The present invention relates to a cache memory control system, and more specifically to a cache memory control device and a pipeline control method effective to avoid the congestion of a pipeline in a control device of cache memory in a pipeline processing system.
Cache memory has been conventionally used in a computer system. The cache memory may predict that data used by a central processing device (CPU) may be soon used again, for example, and hold the data in high-speed cache memory so that the data can be quickly accessed. The cache memory may therefore overcome the restrictions of the access speed of main memory.
However, computer systems have had multiple cores, and a pipeline processing system has been used as an access system to cache memory. As a result, the congestion of a pipeline has become serious and required a workaround to the reduction of the congestion.
In the prior art above, in addition to an essential command requiring access to cache memory when the command is executed, a command not requiring access to the cache memory such as a command for a write or a read of data to and from an ASI (address space identify) register has been provided for an external interface unit through a pipeline from the XP0 to the XP6 illustrated in
Patent document 1 as the prior art relating to the control of cache memory described above discloses a technique of accessing at a high speed a part of an area in main memory specified as a noncache area using no cache memory by switching a bus to a bypass route for providing address and data output from an MPU directly to main memory when an address in a bus is within a predetermined address range.
Next, Patent document 2 discloses a technique of improving a use efficiency of cache memory by checking the value of a frequency of accessing a memory position not in the cache from a processor, executing access to the cache if the value exceeds a threshold, and bypassing the cache if the value does not exceed the threshold.
However, the conventional techniques above cannot solve the problem that the congestion in a pipeline largely increases in a multi-core system having an increasing number of cores when all commands including those not accessing cache memory during the execution of commands are to pass through the pipeline for cache memory access.
The present invention aims at reducing the congestion of a pipeline and improving the system performance by externally bypassing noncache type commands not accessing cache memory during the execution in the commands (for example, the commands may be transmitted from each of a plurality of cores configuring a multi-core system) without using the pipeline for controlling the cache memory.
A cache memory control device according to the present invention controls the pipeline processing for a command that accesses cache memory. The cache memory control device includes at least a cache access determination device and a path switch device.
The cache access determination device determines whether a command provided from an external unit (e.g., each of a plurality of cores configuring a multi-core system) is to access cache memory during the execution. The path switch device puts a command determined as accessing cache memory in the pipeline processing, while the path switch device directly outputs to an external unit a command determined as not accessing the cache memory without putting the command into the pipeline processing.
The pipeline control method according to the present invention is to control the pipeline processing for access to cache memory. As with the cache memory control device described above, it is determined whether an externally provided command is to access cache memory during the execution. A command determined as accessing the cache memory is put in pipeline processing, and a command determined as not accessing the cache memory is directly output to an external unit without being put in the pipeline processing.
As described above, according to the present invention, only a command to access cache memory during the execution is put in pipeline processing in the commands provided from each of the cores configuring, for example, a multi-core system, and a command not to access the cache memory is directly output to an external unit without being put in the pipeline processing.
According to an embodiment of the present invention, the congestion of a pipeline for accessing cache memory can be reduced despite of multiple cores in a multi-core system and a large number of commands provided from each core. The embodiment of the present invention thereby successfully contributes to the improvement of system performance.
The cache access determination unit 2 determines whether a command received from an external unit (for example, each of a plurality of cores configuring a multi-core system) is to access cache memory during the execution of the command.
The path switch unit 3 puts a command determined as accessing cache memory in pipeline processing, and outputs a command determined as not accessing the cache memory directly to an external unit without putting the command in a pipeline.
Each of CPUs 101 and 102 includes two core units 15, a cache unit 16 connected to each core unit 15, and an external interface unit 17 located between the cache unit 16 and the system controller 12, respectively.
The priority determination unit 23 in
The XP0 through the XP6 refer to the stages of the pipeline processing for the cache memory access after the PR0 through the PR3. The processing on the pipeline is performed, and the SC-REQ (e.g., from the move-in buffer 26 illustrated in
In an embodiment, a request (command) transmitted from the core unit is stored in the MI-PORT 20 in the PR0. From among those requests, the cache type requests for which the cache memory is accessed during the execution may be distinguished from the noncache type requests for which the cache memory is not accessed during the execution. As a result, the paths of the requests may be switched in a switching process 30. The paths of the requests are switched by the control in a switching control process 31 using the data of the commands stored in the MI-PORT 20. The switching process 30 may correspond to, but not limited to, the switch of an address bus, a data bus, etc. as a hardware configuration (circuit). In this embodiment, the flow of the process is mainly described.
A request (command) for which the cache memory is not accessed, in other words, a request (command) having no cache, includes noncache type requests and ASI type requests. A noncache type request uses an area in which data is not transferred to the cache memory in the main memory, that is, a noncache area. A noncache type request can be any of the four types, that is, a noncache read, a noncache write, a noncache block read, or a noncache block write. An ASI type request can be a data write/read request to an ASI (address space identify) register in which a write/read can be performed from software, and can be any of the two types, that is, an ASI read and an ASI write.
The requests for which cache access is not performed are bypassed to a path other than a pipeline for access to the cache memory in the switching process 30. The requests are directly provided for the external interface unit as NA (noncache ASI)-REQs after the priorities of cores are determined at the PR1. On the other hand, for the cache type requests for which the cache memory is accessed during the execution, the priorities are determined among the cores at the PR1 and among the ports at the PR2, and put in the pipelines XP0 through XP6 depending on the priorities. Thus, the requests put in the pipeline can be considerably reduced in the multi-core system, and the congestion in the pipeline can be successfully reduced.
On the other hand, the AND gate 36 gets the value of the VLD as is and the inverted values of the 5th bit Thus, the PIPE-REQ indicating that the input request corresponding to the CMD is to be put in the pipeline is output from the AND gate 36 to the switch.
The switch selects and switches a path from which a request is to be output depending on the value of the NA-REQ output from the AND gate 35 or the value of the PIPE-REQ output from the AND gate 36.
Finally described is the effect of the present embodiment. The more number of noncache type commands and ASI type commands, the more effect is provided. For example, when a system is activated, half or more of the commands are noncache type commands, and the remaining commands include ASI type commands. On the other hand, when an OS is normally activated, most of the commands are cache type commands, but ASI type commands are also included; however, a noncache type command is issued if, for example, an external disk is accessed.
Therefore, when a large number of noncache type commands are issued during the activation of a system, or when access is concentrated on external disks even during the normal activation of an OS, the effect of reducing the congestion of a pipeline can be enhanced.
This application is a continuation of PCT application of PCT/JP2007/000660, which was filed on Jun. 20, 2007.
Number | Name | Date | Kind |
---|---|---|---|
5586296 | Bernstein et al. | Dec 1996 | A |
5590368 | Heeb et al. | Dec 1996 | A |
5961632 | Shiell et al. | Oct 1999 | A |
6035424 | Freerksen et al. | Mar 2000 | A |
20060143405 | Ishikawa et al. | Jun 2006 | A1 |
Number | Date | Country |
---|---|---|
0 442 116 | Aug 1991 | EP |
1 220 090 | Jul 2002 | EP |
58-99857 | Jun 1983 | JP |
64-082150 | Mar 1989 | JP |
2-32436 | Feb 1990 | JP |
2-189658 | Jul 1990 | JP |
4-324546 | Nov 1992 | JP |
6-290107 | Oct 1994 | JP |
7-56808 | Mar 1995 | JP |
2735781 | Jan 1998 | JP |
2001-22578 | Jan 2001 | JP |
2006-40090 | Feb 2006 | JP |
2006-185284 | Jul 2006 | JP |
Number | Date | Country | |
---|---|---|---|
20100095068 A1 | Apr 2010 | US |
Number | Date | Country | |
---|---|---|---|
Parent | PCT/JP2007/000660 | Jun 2007 | US |
Child | 12636523 | US |