accessed by every processor. As

there is no need to distribute data or explicitly communicate data between the 3 processors in software, the burden of programming a parallel machine is simpler in a DSM model. In addition, by dynamically partitioning the work, DSM architecture makes it easier to balance the computational load between processors. Finally, as

shared memory is the model provided on small-scale multiprocessors, DSM

architecture facilitates the portability of programs parallelized for a small system to a

larger shared-memory system. In contrast, in a message-passing system, the

programmer is responsible for partitioning all shared data and managing communication of any updates.

The prior art provides numerous examples of DSM architectures. However,

such systems communicate through high bandwidth buses or switching networks, and the shared-memory increases data latency. Latency is defined as the time required to access a memory location within the computer, and describes the bottleneck impeding system performance in multiprocessor systems. Latency is decreased in DSM systems

by memory caching and hardware cache-coherence.

Caching involves placing high-speed memory adjacent to a processor where

the cache is hardware rather than software controlled. The cache holds data and

instructions that are frequently accessed by the processor. A cache system capitalizes

on the fact that programs exhibit temporal and spatial locality in their memory

accesses. Temporal locality refers to the propensity of a program to again access a

location that was recently accessed, while spatial locality refers to the tendency of a

program to access variables at locations near those that were recently accessed.

Cache latency is typically several times less than that of main system memory.

Lower latency results in improved speed of the computer system. Caching is

especially important in multiprocessor systems where memory latency is higher because they are physically larger, but caching does introduce coherence problems 4 between the independent caches. In a multiprocessor system, it becomes necessary to ensure that when a processor requests data from memory, the processor receives the most up-to-date copy of the data to maintain cache coherence.

Protocols incorporated in hardware have been developed to maintain cache coherence. Most small-scale multiprocessor systems maintain cache coherence with a

snoopy protocol. This protocol relies on every processor monitoring (or "snooping")

all requests to memory. Each cache independently determines if accesses made by

another processor require an update. Snoopy protocols are usually built around a

central bus (a snoopy bus). Snoopy bus protocols are very common, and many small- scale systems utilizing snoopy protocols are commercially available.

To increase the processing power of computer systems, manufacturers have

attempted to add more processing units to existing systems. When connecting additional microprocessors to the main bus to help share the workload, processing

power is added linearly to the system while maintaining the cost-performance of the

uni-processor. In such systems, however, bus bandwidth becomes the limiting factor

in system performance since performance decreases rapidly with an increase in the

number of processors.

In order to overcome the scaling problem of bus-based cache coherence

protocols, directory-based protocols have been designed. In directory based systems,

the state of each memory line is kept in a directory. The directory is distributed with

memory such that the state of a memory line is attached to the memory where that line

lives. The caches are kept coherent by a point-to-point cache coherence protocol

involving the memory system and all the processor caches.

U.S. Patent No. 5,029,070 to McCarthy et al. discloses a method for maintaining cache coherence by storing a plurality of cache coherency status bits with 5 each addressable line of data in the caches. McCarthy et al. specifically rejects storing the plurality of cache coherency status bits in the global memory. A plurality of state

lines are hardwired to the bus master logic and bus monitor logic in each cache. The

state lines are ORed so that all the states of all the same type of cache coherency bits

in every cache except for the line undergoing a cache miss appear on the state line.

This allows the bus master to rapidly determine if any other cache has a copy of the line being accessed because of a cache miss.

U.S. Patent No. 5,297,269 to Donaldson et al. discloses a system for point-to- point cache coherence in a multi-node processing system where the coherency is

maintained by each main memory module through a memory directory resident on the

individual memory module. The memories and nodes are coupled together by means of a cross bar switch unit coupled point-to-point to one or more main memory

modules. The memory directory of each main memory module contains a plurality of

coherency state fields for each data block within the module. Each main memory

module maintains the coherency between nodes. The module queries its own directory upon each data transfer operation that affects the coherency state of a data

block.

Sequent (T. Lovett and R. Clapp, StiNG, "A CC-NUMA Computer System for

the Commercial Marketplace," Proceedings of the 23^rd International Symposium on

Computer Architecture, pages 308-317, May 1996) and Data General (R. Clark and K.

Alnes, "An SCI Interconnect Chipset and Adapter," Symposium Record, Hot Interconnects IV, pages 221-235, August 1996) disclose machines that interconnect

multiple quad Pentium Pro nodes into a single shared-memory system. These two

systems both utilize an SCI-based interconnect, a micro-coded controller, and a large

per-node cache. The use of the SCI coherence protocol prevents close coupling of the 6 inter-node coherence mechanism to the intra-node (snoopy) coherence mechanisms.

The mismatch between the two protocols requires the use of a large L3 (node-level)

cache to store the coherence tag information required by the SCI protocol, to correct

the mismatch of cache line size, and to adapt the coherence abstraction presented by

the processing node to that required by SCI. In addition, the complexity of the SCI

coherence protocol invariably leads to programmable implementations that are unable

to keep up with the pipeline speed of the processor bus, and that can only process one

request at a time. The result is a coherence controller that is large, expensive, and slow.

What is needed is an inter-node coherence mechanism that is simple, fast, and well-matched to the pipelined snoopy protocol. Such a mechanism can be very tightly

coupled to the processor bus and can thus achieve higher performance at lower cost.

7 SUMMARY OF THE INVENTION

This invention includes the cache coherence protocol for a sparse directory in combination with multiprocessor nodes, each node having a memory and a data bus

operating under a pipelined snoopy bus protocol. In addition, the invention has the

following features: the current state and information from the incoming bus request

are used to make an immediate decision on actions and next state; the decision

mechanism for outgoing coherence is pipelined to follow the bus; and the incoming coherence pipeline acts independently of the outgoing coherence pipeline.

The invention implements the cache coherence protocol within a cache

coherence unit for use in a data processing system. The data processing system is

comprised of multiple nodes, each node having a plurality of processors with associated caches, a memory, and input/output. The processors within the node are

coupled to a memory bus operating according to a "snoopy" protocol.

Multiple nodes are coupled together using an interconnection network, with

the mesh coherence unit acting as a bridge between the processor/memory bus and the

interconnection network. The mesh coherence unit is attached to the

processor/memory bus and the interconnection network. In addition, it has a

coherence directory attached to it. This directory keeps track of the state information

of the cached memory locations of the node memory. The mesh coherence unit follows bus transactions on the processor/memory bus, looks up cache coherence state

in the directory, and exchanges messages with other mesh coherence units across the

interconnection network as required to maintain cache coherence.

The invention incorporates close coupling of the pipelined snoopy bus to the

sparse directory. In addition, the invention incorporates dual coherence pipelines. 8 The purpose of having dual coherence pipelines is to be able to service network

requests and bus requests at the same time in order to increase performance. Finally,

the invention incorporates a coherence protocol where all protocol interactions have

clearly defined beginnings and endings. The protocol interactions end the transactions on a given line before the interaction of a new line may begin. This process is

achieved by the invention keeping track of all the transient states within the system.

9 BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 A is a block diagram illustrating a multiprocessor system having a

plurality of nodes connected via a mesh coherence unit to an interconnect;

FIG. IB is a block diagram of a node of FIG. 1 A;

FIG. 2A is a block diagram of an individual node of FIG. 1A in further detail;

FIG. 2B is a block diagram of a P6 segment of FIG. 2A; FIG. 3 is a block diagram illustrating one embodiment of a shared memory site of FIG. 1A;

FIG. 4 illustrates the state transitions of the cache coherency protocol during a

remote read miss;

FIG. 5 illustrates the state transitions of the cache coherency protocol for a

remote write miss with clean copies;

FIG. 6 is a block diagram of the mesh coherence unit; and

FIG. 7 is a block diagram illustrating the relationship of the TRAT, ROB, NI,

UP, and DP.

10

DETAILED DESCRIPTION OF THE INVENTION

Referring to FIGS. 1 A and IB, a plurality of nodes 100 are coupled to

Interconnect 110, which enables nodes 100 to share information either by a message-

passing mechanism, a shared-memory mechanism, or a hybrid of the two mechanisms.

In the preferred embodiment, up to four nodes 100 are coupled together in a shared-

memory mechanism to create a shared-memory site 120. The nodes are connected to

Interconnect 110 via a Mesh Coherence Unit 130. Each Mesh Coherence Unit 130 is

coupled to an associated Sparse Directory 140.

Referring now to FIGS. 2A and 2B, node 200 of the cluster shown in FIG. 1 A

is shown. Four processors 210, together with their associated caches 220, are coupled

to memory bus 230. In the present embodiment, memory bus 230 operates according

to a snoopy bus protocol. In addition, associated memory 240 and input/output 250 of

the processors are attached to bus 230. In the preferred embodiment, Quad-P6

segment (P6 segment) 260 contains standard high volume Intel processor-based SMP

nodes made up of four Pentium® Pro processors, up to 1 GByte of DRAM, and two

PCI buses for attaching I/O. The P6 segment 260 is shown in FIG. 2B.

The P6 segment 260 maintains coherency within the segment by the use of a

snoopy bus protocol. Within P6 segment 260, each associated cache 220 snoops or

monitors all transactions with main memory 240 by all other caches 220 within

segment 260. In this way, each cache 220 within the P6 segment 260 is aware of all

memory lines 270 within memory 240 that are transferred from main memory 240 to a

cache 220. Mesh Coherence Unit (MCU) 130 is coupled to both P6 segment 260 via memory bus 11

230 and to Interconnect 110. All inter-node communication is passed through MCU

130. P6 segment 260, together with MCU 130, makes up current node 200. MCU

130 coordinates the flow of instructions and data between current node 200 and the

other nodes 100 connected to Interconnect 110. MCU 130 maintains the cache

coherence between nodes 100 within shared-memory site 120 and extends the P6 bus

functions over the mesh to connect multiple nodes 100 together. Nodes 100, together

with the Interconnect 110, make up a cluster. The nodes within a cluster, both within

shared memory sites 120 and those outside the sites, may be located physically close

to one another or may be distributed at a distance. Within a site 120, coherency is

maintained between nodes 100 by the MCU 130 and coherency is maintained within

node 200 by the standard P6 snoopy bus protocol. In the preferred embodiment,

coherency within site 120 is maintained by hardware; however, those familiar to the

art will recognize that site coherency could also be maintained in software or

firmware.

As will be described in detail below, MCU 130 maintains inter-node

coherency by implementing a directory-based, cache coherence protocol. MCU 130

keeps track of cache lines accessed by remote nodes 100 and the cache line status

within current node 200 with sparse directory 140. The MCU 130 also supports

Message Passing and Memory Copy between nodes both inter- and intra-site. In the

preferred embodiment, MCU 130 is a single custom CMOS chip. Sparse coherence

directory 140 is stored in standard, off-the-shelf SRAM chips. In the present invention, only three 1 Mbit chips are required for a total directory size of less then

0.5 MByte. A comparable full-directory design with 6 bits of state for every cache

line would require 24 MByte of storage per GByte of main memory. 12 Now referring to FIG. 3, one embodiment of the present invention is shown.

Cache coherence within the system is maintained on two levels. On one level, cache

coherence is maintained within individual nodes 200 by a snoopy bus protocol. On

the second level, MCU 130 maintains cache coherence between nodes 100 by using an

invalidating, directory-based cache coherence protocol. Memory line 1900 is one

cache line's worth of data stored in memory 1910 (240) at a particular address. Home

node 1920 is the particular node where a cached memory line physically resides. A

memory line 1900 can be cached locally 1921 at the home node 1920 or remotely in

one or more processor caches 1922 or 1923. When a memory line 1900 is cached

remotely, the line is either unmodified (clean) 1923 or modified (dirty) 1922. Owner

node 1930 is the particular node that has control of memoiy line 1900 with the ability

to update the line. The owner node 1930 is said to own a dirty copy 1922 of the

memory line 1900. A remote node 1950 is said to have a clean copy 1923 of the

memory line 1900 and is referred to as a "sharer" node. A local node 1940 is the node

200 where a current memory request originates. Any node that is not the home or

local node is called a remote node 1950.

Each MCU 1911 (130) maintains a directory to track its own memory lines

1900 in use by remote nodes 1930, 1940, 1950. The home node directory 1960 (140)

tracks those nodes 200 that have requested a read-only copy 1923 of the cache line

1900 so that when a node wants to update a line 1900, the directory 1960 (140) knows

which nodes to selectively invalidate. In addition, the directory 1960 (140) tracks the

node, if any, that owns the memory line 1900. The home node 1920 knows all the

remote nodes 1950 within the shared-memory site 120 that have requested a read-only

copy of the home node's 1920 memory line 1900, and also the remote node 1930 that 13 has requested write access to the memory line 1900. When home node 1920 must

communicate with remote nodes 1930, 1950, it does so selectively based upon which

nodes have accessed the line rather than to all nodes within site 120. Thus, the directory-based cache coherence protocol achieves coherence by point-to-point

communication between nodes 100 rather than broadcast invalidation.

When processor 1941 requests memory line 1900, its cache 1942 is checked to

determine whether the data is present in the cache. If processor 1941 finds the data in

cache 1942 (a "cache hit"), cache 1942 transfers the data to processor 1941. If

processor 1941 does not find the data in cache 1942 (a "cache miss"), the processor

issues a request onto P6 memory bus 230 for the data. The bus request includes the

real address of memory line 1900 and the identification of home node 1920 for

memory line 1900. If memory line 1900 resides on a different node, MCU 1943 (130)

of local node 1940 generates a network request for the cache line 1900 across

Interconnect 110 to home node 1920. When the MCU at home node 1911 (130)

receives the network request, it stages the request onto home node's memory bus 1912

(230). The network request then snoops bus 1912 (230) to determine if memory line

1900 is cached locally 1921. If not cached, the network request obtains control of

memory line 1900 in the node's memory 1910 (240). Thus, the action of MCU 1911

(230) at home node 1920 is like the action of a local processor 210 accessing memory

line 1900 within node 200. Depending on the status of memory line 1900, MCU 1911

(230) will take the appropriate action to give the requested memory line 1900 to local

node 1940. If local node 1940 has requested an update of line 1900, local node 1940

becomes the owner node; if local node 1940 only requests read capacity, it becomes a

sharer. Only one node 200 may be the owner of a memory line 1900 at any one time. 14 When an owner node 1930 finishes accessing a memory line 1900, the owner

node 1930 must notify the home node 1920 that it no longer requires access to line

1900. If owner node 1930 has modified the data, line 1900 must be updated at home

node 1920. This process is termed a "writeback" of the data.

If home node 1920 needs to regain control of memory line 1900, line 1922

may be flushed back to home node 1920 or flushed forward to another remote node

1940 if the remote node requires access. Owner node 1930, upon receiving a flush

back request, returns the control of memory line 1900 to home node 1920. Upon

receiving a flush forward request, owner node 1930 transfers memory line 1900 to a

remote node 1940, which becomes the new owner node.

Referring now to FIG. 4, the state transitions upon a read of a memory line 270

is illustrated. When a remote read by a processor 210 results in a cache miss and the

local node 600 must obtain memory line 270 from home node 610. Local node 600

first determines that memory line 270 is not in cache 220 (read miss). Local MCU

130 requests a fetch 601 from home node 610 for memory line 270 using a fetch

request message. If memory line 270 is not dirty (owned by another node), home

node 630 returns with a data message 602.

Referring now to FIG. 5 the state transitions on write requests by a remote

node to the home node where one or more remote nodes have cached, read-only

copies of the requested line is illustrated. In this transition, only read-only copies are

cached and no node has ownership of the line at the beginning of the sequence. After

a cache miss, local node 720 sends the fetch exclusive request 721 to home node 730.

Home node 730 sends back the data and ownership of line 270 to local node 720 via

the data exclusive message 731. Simultaneously, home node 730 sends invalidation 15 instructions 732 to the remote nodes 740 with cached, read-only copies of memory

line 270. Remote nodes 740 immediately acknowledge the receipt of the invalidation

741 when the invalidation is received in Mesh Interface 350 of the remote nodes 740.

When home node 730 receives acknowledgments 741 from all remote nodes 740 with

cached copies, home node 730 notifies local node 720 with the done message 733. At

this point, local node 720 is the owner of line 270.

Referring now to FIG. 6, a block diagram of the Mesh Coherence Unit (MCU)

130 is shown. The Pipe Control (PC) 310 arbitrates for control of P6 memory bus 230

during different bus phases of Up Pipe (UP) 320 and Down Pipe (DP) 330. Requests

that arrive from P6 bus 230 are turned into network requests by DP 330 and are then

sent to Queue Interface (QI) 340. The network requests are dispatched over the mesh

by Mesh Interface (MI) 350. Remote requests from the mesh are received by MI 350

and passed to Network Interface (NI) 360, which generates either bus requests that are

passed to the UP 320 or network requests that are passed to the QI 340. UP 320

requests P6 bus 230 arbitration from PC 310, and UP 320 sources the requests onto

the P6 bus 230 once arbitration has been won by the PC 310. The Exported Real

Address Table (ERAT) 370 maintains cache coherence information while Temporary

Real Address Table (TRAT) 380 and Remote Outstanding Operations Buffer (ROB)

390 keep track of cache lines in transition.

ERAT 370, TRAT 380, and ROB 390 together make up the sparse directory

140. The sparse directory 140 is a set associative directory. The specific functions of

directory 140 are discussed in detail below.

Down Pipe (DP) 330 transforms the P6 bus requests into network requests.

DP 330 maintains coherence during a P6 bus request in three phases: the request 16 phase, the snoop phase, and the response phase. During the request phase, DP 330 records the P6 request phase signal in a history queue. The history queue holds up to

eight outstanding requests on the bus for a reply and nine requests of data transfer.

The request is held in the queue until the DP 330 is ready to process the request.

DP 330 processes the bus request during the snoop phase. During this phase,

the bus request address is sent to ERAT 370 for lookup and the ERAT data bus is

switched for an ERAT read. The bus requests are translated into network requests with the use of a Coherence Action Table. The network requests are prepared for

output to QI 340 and the ERAT 370, ROB 390, and TRAT 380 entries are updated as

required.

The Queue Interface (QI) 340 creates the packets for transfer to the remote

nodes and passes the completed packets to Mesh Interface (MI) 350. QI 340 receives

a request or command from either DP 330 or NI 360. QI 340 stores the request in an

output queue. If the request contains data, QI 340 must hold the request in abeyance

until it receives the associated data from DP 330. Once QI 340 receives the data, QI

340 consolidates the request and data together into a single packet. The transfer of

data usually occurs sometime after QI 340 receives the request. Requests to transfer

data are always bus 230 initiated requests, while requests initiated by NI 360 are

command requests only.

NI 360 receives packets from Interconnect 110 through MI 350. NI 360 decodes the

packet and determines whether to send the P6 messages to UP 320, send mesh

messages to QI 340, or to update ROB 390 and TRAT 380 entries. The detailed

operation of NI 360 is discussed below.

Up Pipe (UP) 320 receives bus requests from NI 360. UP 320 requests bus 17 arbitration from Pipe Control (PC) 310. Once PC 310 notifies UP 320 that arbitration

has been won, UP 320 sources the bus request onto P6 memory bus 230. Depending

upon the request received from NI 360, UP 320 may also deallocate ROB 390 or

TRAT 380 entries for the incoming transient request.

Referring now to FIG. 7, the interaction between ROB 390/TRAT 380

combination and UP 320, DP 330, and NI 360 is shown. When NI 360 receives a

network request, it reads ROB 390/TRAT 380 entry for the request to determine the

request's current transition state. NI 360 updates or deallocates the ROB 390/TRAT

380 entry as required by the request. During the snoop phase as described above, DP

330 reads the ROB 390/TRAT 380 entries to determine the current transient state of a

given memory line. If no ROB 390/TRAT 380 entry exits for a given line, and the

ROB 390/TRAT 380 is not full, DP 330 may allocate an entry within ROB 390/TRAT

380. Similarly, UP 320 reads the ROB 390/TRAT 380 for a given memory line. UP

320 may deallocate the entry in TRAT 380 depending upon the response that is

received from NI 360. UP 320, NI 360, and DP 330 are each pipelined to handle multiple bus transactions or network requests at the same time. Through the use of

the ROB 390 and TRAT 380, as described above, the UP 320, NI 360, and DP 330

coordinate their actions to operate independently and in parallel when the addresses of

the request they are processing are independent. In this manner, a high-performance,

coherence protocol processing engine is achieved.

The invention has been explained above with reference to a preferred

embodiment. Other embodiments will be apparent to those skilled in the art in light

of this disclosure. For example, the present invention may readily be implemented

using configurations other than those described in the preferred embodiment above. Additionally, the present invention may effectively be used in combination with 18 systems other than the one described above as the preferred embodiment. Therefore, these and other variations upon the preferred embodiments are intended to be covered by the present invention, which is limited only by the appended claims.

Claims

19 WHAT IS CLAIMED IS:

1. A data processing system comprising: a plurality of multiprocessor nodes;

a plurality of coherence units coupled to said multiprocessor nodes; and

a node interconnect coupled to said coherence units.

2. A cache coherence unit for use in a data processing system having multiple

nodes each having a plurality of processors coupled to a memory bus operating

according to a snoopy protocol, each processor having an associated cache memory

for caching information, said cache coherence unit having an address phase, a snoop

phase, and a response phase, and comprising:

a bus interface element coupled to said memory bus;

a coherence control element coupled to said bus interface element and coupled to said cache memory; and a directory coupled to said coherence control element for storing information

of locations in said cache memory.

3. The cache coherence unit as in claim 2, wherein said coherence control

element reads state information from said directory after said address phase and

during said snoop phase, and updates said state information in said directory before a

next snoop phase.

20 4. The coherence unit of claim 2 further comprising a protocol decision pipeline

coupled to said coherence control element for processing requests for cache memory from a remote node.

5. The coherence unit of claim 2 where said directory is a sparse directory.

6. The coherence unit of claim 2 wherein said directory has dual ports.

7. The cache coherence unit of claim 2, wherein said directory stores state

information characterizing a current transient state and a next transient state.

8. The cache coherence unit of claim 7, wherein said coherence control element

updates said information characterizing a current transient state.

9. The cache coherence unit of claim 7, wherein said coherence control element

updates said information characterizing a current transient state and then updates said

information characterizing a next transient state.

21 10. A method for maintaining cache coherence in a cache coherence unit in a data processing system having multiple nodes each having a plurality of processors coupled to a memory bus operating according to a snoopy protocol having an address

phase, a snoop phase, and a response phase, each processor having an associated

cache memory, comprising the steps of:

coupling a bus interface element to said memory bus;

coupling a coherence control element to said bus interface element and to said cache memory; and

storing information of locations in said cache memory locations in a directory

coupled to said coherence control element.

11. The method of claim 10, wherein said coupling step further comprises:

reading state information from said directory after said address phase and

before said snoop phase; and

updating said state information before a next snoop phase.

12. The cache coherence unit of claim 10, wherein said coherence control element

further comprises:

means for reading state information from said directory after said address

phase and before said snoop phase; and

means for updating said state information before a next snoop phase.

22 13. The cache coherence unit of claim 11, wherein said coherence control element further comprises:

means for updating information characterizing a current transient state;

means for reading information characterizing a next transient state; and means for updating said information characterizing a next transient state after

said means for updating information characterizing a current transient state

is completed.

23 14. A data processing system comprising:

a first node having a first memory coupled to a first memory bus, a plurality of processors coupled to said first memory bus, each processor having a

respective cache and said memory bus operated according to a snoopy bus protocol for maintaining coherence between said caches, said snoopy bus

protocol having an address phase, a snoop phase, and a response phase;

a second node having a second memory coupled to a second memory bus, a

plurality of processors coupled to said second memory bus, each processor

having a respective cache and said memory bus operated according to said

snoopy bus protocol; a first internode communication unit coupled to said first memory bus and having a first directory for indicating states of cached blocks of said first

and second memories and a coherence control element coupled to said first

bus and coupled to said first directory for reading state information from

said first directory after said address phase and before said snoop phase

and updating said state information before a next snoop phase; and

a second internode communication unit coupled to said second memory bus

and coupled to said first internode communication unit, said first memory

being accessible to said second node and said second memory being

accessible to said first node, and said second coherence unit having a

second directory for indicating states of cached blocks of said second memory and a coherence control element coupled to said second bus and

coupled to said second directory for reading state information from said

second directory after said address phase and before said snoop phase and

updating said state information before a next snoop phase. 24

15. The data processing system of claim 14 further comprising a plurality of

nodes.

16. The data processing system of claim 14 wherein said first and second

directories are sparse directories.

17. The data processing system of claim 14, wherein said first and second

internode communication units further comprise:

a mesh interface unit; a network interface unit coupled to said mesh interface unit; and a transfer agent coupled to said network interface unit, said memory buses, and

said directories.

18. The data processing system of claim 17, wherein said network interface unit

further comprises: a data transfer unit coupled to an interconnect unit and to said transfer agent

for transferring data and control requests from said interconnect unit to

said transfer agent; and a coherence action unit coupled to said interconnect unit and said directory for

reading state information from said directory.

25 19. A method of maintaining cache coherency within a data processing system

having a plurality of nodes and an interconnect, each node having a plurality of multiprocessors with associated caches and memory, a memory bus controlled by a snoopy bus protocol, a directory for storing state information of cached memory lines,

and a mesh interface controlled by a mesh protocol, comprising the steps of:

receiving a bus request from the said memory bus for the control of said

cached memory line;

reading said state information of said cached line; updating said state information of said cache line;

transforming said bus request into a network request; and

forwarding said network request to said mesh interface.

20. The method of claim 19, further comprising the steps of: receiving a network request from said mesh interface;

arbitrating with said memory bus for scheduling said request onto the said

memory bus;

passing said network request to a transfer agent; and

transferring said network requests from said transfer agent to said memory bus.

26 21. A cache coherence unit within a data processing system having a plurality of nodes, each node having a plurality of processors coupled to a memory bus operating

according to a snoopy bus protocol, each processor having an associated cache, and

coupled to an interconnect, said cache coherence unit comprising:

a control element for arbitrating with said memory bus;

a coherence control unit coupled to said control element for maintaining coherence of cached memory locations of bus requests received from said

memory bus;

a directory coupled to said coherence control unit for storing state information

of said cached memory locations; a first interface unit coupled to said coherence control unit for transforming

said bus requests into mesh-out requests;

a mesh interface coupled to said first interface unit for transferring said mesh-

out requests to said interconnect; a second interface unit coupled to said mesh interface for receiving mesh-in

requests from said interconnect; and

a transfer agent coupled to said second interface unit and coupled to said

memory bus for transferring said mesh-in requests from said second

interface unit to said memory bus.

27 22. The cache coherence unit of claim 21 , wherein said coherence control unit

further comprises: a first agent coupled to said first interface unit for transforming said bus into

mesh-out requests; and

a second agent coupled to said first interface for transferring said mesh-out

requests to said interconnect.

23. The cache coherence unit of claim 21, wherein said coherence control unit

further comprises: a third agent coupled to said first interface for transferring data from said

memory bus to said interconnect.

24. The cache coherence unit of claim 21, wherein said second interface unit

further comprises:

a first input agent for mapping incoming network request from said remote

nodes to bus transactions; and

a second input agent for handling incoming mesh replies and dealing with all

mesh-to-bus data transfers.

25. The cache coherence unit of claim 21, wherein said second interface unit

further comprises: a data transfer unit coupled to said mesh interface and said data storage unit;

and a coherence action unit coupled to said mesh interface unit and said directory.

28 26. A cache coherence unit within a data processing system having a plurality of

nodes and an interconnect, each node having a plurality of multiprocessors with associated caches and memory, a memory bus controlled by a snoopy bus protocol, a

directory for storing state information of cached memory lines, and a mesh interface

controlled by a mesh protocol, comprising:

a control unit coupled to said memory bus for receiving bus requests;

means for reading said state information of said cached line;

means for updating said state information of said cache line; and a mesh interface for forwarding said bus requests to said interconnect.

27. The cache coherence unit of claim 26, further comprising:

means for receiving network requests from said interconnect;

a bus arbitration unit coupled to said memory bus for scheduling said network

requests onto the said memory bus; means for passing said network request to a transfer agent; and

means for transferring said network requests from said transfer agent to said

memory bus.

28. A method for maintaining cache coherence in a multi-node data processing system, comprising the steps of:

receiving a request for a cache line from a local node;

determining the status of said cache line; and

sending a copy of said cache line to said local node.

29 29. The method of claim 28, further comprising the steps of: invalidating the status of remote copies of said cache line; and

notifying said local node that transfer of said cache line is complete.

30. The method of claim 28, further comprising the steps of: requesting a copy of said cache line from a home node; and

receiving a copy of said cache line from said home node.

31. A system for maintaining cache coherence in a multi-node data processing system, comprising: means for receiving a request for a cache line from a local node;

means for determining the status of said cache line; and

means for sending a copy of said cache line to said local node.

32. The system of claim 31 , further comprising:

means for invalidating the status of remote copies of said cache line; and

means for notifying said local node that transfer of said cache line is complete.

33. The system of claim 31 , further comprising:

means for requesting a copy of said cache line from a home node; and

means for receiving a copy of said cache line from said home node.