Electrical computers and digital data processing systems: input/ – Intrasystem connection – Bus interface architecture
Reexamination Certificate
2000-05-12
2002-09-24
Wong, Peter (Department: 2181)
Electrical computers and digital data processing systems: input/
Intrasystem connection
Bus interface architecture
C711S110000, C711S131000, C711S149000
Reexamination Certificate
active
06457087
ABSTRACT:
BRIEF DESCRIPTION OF THE INVENTION
The present invention relates generally to multiprocessor systems and methods. More particularly, the invention relates to a shared memory cache coherent multiprocessor system utilizing a point-to-point interconnect architecture.
BACKGROUND OF THE INVENTION
The aim of parallel processing is to utilize a number of processing elements that can communicate and cooperate to solve a problem. In a highly parallel processing system, hundreds of processing elements are used to solve a problem that is spread over many processing elements. Not all of the processing elements are used to run a single problem and the system can be configured to execute multiple problems simultaneously. By contrast, in a low parallel processing system, tens of processing elements are used to solve an entire problem.
Symmetric multiprocessing (SMP) is one such type of low parallel processing system. A SMP system is characterized by “symmetric” processors that each have an equal share and access to the system resources, including memory and I/O. The processors are managed by a single operating system that provides an application program with a single view of the entire system.
FIG. 1
illustrates one such shared memory SMP
100
. There is shown a number of symmetric processors
102
A-
102
N interconnected by a bus
104
. A main memory
106
is provided that is connected to the bus
104
and shared by each of the processors
102
. In addition, I/O devices
108
are connected to the bus
104
and are accessible by each processor
102
and the main memory
106
. Each of the components of the system
100
are synchronized to a common system clock
110
.
In order to reduce the traffic to the main memory
106
, each processor
102
has a local cache memory
112
that can contain shared data. Since the data in the each processor's cache
112
can be shared by each processor
102
, the problem then becomes one of cache coherency. In most SMW systems, a snoopy bus protocol is used to maintain cache coherency. In a snoopy bus protocol, a memory access transaction, such as a read or write, is broadcasted to all the processors
102
connected to the bus
104
. Each processor
102
monitors or “snoops” the bus
104
for a memory access transaction that pertains to a cache line that is associated with the processor's cache
112
. When the processor
102
finds such a transaction, it takes appropriate action to ensure that each cache line is coherent within the system
100
.
There are several disadvantages with this type of SMW system. The primary disadvantage is the use of the bus as the interconnect structure. Although the use of the bus provides cache coherency, it is a limiting factor for improving the system's throughput. First, the use of the bus constrains the number of transactions that can be processed simultaneously. The same bus is used to process both memory and I/O transactions initiated by each processor. As such, only one transaction can be processed at a time.
Second, the contention for the bus by each processor to access main memory unnecessarily increases the overhead in servicing a memory access transaction. Various approaches have been tried to overcome this limitation such as increasing the width of the bus, running the bus at a higher clock speed, and increasing the size of the caches. However, each of these approaches greatly increases the expense and complexity of the system.
Another limitation with the use of the bus are the well-known transmission line effects associated with buses. These transmission fine effects are attributable to the complicated electrical phenomenon present in the connections made to each device coupled to the bus. These transmission line effects limit the speed at which the bus operates thereby reducing the system's throughput.
Accordingly, there exists a need for a SMP system that overcomes these shortcomings.
SUMMARY OF THE INVENTION
The present invention pertains to a system and method for operating a shared-memory multiprocessing system with cache coherency. The system includes a number of devices including several processors, a multiple accessible main memory, and several external I/O devices. Each device is connected to a flow control unit FCU). The FCU controls the exchange of data between each device in the system. The FCU includes a snoop path for processing a first set of data transactions and one or more data paths that process a second set of data transactions. The snoop path and each of the data paths can operate concurrently thereby providing the system with the capability of processing multiple data transactions simultaneously thereby increasing the system's throughput.
Each device is connected to the FCU by means of a dedicated channel or point-to-point connection. The FCU has a dedicated interface unit for each device. A channel is used by one device and its associated interface unit in the FCU. Since the channel is not a bus, it does not experience the well known transmission line effects associated with buses, and as such, can operate at a high transfer rate. The improved speed of the channel increases the system's throughput.
In addition, the use of the channel does not require an arbitration phase or arbitration logic as is required in bus-based interconnect structures. The elimination of the arbitration logic reduces the complexity of the circuitry and the elimination of the arbitration phase increases the system throughput.
Each of the processors is associated with its own system clock and runs independent of other processor and the FCU. As such, the FCU can receive requests from each of the processors with a high degree of tolerance to clock skewing between the different devices.
In a preferred embodiment, the technology of the present invention can be utilized in a SMP environment. There can be n symmetric processors, n CPU interface units (CIU), l memory control units (MCU), and k bus bridge units (BBU) connected to the FCU. Each processor can have a L2 cache containing data that is shared amongst the processors. Each CIU is coupled to a processor bus and receives memory and I/O requests initiated by the processor to access data that is external to the processor. The CIU translates the processor bus cycles into channel cycles and vice versa.
Each MCU is connected to one or more memory devices and serves to control access between the FCU and the portion of main memory that is under its control. Each BBU serves to provide a communication path between one or more I/O buses interconnected to external I/O devices and the FCU. The BBU receives data requests from the FCU via the channel and from the I/O buses. The BBU converts the I/O bus cycles into channel cycles and vice versa.
The FCU processes memory and I/O transactions received from the devices. The memory transactions can be used to access data that resides in another processor's cache, to access data stored in main memory, to maintain cache coherency, and to access memory mapped I/O. Main memory can include portions designated as either cacheable memory, non-cacheable memory, and I/O addressable memory. The I/O transactions can be used to transfer data between the processor associated with a CIU and an external I/O device.
REFERENCES:
patent: 4315308 (1982-02-01), Jackson
patent: 4438494 (1984-03-01), Budde et al.
patent: 4480307 (1984-10-01), Budde et al.
patent: 5161156 (1992-11-01), Baum et al.
patent: 5271000 (1993-12-01), Engbersen et al.
patent: 5313609 (1994-05-01), Baylor et al.
patent: 5335335 (1994-08-01), Jackson et al.
patent: 5440698 (1995-08-01), Sindhu et al.
patent: 5505686 (1996-04-01), Willis et al.
patent: 5511226 (1996-04-01), Zilka
patent: 5513335 (1996-04-01), McClure
patent: 5524234 (1996-06-01), Martinez, Jr. et al.
patent: 5526380 (1996-06-01), Izzard
patent: 5535363 (1996-07-01), Prince
patent: 5537569 (1996-07-01), Masubuchi
patent: 5537575 (1996-07-01), Foley
patent: 5553310 (1996-09-01), Taylor et al.
patent: 5561779 (1996-10-01), Jackson
patent: 5568620 (1996-10-01), Sarangdhar et al.
patent: 5574
Conexant Systems Inc.
Hale Kelly H.
Kind Keith
Vo Tim
Wong Peter
LandOfFree
Apparatus and method for a cache coherent shared memory... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Apparatus and method for a cache coherent shared memory..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Apparatus and method for a cache coherent shared memory... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-2821575