The present technique relates to an apparatus, memory controller, memory module and method for controlling data transfer in memory.
Within modern data processing systems, it is known to provide a memory controller for controlling access to a portion of memory. Hence, access requests from one or more sources may be routed to the memory controller, which then sends appropriate commands to the portion of memory to cause read and write operations to be performed in order to process the access requests. For a read operation, the read data is returned from the memory to the memory controller, and for a write operation the write data is output from the memory controller to the memory.
Often the portion of memory controlled by the memory controller will be formed of a plurality of memory modules, and the memory controller can perform read or write operations to each of the memory modules.
Memory access operations can incur significant power consumption and introduce significant latency, and accordingly it is desirable to seek to provide improved techniques for performing data transfer in memory.
In one example configuration, there is provided an apparatus comprising: a memory controller and a plurality of memory modules; wherein: the memory controller, in order to control direct data transfer, is configured to: transmit a first direct transfer command to a first memory module of the plurality of memory modules, wherein the first direct transfer command comprises information indicating that the first memory module should transmit data bypassing the memory controller; and transmit a second direct transfer command to a second memory module of the plurality of memory modules, wherein the second direct transfer command comprises information indicating that the second memory module should store the data received directly from the first memory module; the first memory module is configured to: receive the first direct transfer command from the memory controller; and directly transmit the data for receipt by the second memory module in dependence on the first direct transfer command; and the second memory module is configured to: receive the second direct transfer command from the memory controller; receive the data from the first memory module directly; and store the data in dependence on the second direct transfer command.
In another example configuration, there is provided a memory controller configured, in order to control direct data transfer, to: transmit a first direct transfer command to a first memory module, wherein the first direct transfer command comprises information indicating that the first memory module should transmit data bypassing the memory controller; and transmit a second direct transfer command to a second memory module, wherein the second direct transfer command comprises information indicating that the second memory module should store data received directly from the first memory module.
In a further example configuration, there is provided a memory module configured, in order to implement direct data transfer, to: receive a direct transfer command from a memory controller, wherein the direct transfer command comprises information indicating that the memory module should transmit data bypassing the memory controller; and directly transmit data for receipt by a further memory module in dependence on the direct transfer command, such that the transmitted data bypasses the memory controller.
In a yet further example configuration, there is provided a memory module configured, in order to implement direct data transfer, to: receive a direct transfer command from a memory controller, wherein the direct transfer command comprises information indicating that the memory module should store data received directly from a further memory module; receive data from the further memory module directly that has bypassed the memory controller; and store the data in dependence on the direct transfer command.
In a still further example configuration, there is provided a method comprising: transmitting a first direct transfer command from a memory controller to a first memory module, wherein the first direct transfer command comprises information indicating that the first memory module should transmit data bypassing the memory controller; transmitting a second direct transfer command from the memory controller to a second memory module, wherein the second direct transfer command comprises information indicating that the second memory module should store data received directly from the first memory module; transmitting data from the first memory module to the second memory module directly in dependence on the first direct transfer command; and storing the data in the second memory module in dependence on the second direct transfer command.
The present technique will be described further, by way of example only, with reference to embodiments thereof as illustrated in the accompanying drawings, in which:
Before discussing the embodiments with reference to the accompanying figures, the following description of embodiments is provided.
In accordance with one embodiment, an apparatus is provided that comprises a memory controller and a plurality of memory modules. Direct data transfer between the memory modules is supported. In particular, the memory controller is arranged to control such a direct data transfer by transmitting a first direct transfer command to a first memory module and a second direct transfer command to a second memory module. Whilst in one embodiment the first direct transfer command may be a separate command to the second direct transfer command, with the first memory module receiving the first direct transfer command and the second memory module receiving the second direct transfer command, in an alternative embodiment the first and second direct transfer commands may in fact be the same command which is received by both the first memory module and the second memory module, that single command providing the information required by both the first memory module and the second memory module.
The first direct transfer command comprises information indicating that the first memory module should transmit data bypassing the memory controller, and the second direct transfer command comprises information indicating that the second memory module should store the data received directly from the first memory module.
Following receipt of the first direct transfer command from the memory controller, the first memory module is configured to directly transmit the data for receipt by the second memory module in dependence on that first direct transfer command. The second memory module is then arranged, following receipt of the second direct transfer command, to receive the data from the first memory module directly and to store that data in dependence on the second direct transfer command.
In typical systems it is traditionally the case that all of the memory modules are essentially identical, and together form a uniform address space, potentially interleaved across the various memory modules. Within such an environment, the memory controller is arranged to process read requests by reading data from the relevant memory module and providing that data to a request source, and is arranged to process write requests by outputting write data from a request source to the relevant memory module for storage therein. Which memory module is accessed will depend on the address specified by the request source.
However, the inventors realised that with emerging technologies there is likely to be a need to transfer data from one memory module to another. Using existing techniques, this would involve the memory controller issuing a read command to read data from one memory module, and then, once that data is received by the memory controller, the memory controller would subsequently output a write command along with that data in order to cause that data to be written to a different memory module.
The technique of the described embodiments aims to alleviate the latency and power consumption associated with such data transfers, by providing a mechanism which avoids the data itself needing to be routed through the memory controller, whilst allowing the memory controller to still orchestrate the data transfer.
In particular, in accordance with the present technique, the memory controller still communicates directly with the first memory module and the second memory module in order to cause the data transfer to take place, but the data itself is not routed via the memory controller, and instead the operations of the first and second memory modules are coordinated so that the data output by the first memory module can be received directly by the second memory module for storage therein.
The first and second direct transfer commands can take a variety of forms. Whilst in one embodiment the first direct transfer command may explicitly identify to the first memory module that the second memory module is to receive the output data, and the second direct transfer command may explicitly identify to the second memory module that it is to receive data from the first memory module, in other embodiments there is no requirement for the first memory module to know the recipient for the data that it is to output in response to the first direct transfer command, or for the second memory module to explicitly know the source memory module from which it is receiving data when it processes the second direct transfer command.
In particular, in some embodiments the first and second direct transfer commands may cause the timing of the operations performed by the first memory module and the second memory module to be coordinated so as to ensure that when the first memory module outputs the data onto a communications infrastructure accessible to the second memory module as a result of processing the first direct transfer command, the second memory module is at that point ready to receive that data from the communications infrastructure via its processing of the second direct transfer command. One or both of the first memory module and the second memory module may also be arranged to be responsive to the direct transfer commands to perform other ancillary operations required by the data transfer that would naturally be performed by the memory controller were the data to be routed through the memory controller in the standard manner, but which cannot readily be performed by the memory controller when the data is being transferred directly between the memory modules. For example, as will be discussed in more detail below, at least one of the memory modules may perform one or more operations to ensure compliance of the direct data transfer with communication protocol rules of the communications infrastructure.
It should be noted that use of the term “directly” herein, for example when describing the first memory module as directly transmitting the data for receipt by the second memory module, and describing the second memory module as receiving the data from the first memory module directly, is intended to refer to the functionality achieved via performance of the first and second direct transfer commands, whereby the data passes from the first memory module to the second memory module without being routed via the memory controller, which can reduce latency and power consumption, and result in higher efficiency in the use of the communications infrastructure. However, there is no restriction on how the first and second memory modules are physically interconnected in order to achieve the described direct transfer functionality. Hence, whilst in one embodiment a direct physical link may be provided between the first and second memory modules, this is not a requirement, and any suitable path may be used by which the data is directly routed from the first memory module to the second memory module without that data being routed through the memory controller. Hence, for example, one or more intervening components (such as additional memory modules) may be located in the physical path between the first memory module and the second memory module whilst still allowing the first memory module to output data that can be directly received by the second memory module, i.e. without that data first having to be returned to the memory controller for subsequent onward propagation to the second memory module.
The memory controller and the plurality of memory modules can be arranged to communicate in a variety of ways. In one embodiment the apparatus further comprises a communications infrastructure interconnecting the memory controller and the plurality of memory modules, with the memory controller having a shared interface to communicate with said plurality of memory modules. The memory controller may be arranged to control the direct data transfer between the first memory module and the second memory module by issuing the first and second direct transfer commands via the shared interface, and the first memory module and the second memory module may be arranged to operate in response to said first and second direct transfer commands so as to cause the data to be transferred over the communications infrastructure from the first memory module to the second memory module bypassing the memory controller.
By using the direct data transfer approach of the described embodiments, this can significantly reduce bandwidth and power consumption that would be incurred were the traditional approach to be adopted whereby the memory controller would read the data from the first memory module back through its shared interface, and then output that data via a separate write command from the shared interface to the second memory module.
There are a number of ways in which the first and second memory modules can be arranged to respond to the first and second direct transfer commands, respectively. In one embodiment, issuance by the memory controller of said first and second direct transfer commands is arranged to cause the first memory module to perform a read operation to output the data onto the communications infrastructure, and to cause the second memory module to perform a write operation to store the data output onto the communications infrastructure by the first memory module, a relative timing of the read operation and the write operation being controlled to ensure that at the time the write operation is performed by the second memory module, the first memory module is outputting the data onto the communications infrastructure. By controlling the relative timing of the operations performed by the first memory module and the second memory module, this can allow the second memory module, when performing its write operation in response to the second direct transfer command, to be directly provided at that point in time with the data read by the first memory module. In some embodiments, there is hence no requirement for the second memory module to be aware of where the data it is writing originated from.
There are a number of ways in which the relative timing may be specified. In one embodiment, the first direct transfer command comprises a read command issued to the first memory module and the second direct transfer command comprises a write command issued to the second memory module, use of the read command and the write command causing said relative timing to be selected for the read operation and the write operation. The relative timing can be encoded in a variety of ways. For example, in one embodiment the read and write commands could be modified read and write commands, and the relative timing may be encoded directly into fields of the modified read command and/or modified write command. However, in an alternative embodiment, such relative timing values may be pre-programmed into a storage element such as a mode register, which will then be referenced by the first and second memory modules when performing the read and write operations specified by the read and write commands.
Whilst the direct data transfer technique of the described embodiments allows the memory controller to still orchestrate the data transfer, whilst avoiding the need for the data itself to be routed through the memory controller, thereby giving rise to bandwidth and performance benefits, in some implementations certain related activities that a memory controller might normally perform in association with standard read and/or write operations can become difficult for the memory controller to perform when the data itself is not passed through the memory controller. In one embodiment, certain of these tasks can be devolved to either or both of the first and second memory modules, with the first and/or second direct transfer commands causing those tasks to be performed by the memory module(s). For example, in one embodiment, at least one of the first memory module and the second memory module comprises direct transfer handling circuitry to perform at least one operation to ensure compliance of the direct data transfer with communication protocol rules of the communications infrastructure.
In one embodiment, the direct transfer handling circuitry is provided in at least the first memory module and is arranged, when the first memory module responds to the first direct transfer command by performing a read operation, to perform a check code generation operation to generate a check code for that data which is then output to the second memory module in association with the read data. In situations where the memory controller itself issues a write command along with associated write data to be written into a memory module, the memory controller can generate a check code (for example a cyclic redundancy check (CRC) code) in association with the write data, in order to enable the recipient memory module to check that the data it receives is correct. However, when adopting the direct data transfer technique of the described embodiments, such functionality would become difficult/impractical to implement by the memory controller, and instead in accordance with the described embodiment the first memory module is configured to generate a check code when performing a read operation in response to the first direct transfer command. This then enables the second memory module to perform the usual checks that it would perform when processing a write operation.
In one embodiment, the first and second memory modules comprise a physical layer interface for coupling to the communications infrastructure. Similarly, the memory controller will also have a physical layer interface for coupling to the communications infrastructure. The physical layer interface is provided for managing the communication over the communications infrastructure to take into account the nature of the physical connections between the memory controller and the memory modules. That physical connection will typically comprise multiple relatively long physical wire connections, and may for example be required to operate as a parallel interface at high speed. The physical layer interface operates in a manner to maintain synchronised communication across that physical connection between the memory controller and the memory modules.
Due to the nature of the physical connection, time varying effects can be introduced that can affect the ability to correctly synchronise the various signals, and hence it is typically the case that such a physical layer interface needs to be subjected to a periodic training operation during which test commands are issued over the physical connections in order to enable performance of any necessary recalibration of the phase delay circuits and the like used to ensure synchronised communication between the memory controller and the memory modules.
Typically, such a training operation is performed by the memory controller, since all of the relevant paths to be tested either emanate from, or terminate at, the memory controller. However, in accordance with the direct data transfer technique of the described embodiments, there are additional paths via which data is routed in a manner that bypasses the memory controller. In accordance with one embodiment, the direct transfer handling circuitry within one or both of the first and second memory modules can be arranged to also perform training operations in order to ensure that synchronised communication is maintained across the various paths used during direct data transfer. In particular, in one embodiment the direct transfer handling circuitry comprises training circuitry, responsive to a training event, to perform a training operation of the corresponding physical layer interface in order to determine control values used to control timing of signals issued during performance of a direct data transfer between the first memory module and the second memory module.
Furthermore, in one embodiment the direct transfer handling circuitry further comprises synchronisation control circuitry to store the control values determined during the training operation, and to use the stored control values to control the timing of signals issued during a subsequent performance of the direct data transfer. Hence the control values required are stored locally within the memory modules so as to facilitate correct synchronised communication when directly transferring data between the memory modules.
It is known to employ termination circuits within the memory controller and the memory modules in order to improve signal integrity during data transfer over the communications infrastructure. In accordance with such termination techniques (often referred to as on-die termination (ODT)), termination resistor circuitry within the memory controller and within the memory modules is used for impedance matching in the transmission lines. However, when supporting the direct data transfer techniques of the described embodiments, the issue of termination becomes more complex, due to the different paths by which data may be communicated. In one embodiment, this problem is alleviated by the direct transfer handling circuitry being arranged to comprise termination control circuitry to perform a termination setting operation to configure termination components within the associated memory module dependent on whether a currently processed command identifies the direct data transfer or identifies a transfer to be performed between the memory controller and the associated memory module. Hence, the termination settings applied within the memory modules can be varied dependent on whether normal transfers via the memory controller are taking place, or instead direct data transfers are taking place.
Furthermore, in one embodiment, when the currently processed command identifies the direct data transfer, the termination setting operation is further arranged to configure the termination components within the associated memory module dependent on whether the associated memory module is the first memory module or the second memory module. Hence, the termination settings applied by the source memory module and destination memory module may be different. The termination settings can also be varied amongst the other components connected to the communications infrastructure, whether they be additional memory modules not party to the current direct data transfer, or the memory controller itself which is being bypassed during the direct data transfer.
There are a number of ways in which the memory controller can be arranged to generate the necessary first and second direct transfer commands required to cause the direct data transfer to be performed. In particular, the memory controller may receive access requests from a variety of sources, and in one embodiment can be arranged to detect situations where such an access request requires data to be transferred between the first and second memory modules, and in that instance to cause the required direct transfer commands to be generated. In particular, in one embodiment the memory controller comprises direct transfer control circuitry that is responsive to a transfer access request received by the memory controller that requires data to be transferred between the first and second memory modules, to control generation of the first and second direct transfer commands required to cause the direct data transfer to be performed.
In one embodiment, the memory controller comprises a pending access requests buffer to store pending access requests to be processed, and the direct transfer control circuitry is responsive to the presence of a transfer access request in the pending access requests buffer that requires data to be transferred between the first and second memory modules, to update the pending access requests buffer to identify the first and second direct transfer commands required to cause said direct data transfer to be performed. Hence, in such an embodiment the direct transfer control circuitry can monitor the contents of the pending access requests buffer in order to detect situations where the source and destination addresses for a transfer access request reside within the first and second memory modules, respectively, and in that instance can update the contents of the pending access requests buffer so as to cause the earlier-mentioned first and second direct transfer commands to subsequently be issued when that transfer access request is processed.
As mentioned earlier, the inventors realised that with emerging technologies there is likely to be a need to transfer data from one memory module to another. For example, it is possible that the first memory module will have a different memory characteristic to the second memory module. The memory characteristic that is different can take a variety of forms, for example the relative performance of the two memory modules may be different, the density of the memory storage elements may be different, the physical construction itself may be different, the persistency of data within the memory storage elements of the different memory modules may be different, etc. This is likely to become more common with the advent of additional types of memory that can be used in addition to the more standard memory such as dynamic random access memory (DRAM).
For example, with the advent of Storage Class Memory (SCM), new types of memory modules are emerging that contain a different type of memory, such as Phase Change Memory (PCM), 3D XPoint, etc. These SCM memory modules may use a different protocol, for example NVDIMM-P, or the proprietary DDR4-T, to talk to the System-on-Chip (SoC). But those memory modules will still use the same physical interface and are ultimately able to occupy the same memory channel as DRAM memory modules. It would hence for example be possible to construct a memory channel where a memory controller communicates with one or more DRAM memory modules and more or more SCM memory modules. Accordingly, these different types of memories may coexist on the same memory channel, and thus on the same data bus, but on different memory modules within that memory channel.
With these emerging technologies, the inventors realised that this provides a number of options as to how the memories are organised and managed. For example, it would be possible to arrange for one of the first and second memory modules to be used as a cache for data stored in the other one of the first and second memory modules. For instance, if we consider the particular example above of a DRAM memory module and an SCM memory module, it may be desirable to treat the DRAM memory module as a cache for the typically larger, cheaper and slower SCM memory module. When adopting such an approach, data will inevitably have to move between DRAM and SCM, for cache linefills, and writebacks. In such an approach, the use of the direct data transfer technique of the described embodiments can give rise to significant benefits in terms of reduced latency, reduced power consumption and improved efficiency.
However, the direct data transfer techniques of the described embodiments are not limited to use in such situations. As another example use case, it can be desirable to provide for in-field replacement of memory modules, so that when a memory module is identified to be failing, data can be copied to a backup memory module. Using existing techniques, this requires the data to be read back to the memory controller, from where it is subsequently output via suitable write commands to the backup memory module. However, when using the direct data transfer technique of the described embodiments, the process could be performed much more efficiently using a direct memory module to memory module transfer, thus saving both system bandwidth and reducing margin required for detection of failure. In particular, the process is ideally performed with minimal loss of service to the system and in minimal time to pre-empt total memory module failure. Due to the significant reduction in time required to perform the in-field replacement when using the direct data transfer technique, the margin can be reduced that is used to trigger the in-field replacement.
In one embodiment where one of the memory modules is used as a cache for the other memory module, the apparatus may further comprise a cache controller to control use of said one of the first and second memory modules as a cache, the cache controller being arranged to apply a cache allocation policy, and to issue at least one transfer access request to the memory controller when the cache allocation policy indicates that a block of data should be transferred between the first and second memory modules. Hence, the cache controller can insert transfer access requests into the series of requests received by the memory controller as and when needed, in order to cause data to be transferred between the memory modules as required.
The direct data transfer may take a variety of forms, and hence for example may in one embodiment comprise a data move operation, whilst in another embodiment it may comprise a data copy operation (i.e. where the source data is left in place within the first memory module whilst a copy is stored within the second memory module).
In accordance with one embodiment, there is provided an apparatus comprising: a memory controller; a plurality of memory modules; and a communications infrastructure interconnecting the memory controller and the plurality of memory modules; the memory controller having a shared interface to communicate with said plurality of memory modules; the memory controller arranged to control direct data transfer between a first memory module and a second memory module of said plurality of memory modules, by issuing at least one direct transfer command via the shared interface, and the first memory module and the second memory module being arranged to operate in response to said at least one direct transfer command so as to cause target data to be transferred over the communications infrastructure between the first memory module and the second memory module without the target data being routed via the memory controller.
It will be appreciated that the memory controller and various memory modules employed within an apparatus as described above may be manufactured and sold separately. The memory controller will be configured to be able to transmit the earlier-mentioned first direct transfer command and the second direct transfer command, in order to cause the first and second memory modules to coordinate their activities so as to transfer data directly between them, hence allowing the memory controller to still orchestrate the data transfer, but with the data itself bypassing the memory controller, thus enabling a reduction in the bandwidth requirements, an increase in performance, etc.
In one embodiment, the memory controller has a shared interface to communicate with a plurality of memory modules via a communications infrastructure, and also includes direct transfer control circuitry to control the direct data transfer between the first memory module and the second memory module by causing the required first and second direct transfer commands to be issued via the shared interface. The direct transfer control circuitry can be arranged to be responsive to a transfer access request received by the memory controller that requires data to be transferred between the first and second memory modules, to control generation of the first and second direct transfer commands required in order to cause the direct data transfer to be performed, rather than requiring the data to be returned to the memory controller for onward propagation to the second memory module.
The individual memory modules may be configured so that they can operate either as a source or as a destination for the above described direct data transfer, and hence be responsive to both the first direct transfer command and the second direct transfer command. Alternatively, some memory modules may be constructed so that they always act as the source during a direct data transfer, or always act as a destination during a direct data transfer. Each memory module may comprise direct transfer handling circuitry arranged to perform at least one operation to ensure compliance of the direct data transfer with communication protocol rules of the communications infrastructure interconnecting the memory module with the memory controller and one or more additional memory modules.
Each memory module may consist of a plurality of separate memory devices (also referred to as memory chips). Whilst in some embodiments the direct transfer handling circuitry may be provided centrally for use by all such memory devices, in an alternative embodiment the functionality of the direct transfer handling circuitry may be distributed, so that that direct transfer handling circuitry effectively comprises a separate direct transfer handling component for each of the memory devices.
Particular embodiments will now be described with reference to the Figures.
There are a number of ways in which the requesting devices 10, 15, 20 may be coupled with the memory controller 30, but in the arrangement shown in
The portion of memory 40 can take a variety of forms, but in one embodiment consists of a plurality of memory modules 42, 44, which may for example take the form of dual in-line memory modules (DIMMs). Each DIMM may contain a plurality of separate memory devices (also referred to herein as memory chips).
Whilst in one embodiment a single memory controller 30 may control the entire memory, in some embodiments the memory is partitioned into multiple parts, with each part having an associated separate memory controller and associated physical layer interface. As shown in
In one embodiment, the various requesting devices 10, 15, 20, the interconnect 25, and the memory controllers 30, 45 may all be provided within a System-on-Chip (SoC). The physical layer interface 35, 50 may in one embodiment also be provided on-chip, or may be provided off-chip at the interface with the memory controller 30, 45. In one embodiment, the portions of memory 40, 55 are provided off-chip and connected via connection wires with the physical layer interface 35, 50.
Typically there will be an on-chip communication protocol used for communications routed between the various components. In one embodiment the protocol used on-chip may be different to the protocol used for communication with the portion of memory, which is illustrated schematically in
In one particular example embodiment, the on-chip communication protocol may be the AMBA eXtensible Interface (AXI) protocol developed by ARM Limited, Cambridge, United Kingdom.
The memory controller 30, 45 will in one embodiment not only handle the scheduling of the pending access requests, but also perform protocol conversion to convert from the on-chip communication protocol to the communication protocol used to communicate with the off-chip portions of memory 40, 55. The communication protocol used to communicate with the memory modules will depend on the type of memory used for those memory modules, but in general may require the issuance of a series of commands of different types over a command propagation path, along with the propagation of write and read data over a data path.
The number of DIMMs connected to each channel may be varied dependent on embodiment, and hence in some embodiments more than two DIMMs may share the same channel. Further, the number of separate channels provided can vary dependent on embodiment.
In typical systems it is traditionally the case that all of the memory modules are essentially identical, and together form a uniform address space, potentially interleaved across the various channels and memory modules. Within such an environment, each memory controller is arranged to process read requests by reading data from the relevant memory module and providing that data to a request source, and is arranged to process write requests by outputting write data from a request source to the relevant memory module for storage therein. Which memory module is accessed will depend on the address specified by the request source.
However, the inventors realised that with emerging technologies there is likely to be a need to transfer data from one memory module to another. As an example, within each DIMM, the memory devices may be of the same form, but the memory used within each of the separate DIMMs need not be the same, and indeed within one DIMM the memory used may have a different characteristic to the memory used in another DIMM provided within the same memory channel. As a particular example, with the advent of Storage Class Memory (SCM), it would be possible for at least one of the DIMMs to be constructed using SCM, whilst at least one other of the DIMMs connected to the same memory controller is constructed using a different memory technology, such as traditional DRAM.
Accordingly, it is possible that the DIMMs sharing the same memory channel may have different memory characteristics, whether that be in terms of access performance, density, physical construction, persistency, etc. In such situations, it may be desirable to allow for data to be transferred from one DIMM to another. Using existing techniques, this would involve the memory controller issuing a read command to read data from one DIMM, and then, once that data is received by the memory controller, the memory controller would subsequently output a write command along with that data in order to cause that data to be written to a different DIMM.
As will be discussed in more detail with reference to the subsequent figures, the technique of the described embodiments aims to alleviate the latency and power consumption associated with such data transfers, by providing a mechanism which avoids the data itself needing to be routed through the memory controller, whilst allowing the memory controller to still orchestrate the data transfer.
As mentioned earlier, there may be a variety of reasons to wish to support such a data transfer, but it is expected that such transfer operations will be particularly useful in systems where individual DIMMs within the same memory channel may have different memory characteristics. This could for example allow one relatively small, fast access, DIMM to be used as a cache for another larger, slower access, DIMM.
In the example shown in
In addition to the modified read command sent to the DIMM that is acting as a source DIMM for the transfer, in this example DIMM 144, a modified write command is also issued by the memory controller to the DIMM acting as the destination DIMM for the transfer, in this example the DIMM 042, as indicated by the arrow 85 in
The modified read and write commands comprise information sufficient to cause the source and destination DIMMs to coordinate their activities such that the data read by the source DIMM and output on to the communications infrastructure is received by the destination DIMM 42, thereby avoiding the need for the data to first be returned to the memory controller for subsequent output to the destination DIMM as part of a separate later write command. However, the memory controller 30 still orchestrates the data transfer. In particular, the modified read command and modified write command will cause the source and destination DIMMs 44, 42 to operate somewhat differently to how they would operate in response to normal read and write commands. The use of the modified read and write commands will cause relative timing to be managed between the performance of the read operation by the source DIMM and the performance of the write operation by the destination DIMM to thereby ensure that by the time the destination DIMM performs the write operation, the data that it is required to store is being presented on the communications infrastructure from the source DIMM. The relative timing may be encoded directly within the modified read command and the modified write command, or alternatively that relative timing may be encoded in advance, through programming of an appropriate mode register 48 (shown in
For example, the relative timing information may specify a delay between receipt of a modified read command, and the outputting of the read data on the communications infrastructure, and similarly may identify a delay between receipt of a modified write command, and the performance of the write operation. The memory controller can then coordinate the timing at which it issues the modified read command and the modified write command, so that the relative timing of the read and write activities performed by the source and destination memory modules is such that at the time the destination memory module is seeking to perform its write operation, the required data is presented on the communications infrastructure by the source memory module.
A data transfer controller 90 may be provided within the memory controller to control use of the modified read and modified write commands. In particular, the system shown in
In addition to the relative timing of the operations performed by the source and destination DIMMs being controlled in dependence on the use of the modified read and write commands, the modified read and write commands can also cause one or both of the source and destination DIMMs to perform various other operations aimed at ensuring compliance of the direct data transfer with the communication protocol rules of the communications infrastructure. Normally, the memory controller could directly ensure compliance with those rules, but given that the memory controller, whilst still orchestrating the data transfer, does not itself handle the actual transfer of the data, it is useful to devolve certain operations to the DIMMs themselves in order to ensure compliance of the direct data transfer with the communication protocol rules.
In particular, as shown in
Whilst in
Further, when more than two DIMMs are provided within the same memory channel, it will be appreciated that any one of them may be used as a source and any one of them may be used as a destination for a direct data transfer, as appropriate. Hence, data could for example be transferred via a direct data transfer process, between DIMMs 42 and 46 shown in
A scheduler circuit 120 is used to perform scheduling operations in respect of the contents of the pending access request buffer(s) in order to determine the order in which to issue the pending access requests to the portion of memory. This can take into account the structure and organisation of the portion of memory in order to seek to optimise the performance of the memory accesses. For example, when a particular page is open in DRAM memory, it is typically more efficient to continue to access the contents of that page rather than closing the page and opening another page. Hence, pending access requests can be reordered by the scheduler with the aim of seeking to perform multiple accesses to an open page in the memory. Access execution circuitry 130 is used to issue selected access requests from the buffer 110 (as ordered by the scheduler 120) to the physical layer interface for onward propagation over the physical connection medium to the portion of memory 40.
The physical layer interface 35 includes a command module 145 for issuing the command signal which will identify the type of access and the address being accessed, and a data module 140 for outputting write data to the portion of memory, or for receiving read data from the portion of memory. In some embodiments, enable signals can be used in order to activate the relevant DIMM for any particular command.
A synchronisation control component 150 controls the operation of the command module and the data module in order to maintain synchronous communication over the physical connection between the physical layer interface 35 and the portion of memory 40. The synchronisation control block will include a number of circuits such as phase-lock-loop circuits and delay-lock-loop circuits that are configured to seek to take account of delays, both relative and absolute, in the paths between the physical layer interface and the portion of memory. In one embodiment, a clock signal is issued in association with the command signal from the command module to provide a timing reference, and a strobe signal is issued in association with the data signals to ensure that the data is latched on appropriate clock edges of the clock signal.
Over time, drift can be introduced into the signals propagating over the connection between the physical layer interface 35 and the portion of memory 40 due to a variety of factors, for example changes in operating temperature and voltage, etc. Hence, periodically a training operation may need to be performed on the physical layer interface 35 in order to update the configuration of the synchronisation control block 150 (to adjust phase delays and the like). Physical layer training circuitry 160 can be provided for performing such a training operation as and when required, and can apply a number of processes such as read/write levelling, data eye training, per-bit data deskew control, and PVT (process, voltage, temperature) compensation using any of a number of known techniques. The training circuitry 160 may be arranged to initiate the training operation periodically, for example on elapse of a predetermined time period, or may be arranged to actively monitor the quality of the signals being communicated over the connection with the portion of memory so as to seek to detect when such drifting issues are reaching a point where there is an unacceptable probability that an error may soon occur. There are a number of properties in the signals that could be monitored, for example the signal to noise ratio on data capture. When a threshold is exceeded, then the training operation could be initiated.
Whenever the training operation is performed, it is necessary to stall performance of the pending access requests, and accordingly during the period in which the physical layer interface 35 is undergoing a training operation, no pending access requests will be issued from the access execution circuitry 130.
The training circuit 160 may be a dedicated hardware circuit used to perform the training operation, by liaising with the synchronisation control block 150 in order to cause a series of test commands to be issued to the memory, with the various signals being monitored in order to determine how best to adjust the various synchronisation control parameters maintained within the synchronisation control block 150. However, in an alternative embodiment, the training circuitry 160 may be implemented by a more general purpose processor executing firmware in order to perform the training operation.
As will be discussed in more detail later, in one embodiment the transfer handler circuitry provided within the DIMMs may also be arranged to perform certain physical layer training operations, and may have associated sync control circuitry to store the control values determined during such training operations. This hence enables effective training of the paths within the communications infrastructure used during direct data transfer between memory modules.
As shown in
However, if at step 205 it is determined that the source and destination addresses are in different DIMMs, then at step 215 the data transfer controller 90 can replace the transfer request with modified read and write commands within the buffer(s) 110. In addition, at step 220, it will inform the scheduler that the modified read command and the modified write command need to be scheduled together, in order to ensure that a desired relative timing between those read and write commands occurs, so that the data output by the source DIMM onto the communications infrastructure will be present on the communications infrastructure at the correct time for it to then be received by the destination memory module whilst processing the modified write command.
The modified read and write commands can take a variety of forms, but an example format for each is illustrated schematically in
Another portion 254 of the command 250 will provide an indication of the source address. As with normal read and write commands, certain control information can also be provided in a field 256 if desired, for example to identify if auto precharge is to be performed in order to close a memory page once the access is finished, to provide a burst chop indication to indicate whether a burst is to be split into multiple portions, etc. Additionally, in one embodiment an optional field 258 may provide an indication of an additive timing to be used, for example to identify the delay between receipt of the read command and the clock cycle in which the read data is to be output onto the communications infrastructure.
As shown in
Whilst in
The various DIMMs connected to the memory channel can take a variety of forms, but will typically consist of multiple memory devices (also referred to herein as chips) which collectively form the memory of the DIMM.
In embodiments where a wide data bus is used, then it is possible in an alternative configuration to distribute the transfer handling functionality amongst the separate memory devices, as illustrated in
In a yet further alternative embodiment, a hybrid between the approaches of
Once a command has been received, then at step 455 it is determined whether the received command indicates a direct data transfer operation, which in one embodiment can be determined from the form of the command received. For example, if a modified read command or a modified write command is received, then this indicates that a direct data transfer operation is required. If not, then at step 460 the DIMM performs normal read/write processing and the CRC generator is not required to perform any CRC code generation activity.
If at step 455 it is determined that the command does indicate a direct transfer operation, then at step 465 it is determined whether the command is one of the earlier mentioned modified read commands. If so, then the process proceeds to step 470 where the required read data is obtained from the source address, i.e. is read from the relevant memory device within the DIMM. In addition, the DIMM will refer to the required additive timing information, whether that is specified directly within the modified read command, or is specified in a particular mode register referred to by the DIMM when performing modified read operations, in order to determine when the read command needs processing, and in particular when the read data should be transmitted on to the communications infrastructure interconnecting the memory modules and memory controller.
Once the read data has been obtained, then at step 475 the CRC generator 405 generates a CRC code for the read data, whereafter at step 480 the read data is transmitted over the communications infrastructure during the clock cycle identified by the additive timing, so that that read data, can then be obtained by the DIMM containing the destination address. That transmitted read data will also be accompanied by the generated CRC code. The time at which the CRC code is transmitted will vary dependent on embodiment, but in one embodiment may be transmitted in the clock cycle following the read data, with the destination DIMM then receiving the read data and subsequently receiving the CRC code so that it can perform a check as to the accuracy of the data obtained from the communications infrastructure.
If it is determined at step 465 that the command is not a modified read command, then in one embodiment this will mean that the command is a modified write command, and the process will proceed to step 482 where the destination DIMM processing the modified write command will await receipt of the data and associated CRC code from the source DIMM. Again, the required additive timing will be referred to in order to determine at what point the destination DIMM should expect to receive the data from the communications infrastructure interconnecting the source and destination DIMMs. Once the data and CRC code have been received, then at step 484 the destination DIMM will perform a CRC code check in order to then determine at step 486 whether the CRC is passed or not. Assuming the check is passed, then the data is written to the destination address at step 488, whereas otherwise an error is raised at step 490. The error may be raised in a variety of ways, but in one embodiment involves notifying the memory controller via an alert pin provided at the memory controller's interface. Such a signal will then for example cause the memory controller to seek to fix the problem that caused the data transmission to fail, for example by causing the relevant paths within the physical interface to be retrained, whereafter the data transfer can be retried.
The physical training circuit and sync control blocks within the individual DIMMs operate in an analogous manner to the physical training circuitry 515 and sync control block 512 associated with the memory controller 505, and accordingly can perform the general functions discussed earlier when discussing
Whilst for ease of illustration
When the trigger condition is detected, then at step 605 the relevant phy training circuit advises the memory controller that it is entering the phy training mode. This causes the memory controller to temporarily suspend propagation of commands over the communications infrastructure so as to enable the training to take place.
At step 610 it is then determined which path or paths physical training needs to be performed upon. This may be directly determinable from the nature of the trigger condition, which as mentioned earlier could for example be associated with the signal quality on a particular path deteriorating to a threshold level, or could be specified in a variety of different ways.
Following a determination of the paths to be trained, then for a first path to be tested a training sequence can be performed in order to cause a series of test commands to be issued over the relevant path, with the various signals being monitored in order to determine how best to adjust the various synchronisation control parameters maintained within the associated synchronisation control block. Accordingly as a result of performing the training sequence, the timing control parameters required to be maintained by the synchronisation control block can be assessed. Thereafter, at step 620 the synchronisation control settings are updated within the synchronisation control block for the associated tested path, whereafter at step 625 it is determined whether there are any more paths to test. If so, the process returns to step 615 to test the next path, whereas otherwise the process proceeds to step 630 where the memory controller is advised that the physical training mode is being exited, whereafter the process returns to step 600.
The exact form of the training sequences required will depend on the way in which the commands and associated data are propagated over the interconnecting wires between the memory controller and associated memory modules. In the example discussed earlier with reference to
The exact mechanisms employed within the physical training circuits associated with the DIMMs can vary dependent on embodiment. For example, considering modern DRAMs (e.g. DDR4), a “write levelling” mode is offered, which allows the DRAM device to measure the compensation required for flight-time skew between clock and strobe by feeding back the clock signal sampled using the strobe signal onto the data pins, thus allowing an iterative loop to adjust compensation until capture is aligned. Similarly, LRDIMMs (which buffer the data bus between the host and the DRAM) offer equivalent capability. This process requires the transmitter to send pulses on the strobe signals whilst not driving the data bus. To facilitate DIMM to DIMM communications as required by the above described direct data transfer, a new mode could be added to allow a DIMM to generate such pulses so that the receiver DIMM can be trained to the transmitter DIMM. In one embodiment, the memory controller could still be used to evaluate the data responses and adjust calibration settings during the training if desired. However, the evaluation mechanisms could also be moved to the DIMMs themselves. In such an embodiment, the memory controller could then be provided with the ability to train its strobe and data generation to match using a similar process.
As discussed earlier, another function that requires some modification to deal with the fact that during direct data transfer data is communicated directly between DIMMS rather than via the memory controller is that of ODT, which sets resistive values applied by termination components within the memory modules and the memory controller for impedance matching purposes. This can be used to improve signal integrity during data transfer. However, the termination settings required by the memory controller and the various memory modules connected to the same memory channel will vary dependent on whether normal read and write commands are being processed where data is moving between the memory controller and a memory module, or direct data transfer is being performed where data is moving between memory modules without being routed back through the memory controller.
However, if at step 700 it is determined that a direct data transfer operation is required, then at step 710 the DIMMs forming the source and destination for the transfer are identified. Thereafter, at step 715 appropriate termination settings are selected for each termination point coupled to the communications infrastructure. It should be noted that the relevant termination points are in one embodiment not only the source and destination DIMMs, but also any other DIMMs and indeed the memory controller itself, since the termination settings of all of these components may need to be adjusted to take account of the direct data transfer taking place.
Following step 715, then the termination settings are output for each termination point, and in order to cause transfer termination to be employed that is tailored to the particular direct data transfer being performed.
Once at step 755 it is determined that a direct data transfer operation is required, then the termination points determines whether the source address is within the address range serviced by that termination point. If so, then preconfigured termination settings are applied at step 770 that have been set up for use when the DIMM is acting as the source DIMM.
If the source address it not within the address range of the DIMM performing the process shown in
If it is determined at step 775 that the destination address is also not associated with the address range of the DIMM, then at step 785 the DIMM applies preconfigured termination settings for use when that DIMM is neither a source or a destination for the direct data transfer. In one embodiment, the DIMM may be arranged to use the same preconfigured termination settings at step 785 irrespective of which other memory modules are being used as the source memory module and the destination memory module, but alternatively it is possible that the termination settings applied may be dependent on which memory module is being used as a source and which memory module is being used as a destination.
It will be appreciated that various types of termination methodologies may be used depending on how the resistors implementing the termination are connected to the relevant transmission line of the communication infrastructure. The termination settings will be determined accordingly, in order to ensure that the correct amount of resistance is applied at each termination point to seek to ensure that the impedance of every location on the transmission line is uniform throughout its length in order to thereby minimise distortion of the signal propagated over the transmission line.
As shown in
In addition, the cache controller can apply normal cache allocation policies in order to decide when to allocate data into the DRAM DIMM, and when to evict data from the DRAM DIMM back to the SCM DIMM. Data transfer requests can be issued by the cache controller to the memory controller in such instances, specifying the required source and destination addresses, where either the source or the destination address is associated with the DRAM DIMM, and the other of the source and destination address is associated with the SCM DIMM. This process is illustrated schematically by the flow diagram of
In particular, at step 850 the cache controller 805 applies the cache allocation policy in order to identify whether data should be moved into or out of the DRAM DIMM. As soon as it is determined that such a move should take place, then at step 855 an appropriate move access request is created and forwarded to the memory controller. As discussed earlier with reference to
Whilst the approach of
From the above described embodiments, it will be seen that such embodiments, through the provision of the direct data transfer technique described, enable lower latency, lower power consumption and higher bus efficiency in systems where memory modules having different memory characteristics occupy the same memory channel. For example, in such instances one of the memory modules may be used as a software or hardware managed cache for another one of the memory modules.
However, the direct data transfer technique can also have benefits in other embodiments where it may not necessarily be the case that the individual memory modules have different characteristics. For example, it is known to provide for in-field replacement of DIMMs. In accordance with known techniques, when a DIMM is identified to be failing, typically the data is copied to a back-up DIMM. This process is ideally performed with minimal loss of service to the system and in minimal time to pre-empt total DIMM failure. However, in accordance with existing techniques, it would be necessary to issue a series of read and write commands in order to read the data out of one DIMM back to the memory controller, and then to subsequently issue that data from the memory controller to the replacement DIMM. However, this “copy to replacement” process could be achieved much more quickly using the direct data transfer mechanism of the described embodiments, thus saving both system bandwidth and reducing the margin required before triggering the in-field replacement process.
The enhanced training capabilities offered by certain of the described embodiments can also offer additional benefits. Using existing techniques, each DIMM is typically trained individually, and the memory controller may require additional cycles between transfers to separate DIMMs so as to adjust its delay circuitry for different settings. However, by providing the refined training mechanisms described earlier where individual DIMMs may also initiate training sequences, this may offer the ability to reduce the need for differing delays, thereby improving bus utilisation.
In the present application, the words “configured to . . . ” are used to mean that an element of an apparatus has a configuration able to carry out the defined operation. In this context, a “configuration” means an arrangement or manner of interconnection of hardware or software. For example, the apparatus may have dedicated hardware which provides the defined operation, or a processor or other processing device may be programmed to perform the function. “Configured to” does not imply that the apparatus element needs to be changed in any way in order to provide the defined operation.
Although illustrative embodiments of the invention have been described in detail herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various changes, additions and modifications can be effected therein by one skilled in the art without departing from the scope and spirit of the invention as defined by the appended claims. For example, various combinations of the features of the dependent claims could be made with the features of the independent claims without departing from the scope of the present invention.