Out of order dram sequencer

FIELD OF THE INVENTION

0001The present invention relates to the architecture and operational method of a memory controller for controlling memory access operations to achieve an increased effective memory bandwidth.

BACKGROUND OF THE INVENTION

0002In most computer or data processing systems, the main active memory, or random access memory (RAM), is a dynamic random access memory (DRAM). The structure of a DRAM is generally composed of a number of memory cells organized into a plurality of banks. Each bank corresponds to an array of the memory cells with each cell being respectively associated with a unique memory address. In particular, memory addresses within a bank are each designated by a row address and a column address, wherein each row address is defined as a memory page. Each page of memory, therefore, contains several memory locations corresponding to the different column designations within the page.

0003When performing a series of access requests, if a currently requested page is found in a same bank currently having another page open, such condition is known as a "page conflict," whereupon the previously opened page must first be closed, or "precharged." After precharging, the requested page may then be opened, or "activated," and then the read or write operation is performed. A "page miss" occurs if the currently requested page is found in a bank which has no page open, thus requiring an activation procedure to be performed. A "page hit" is said to occur when a current memory access request is for a page which is already open from a previous memory access request.

0004Due to the extra processing which must be performed for page conflict and page miss memory accesses relative to page hit requests, the time needed to perform the former two processes is significantly greater than for the latter. In early stages of microprocessor technology development, requests to access a DRAM memory page, for both read and write operations, were received and fulfilled on a first in, first out basis. Such processing tends to be very inefficient, resulting in a large number of page misses and conflicts, and thus requiring an extensive dedication of processor and/or memory controller resources to precharging and activating memory pages.

0005More recently, more advanced processing methods have been developed in which memory access is based on priority. The priority of the access request may be based on various factors such as the type of device sending the request, the type of access requested, the memory address desired to be accessed by the request, etc. The problem with providing memory access strictly on priority, however, is that low priority requests may be denied access for unacceptably long periods of time.

0006Moreover, as each new generation of computers evolves, memory clock speeds, are increased significantly. As the speed of a memory's clock increases, the potential occurrences of and the time penalty for page miss memory operations, bank busy conflicts, and other conflicts also become increasingly significant. In particular, the data bus used to transfer information to and from each accessed memory location is idle during precharging, activating, waiting for bank availability, etc.

0007A solution is therefore needed to mitigate the drawbacks discussed above. In particular, memory processing efficiency would be greatly improved if the order of a sequence of received memory access requests could be rearranged to avoid or reduce conflicts. By avoiding or reducing conflicts, the memory data bus is more efficiently utilized in that idle time in the memory data bus is reduced or eliminated, which thereby effectively increases the memory bandwith of the memory system and enables more memory access transactions to occur in a shorter amount of time than previously possible. <patcit id="pcit0001" dnum="US6385708B"><text>US 6,385,708</text></patcit> discloses a memory controller which delays the issuance of a memory access command if a conflict is detected with a previously issued command. <patcit id="pcit0002" dnum="US5630096A"><text>US 5,630,096</text></patcit> discloses another example of a memory controller, for controlling a synchronous DRAM.

0008The present invention seeks to address the problems identified in the prior art by rearranging the sequentially received order of DRAM access requests to minimize conflicts and delays such as those discussed above, while returning the requested information to requesting units in the same order in which the requests were originally received.

0009In the present invention, memory access requests are successively received in an input queue of a memory controller. A sequence matrix is arranged after the request input queue, whereupon conflicts or potential delays between sequential requests are identified by a conflict detector. The conflict detector re-orders the memory core access requests to optimize the flow of data to and from the data bus. For example, if a bank busy condition or other delay is recognized by the conflict detector in the sequentially received memory requests, the memory controller rearranges the order in which the pending memory requests will be executed to eliminate the conflict or delay, if possible, or otherwise to minimize the delay.

0010Accordingly, in a first aspect of the invention there is provided a memory controller comprising: an input queue for receiving a plurality of memory access requests in an originally received order from at least one requestor, each memory access request including a read or write command and a data control command; a sequence matrix for storing the plurality of memory access requests in a sequenced order; a sequencing unit for arranging the plurality of memory access requests in the sequence matrix into the sequenced order; a conflict detector for detecting whether a conflict or delay would occur among the memory access requests in the sequence matrix; an execution queue for storing a plurality of commands, the plurality of commands being derived from the memory access requests in the sequence matrix; a read return queue for keeping track of the originally received order of the memory access requests after execution of the memory access requests from the execution queue in the execution order, and for returning requested data to the at least one requestor in the originally received order of the memory access requests; and a returned data buffer for holding requested data read from a memory until the data is to be returned to the requestor of the associated memory access request; wherein said conflict detector and said sequencing unit are operable such that, in use, the conflict detector reports detected conflicts or delays to said sequencing unit to cause said sequencing unit to rearrange the order of the memory access requests in the sequence matrix by moving ahead a memory access request involved in a detected conflict; and wherein the memory controller is operable to arrange the plurality of commands in the execution queue such that, in use, a data control command relating to a first memory access request is executed before a read or write command relating to a second memory access request, the second memory access request preceding the first memory access request in the sequenced order.

0011In another aspect, the invention provides a method for controlling memory access operations, comprising: receiving a plurality of memory access requests in an originally received order; detecting memory access conflicts involving one or more of the received memory access requests; rearranging the memory access requests into a sequenced order based on the detection result, said rearranging including moving ahead a received memory access report involved in a detected conflict or delay; extracting a plurality of commands from the plurality of memory access requests, including the step of parsing each memory access request into at least a read or write command and a data control command; rearranging the read or write command and the data control command for each request such that a data control command relating to a first memory access request is executed before a read or write command relating to a second memory access request, the second memory access request preceding the first memory access request in the sequenced order; and executing the plurality of commands, including the step of selecting at least one command at a time from the rearranged sequence of commands.

0012In a further aspect, the invention provides a computer system comprising: a processor; at least one memory device for exchanging data with said processor; and a memory controller as aforesaid, for managing memory access requests from the processor to the at least one memory device.

0013Further features, aspects, and advantages of the present invention will become apparent from the following detailed description of a preferred embodiment of the invention, described with reference to the accompanying drawings, wherein:

0014<figref idref="f0001">FIG. 1</figref> is a block diagram of a memory controller in accordance with the present invention;

0015<figref idref="f0002">FIG. 2</figref> is a flowchart illustrating the process for returning requested data obtained by the re-ordeaed execution of read requests to their respective requestors in the originally received order of the requests.

0016<figref idref="f0003">FIG. 3</figref> is a first embodiment of a conflict re-ordering process performed by the sequence load logic unit in the memory controller upon detection of a conflict or delay in a current timing matrix;

0017<figref idref="f0004">FIG. 4</figref> is a second embodiment of a conflict re-ordering process performed by the sequence load logic unit in the memory controller upon detection of a conflict or delay in a current timing matrix;

0018<figref idref="f0005">FIG. 5</figref> is a third embodiment of a conflict re-ordering process performed by the sequence load logic unit in the memory controller upon detection of a conflict or delay in a current timing matrix;

0019<figref idref="f0006">FIG. 6</figref> is a timing diagram for illustrating the operation of the present invention.

0020<figref idref="f0007">FIG. 7</figref> is a block diagram of a processing system in which the memory controller of the present invention may be utilized.

DETAILED DESCRIPTION OF THE INVENTION

0021For ease of description, the preferred embodiments of the present invention are discussed below as being used in conjunction with dynamic random access memory (DRAM) devices. Nevertheless, it should be understood that the present invention is not limited to applications involving DRAM. Rather, it is emphasized that the memory controller and methods of the present invention may be used in conjunction with other types of random access memories, such as static RAMs (SRAM) and the many different subspecies of DRAMs, including, for example, fast page mode DRAM (FPM DRAM), extended data out DRAM (EDO DRAM), burst EDO DRAM, synchronous DRAM (SDRAM), double data rate DRAM (DDR DRAM), Rambus DRAM (RDRAM), etc.

0022<figref idref="f0001">Fig. 1</figref> shows a memory controller 10 in accordance with the present invention, and which includes an input queue 12, a command parser 14, a sequencing unit 16, a sequence matrix 18, a conflict detector 20, a command sequencer 22, an execution queue 24, a command selector 26, an input/output buffer 28, a read return queue 30, and a returned data buffer 32.

0023Memory access requests enter the memory controller 10 and are received in input queue 12 on a first-in first-out basis. The received requests are then sequentially processed by command parser 14 to obtain relevant information from each request signal, such as memory address (MA) data, a chip select (CS) command (indicating a requested memory bank to be accessed), a row address select (RAS) command, and a column address select (CAS) command, and a write enable (WE) state for indicating whether the request is a read or a write operation.

0024The information obtained is then provided to the sequencing unit 16, which places the received memory access requests into a sequence matrix 18 in accordance with a clock signal received in the sequence matrix 18. It is noted that upon startup of the process, the access requests are loaded into the matrix in the order in which they are received in the sequencing unit 16.

0025Conflict detector 20 monitors the information in the sequence matrix 18 and checks for any conflicts or delays that may occur if the sequence of requests in the matrix were to be executed in the current order in the sequence matrix. Any conflict or delay detected by conflict detector 20 is reported back to sequencing unit 16, which then rearranges the order of the requests in the matrix to minimize or eliminate the time the memory data bus is idle due to the identified conflicts or delays. Conflicts which may be detected by conflict detector 20 include, but are not limited to, page conflicts and bank busy conditions in which a memory bank is busy performing another read or write operation, for example. Delays identified by the conflict detector are conditions which, while not necessarily a conflict with the execution of another memory access request, would require the performance of preparatory steps during which time the memory data bus is idle. Such delays include, for example, page conflicts, page misses, etc.

0026Generally, both read and write type memory access requests may be re-ordered in the sequence matrix 18. However, it is preferable that write requests only be rearranged if necessary to the extent that the re-ordered sequence does not create any addressing conflicts with earlier received requests in to the input queue or otherwise interfere with the data stored or to be stored in the relevant memory locations in connection with any other memory access requests in the matrix.

0027As additional memory access requests are moved into the sequence matrix 18, while also being rearranged to resolve conflicts or reduce delays by sequencing unit 16, the requests at the front of the sequence are moved into the execution queue 24, which serves as a transfer buffer where the rearranged requests await execution. Depending on the conflict resolution process used in the sequencing unit 16, the requests may be moved into the execution queue 24 either on a continuous basis in accordance with a clock signal, in batches of a predetermined number of access requests, or based on a predetermined cumulative size of the requests. Preferably, the clock (CLK) for the execution queue 24 is the same clock (CLK) guiding the loading of sequence matrix 18.

0028Command sequencer 22 arranges the various commands associated with each request transferred to execution queue 24 from the sequence matrix 18, as it may be necessary to insert and/or temporally separate data control commands from the read or write command of a particular memory access request. For example, if a read or write request sent to execution queue 24 requires a precharging and/or activation operation, a data control command to initiate the precharging and/or activation operation is placed in the execution queue ahead of the relevant read or write operation, with at least one other read or write command associated with a different access request positioned between the precharge and/or activate command and the associated read or write command.

0029An advantage of the present invention is realized by separating the read or write commands from such data control commands in this manner. Specifically, in the above example, the precharge and/or activation operation in the above example can be performed while the read or write command for another memory access request can be immediately executed. Thus, the memory data bus does not have to be idle during the time the pre-charge and/or activation operation is perfomed.

0030When each read request is transferred to the execution queue, a tag is temporarily added to the data control commands for that request, for identifying the original relative placement of the each request as received in the input queue 12. Alternatively, each read request may be assigned a respective buffer location in a read buffer 32, which will be described in more detail later.

0031In accordance with a command select clock (CMD CLK) signal fed into command selector 26, one or more memory access requests from the front of the execution queue 24 is (are) selected for execution in the appropriate DRAM bank(s). For example, if the command select clock signal is four times the speed of the clock speed at which requests are loaded into sequence matrix 18, then four access requests are removed from execution queue 24 for each clock signal of queue 24. In this case, command selector 26 regulates the request selection process so that the four access requests which are all selected at one time from queue 24 are executed at even intervals. If the commands are selected from the execution queue for execution at the same clock speed at which new requests are entered into the sequence matrix 18, then command selector 26 may be omitted from memory controller 10.

0032I/O buffer 28 is a transition buffer used during the read or write operation specified in each access request. If a current access request to be executed is a write operation, the data to be written into the selected memory cell is temporarily written into I/O buffer 28. Similarly, data read from a selected memory cell in a read operation is temporarily stored in I/O buffer 28.

0033Upon execution of the requested memory access, each read request is placed into a read return queue 30. Read return queue 30 manages the requested data read from the DRAM upon execution of the read requests, and returns the requested data to the respective requestors in the order in which they were received in the input queue. Data read from the DRAM is either transferred directly to the requestor or is placed into returned data buffer 32, depending on the tag or assigned buffer location associated with the returned request data.

0034Referring now to <figref idref="f0002">Fig. 2</figref>, after executing a read request, the executed read request is returned to the read request queue 30, with the data obtained by the request being temporarily held in I/O buffer 28 (step 100). If the read request queue 30 determines at step 110 that the returning read request is associated with the most current tag or assigned buffer location, the data obtained by that request is returned to the requestor at step 120, and the current tag/buffer location information is updated at step 130.

0035A tag or buffer location is "current" if it is assigned to or associated with the read request having the longest latency in the memory controller 10. Read return queue 30 may keep track of the most current tag or buffer location, for example, by incrementing a count value which represents the current tag or buffer location each time a returned data is transferred to its requestor.

0036If the returning read request has a tag or an assigned buffer location which is not current, the data is placed into the read data buffer 32 (step 140) until the associated tag or buffer location becomes current. After placing the returned data into the buffer 32 in step 140 or after updating the current tag or buffer location in step 130, the read return queue 30 determines whether or not returned data corresponding to the current tag or buffer location can be found in the returned data buffer 32. If "yes," the process returns to step 120, where the current returned data is transferred to the requestor of that data, and the current tag or buffer location is again updated in step 130. If returned data corresponding to the current tag or buffer location is not found in the buffer 32, the process returns to step 100 to receive the returned data obtained by the execution of the next read request in the execution queue. Due to this process shown in <figref idref="f0002">Fig. 2</figref>, the read data is returned to the respective requestors in the order in which the requests were received into the input queue 12.

0037If tags are used to indicate the originally received order of the read requests, the tags are temporarily inserted among the data control commands of each request, but are not included in the read data returned to the requestors. Preferably, the functions of the read return queue 30 and of the returned data buffer 32 are performed according to a clock speed corresponding to the clock speed of the DRAM or an integral multiple thereof. As such, returned read requests and returned data received in the read return queue 30 and returned data buffer 32 can be coordinated with the updating of the current tag/buffer location and the transferring of the returned data, respectively.

0038A first embodiment of a conflict re-ordering process which the sequencing unit 16 may use to reschedule a memory access request due to a detected conflict or delay will be described with reference to the flowchart shown in <figref idref="f0003">Fig. 3</figref> At step 200, the last memory access request to be parsed is placed in the sequence matrix 18. At step 210, the status of the memory bank desired to be accessed is checked to determine if any conflicts or delays would occur if the newly arrived memory access request is executed at its present position in the sequence matrix 18.

0039If sequencing unit 16 detects a conflict or delay with respect to the newly arrived request, sequencing unit 16 determines at step 240 whether a more suitable timing position can be found among the sequence of memory access requests ahead of its current position. Specifically, sequencing unit 16 first determines whether any unresolved conflicts or delays are present in the sequence ahead of the new access request, and if so, whether or not the new access request can be performed during that time without conflict. If there are no pending conflicts or delays, sequencing unit 16 checks whether the new access request may be rescheduled at any point in the matrix without causing any new conflicts or delays among the previously scheduled requests. If not, the process is redirected to step 230, whereupon the new access request is left at the end of the current timing sequence with the unresolved condition. If a suitable timing position can be found ahead in the sequence, the newly arrived request is inserted into sequence at that position (step 250). The process is then returned to step 200 to be repeated for the next incoming memory access request.

0040If, on the other hand, no conflict or delay is detected with respect to the newly arrived access request, the conflict detector 20 next determines at step 220 whether any unresolved conflicts or delays are present in the timing matrix ahead of the current position of the newly arrived request. If an unresolved conflict or delay is found, the process is redirected to step 240 discussed above. If no existing conflicts are found, the sequencing unit 16 leaves the request in its current position in sequence matrix 18 at step 230, and then returns to step 200 to repeat the process for the next incoming memory access request.

0041In this embodiment, any unresolved conflicts or delays may or may not be later resolved with the arrival of a new access request with a subsequent iteration of the sequencing unit 16. If no suitable request arrives to alleviate the conflict or delay, the memory access requests will continue to be processed in the designated order, but there will be some inefficiency in utilization of the memory data bus line due to the unresolved conflict(s) or unmitigated delay(s).

0042A second embodiment of the conflict re-ordering process performed by sequencing unit 16 is illustrated in the flowchart shown in <figref idref="f0004">Fig. 4</figref>. As another memory access request is moved out of the sequence matrix 18 to the execution queue 24, the access request next in line in the sequence matrix is moved into the first location of the sequence matrix at step 300. At step 310, conflict detector 20 determines whether a conflict or delay is present with respect to the access request at the head of the sequence matrix (i.e., the earliest one in among the requests in the matrix). If no conflict is found, the access request is passed onto the execution queue 24 at step 320, and the process returns to step 300.

0043On the other hand, if a conflict or delay is detected in step 310, the conflict detector 20 turns its attention to the next access request in the sequence matrix, and determines at step 330 whether or not that request can be performed at that time position without any conflicts. If no conflicts or delays would be created by scheduling that access request at that time position, the request is sent to the execution queue 24 at step 320, and the process returns to step 300. If a conflict or delay is found, the process repeats step 330 until a request is found which may be suitably executed at that time position.

0044In this embodiment, if a conflict or delay is found to exist with respect to the memory access of the request at the head of the sequence matrix, and any subsequent requests, the rejected requests remain in their current position in the sequence until a suitable time slot is found for the request. With this process, no memory access will be sent to the execution queue 24 with a conflict or delay condition. Also, each request is given priority based on latency, and will be executed at the earliest possible time slot in which no conflict condition is created by the timing of that request.

0045A variation of the embodiment discussed above with reference to <figref idref="f0004">Fig. 4</figref> is shown in <figref idref="f0005">Fig. 5</figref>, wherein if a conflict or delay is found with respect to the current timing position of the access request at the head of the sequence matrix, the offending request is sent to the back of the sequence, rather than being left in the sequence at its current location. Specifically, if a conflict is found at step 310, step 420 is executed in which the request having the conflict or delay is sent to the back of the sequence. Then, conflict detector 20 moves to the next access request in line to determine if any conflicts or delays are found with respect to that request (step 430). If no conflicts or delays are detected, the request is sent to the execution queue 24, similarly to the process shown in <figref idref="f0004">Fig. 4</figref>. If a conflict or delay is detected, the process returns to step 420.

0046Referring now to <figref idref="f0006">Fig. 6</figref>, a timing chart is shown illustrating the result obtained upon the operation of the present invention. In this example, read requests RD₀, RD₁, and RD₂ are received in the input queue in the order as listed. Assume that a page conflict condition has been detected with respect to RD₀, and that RD₁ and RD₂ are both page hits in other memory banks. Because a delay would have resulted if RD₀ were allowed to execute to completion before RD₁ and RD₂, the requests have been rearranged so that while the precharge operation is being performed for RD₀, read memory accesses for RD₁ and RD₂ are executed. Rearranging the commands in this manner minimizes the time that the memory data bus would have been idle while waiting for the precharge operation for RD₀ to finish executing, thus resulting in a more efficient utilization of the memory data bus.

0047After the data for RD₁ and RD₂ halve been read from the DRAM, the activation operation for RD₀ is performed, and then the read access command for RD₀ is performed. The read data "1111" and "2222" are transferred to the returned data buffer in the order the read operations are executed. After the "0000" data is obtained from the appropriate memory cell, however, the data is returned first to the requestor of RD₀. Then, the data "1111" and "2222" are returned to their respective requestors, in that order.

0048<figref idref="f0007">Fig. 7</figref> illustrates an exemplary processing system 900 which may utilize the memory controller 10 of the present invention. The processing system 900 includes one or more processors 901 coupled to a local bus 904. Memory controller 10 and a primary bus bridge 903 are also coupled the local bus 904. The processing system 900 may include multiple memory controllers 10 and/or multiple primary bus bridges 903. The memory controller 10 and the primary bus bridge 903 may be integrated as a single device 906.

0049The memory controller 10 is also coupled to one or more memory data buses 907. Each memory bus accepts memory components 908 which include at least one memory device 902. The memory components 908 may be formed as a memory card or a memory module. Examples of memory modules usable in the system 900 include single inline memory modules (SIMMs) and dual inline memory modules (DIMMs). The memory components 908 may include one or more additional devices 909. For example, in a SIMM or DIMM, the additional device 909 might be a configuration memory, such as a serial presence detect (SPD) memory.

0050The memory controller 10 may also be coupled to a cache memory 905. The cache memory 905 may be the only cache memory in the processing system. Alternatively, other devices, for example, processors 901 may also include cache memories, which may form a cache hierarchy with cache memory 905. If the processing system 900 include peripherals or controllers which are bus masters or which support direct memory access (DMA), the memory controller 10 may implement a cache coherency protocol. If the memory controller 10 is coupled to a plurality of memory buses 907, each memory bus 907 may be operated in parallel, or different address ranges may be mapped to different memory buses 907.

0051The primary bus bridge 903 is coupled to at least one peripheral bus 910. Various devices, such as peripherals or additional bus bridges may be coupled to the peripheral bus 910. These devices may include a storage controller 911, an miscellaneous I/O device 914, a secondary bus bridge 915, a multimedia processor 918, and an legacy device interface 920. The primary bus bridge 903 may also coupled to one or more special purpose high speed ports 922. In a personal computer, for example, the special purpose port might be the Accelerated Graphics Port (AGP), used to couple a high performance video card to the processing system 900.

0052The storage controller 911 couples one or more storage devices 913, via a storage bus 912, to the peripheral bus 910. For example, the storage controller 911 may be a SCSI controller and storage devices 913 may be SCSI discs. The I/O device 914 may be any sort of peripheral. For example, the I/O device 914 may be an local area network interface, such as an Ethernet card. The secondary bus bridge may be used to interface additional devices via another bus to the processing system. For example, the secondary bus bridge may be an universal serial port (USB) controller used to couple USB devices 917 via to the processing system 900. The multimedia processor 918 may be a sound card, a video capture card, or any other type of media interface, which may also be coupled to one additional devices such as speakers 919. The legacy device interface 920 is used to couple legacy devices, for example, older styled keyboards and mice, to the processing system 900.

0053The processing system 900 illustrated in <figref idref="f0007">Fig. 7</figref> is only an exemplary processing system with which the invention may be used. While <figref idref="f0007">Fig. 7</figref> illustrates a processing architecture especially suitable for a general purpose computer, such as a personal computer or a workstation, it should be recognized that well known modifications can be made to configure the processing system 900 to become more suitable for use in a variety of applications. For example, many electronic devices which require processing may be implemented using a simpler architecture which relies on a CPU 901 coupled to memory components 908 and/or memory devices 902. These electronic devices may include, but are not limited to audio/video processors and recorders, gaming consoles, digital television sets, wired or wireless telephones, navigation devices (including system based on the global positioning system (GPS) and/or inertial navigation), and digital cameras and/or recorders. The modifications may include, for example, elimination of unnecessary components, addition of specialized devices or circuits, and/or integration of a plurality of devices.

0054Although the present invention has been described in relation to particular embodiments thereof, many other variations and modifications and other uses will become apparent to those skilled in the art. It is preferred, therefore, that the present invention be limited not by the specific disclosure herein, but only by the appended claims.