CN110049104A - Hybrid cache method, system and storage medium based on layering on-chip interconnection network - Google Patents
Hybrid cache method, system and storage medium based on layering on-chip interconnection network Download PDFInfo
- Publication number
- CN110049104A CN110049104A CN201910198728.2A CN201910198728A CN110049104A CN 110049104 A CN110049104 A CN 110049104A CN 201910198728 A CN201910198728 A CN 201910198728A CN 110049104 A CN110049104 A CN 110049104A
- Authority
- CN
- China
- Prior art keywords
- data
- kernel
- controller
- cluster
- processor
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L67/00—Network arrangements or protocols for supporting network services or applications
- H04L67/50—Network services
- H04L67/56—Provisioning of proxy services
- H04L67/568—Storing data temporarily at an intermediate stage, e.g. caching
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Memory System Of A Hierarchy Structure (AREA)
Abstract
The invention discloses hybrid cache method, system and storage medium based on layering on-chip interconnection network, method includes: the request of data sent according to the first kernel, the query caching data in the first cache unit;Confirmation do not inquire it is data cached after, send request of data, and query caching data in other kernels to shared bus;Confirmation do not inquire it is data cached after, by first node controller to contents controller send request of data;The second kernel cluster is confirmed by contents controller;It is sent a query to by second node controller to contents controller data cached;By contents controller by it is described inquire data cached feed back to first node controller;By first node controller by it is described inquire data cached feed back to the first kernel.Present invention reduces storage overhead and and realize difficulty, can be widely applied to technical field of integrated circuits.
Description
Technical field
The present invention relates to technical field of integrated circuits, be based especially on layering on-chip interconnection network hybrid cache method,
System and storage medium.
Background technique
Cache is the single-level memory being present between main memory and processor, for mitigate processor and memory it
Between gaps between their growth rates.In multi-core processor, since multiple kernels may be written and read same data block, when in some
When the shared data block of verification carries out write operation, the copy of the data block will obsolesce in the cache of other kernels
Data, at this moment the data block in system in the cache of each kernel just will appear inconsistent, that is, the high speed often said is slow
Deposit inconsistence problems.For this problem, bus monitoring agreement and the cache coherence based on catalogue are generally used at present
Agreement solves, in which:
1, bus monitoring agreement is by way of bus monitoring come maintaining cached consistency.It allows system
In each processor can monitor all operations carried out to memory, if these operations destroy in local cache number
According to coherency state, local kernel will to bus send control signal, other kernels listen to control signal after can make
Corresponding consistency movement.Each cache has the status indicator of oneself to show whether current cache line data have
Effect.When local kernel will be written and read, whether the corresponding cache line data of first lookup is effective, carries out if effectively
Read-write operation, while an invalid signals are broadcasted in bus, it, will be corresponding high when other kernels listen to the invalid signals
Fast cache line data deactivates.Otherwise when local kernel will be written and read, if the number of the cache in local kernel
It, then will be to one data request signal of bus broadcast, when other kernels listen to the signal, and the height of corresponding address according to invalid
Fast cache lines are effective, then send corresponding data to request data local kernel.
It 2, is that cache coherence maintenance command is only issued into storage phase based on the cache coherent protocol of catalogue
With those of data block duplicate cache.Catalogue in the cache coherent protocol based on catalogue is used to store system
In it is all it is interior check the cachelines occupy situations, and asking for the same data access of all interior verifications is guaranteed by catalogue
Ask serial execution.When local kernel reads and writes data to local cache, if not corresponding in local cache
Data copy then sends request signal to catalogue, and contents controller will determine which of system high speed by searching for catalogue
Caching possesses the data copy, and request signal is forwarded to the cache.When the cache for possessing the data copy connects
When receiving this data request signal, it is slow by the local high speed that latest data is transmitted to request data by data transmission network
It deposits.Cache coherent protocol based on catalogue is suitable for the multi-core processor using on-chip interconnection network framework.
In on-chip interconnection network framework and its corresponding cache coherent protocol based on catalogue, each node it
Between using grouping routing by the way of be interconnected communication, provide the ability of good scalability and parallel communications for system,
So that bandwidth increases several orders of magnitude, the problem of bus architecture bandwidth limits has been well solved.And on piece interconnects
Network has changed the long interconnection line in bus architecture into short interconnection line between alteration switch, so that lower power consumption.
Currently, in the case where processor cores are less, it is general using shared bus architecture and using bus monitoring association
It discusses and carrys out maintaining cached consistency;In the case where processor cores are more, then use on-chip interconnection network framework and adopt
With the cache coherent protocol based on catalogue.
But when number of cores increases again, the catalogue in the cache coherent protocol based on catalogue stores will
Very big memory space can be occupied, storage overhead is big and realizes that difficulty is big.
Summary of the invention
In order to solve the above technical problems, it is an object of the invention to: provide a kind of storage overhead it is small and realize difficulty it is small,
Hybrid cache method, system and storage medium based on layering on-chip interconnection network.
On the one hand, the embodiment of the invention provides a kind of hybrid cache methods based on layering on-chip interconnection network, including
Following steps:
According to the request of data that the first kernel is sent, the query caching number in corresponding first cache unit of the first kernel
According to;
Confirm do not inquired in the first cache unit it is data cached after, to the shared bus of the first kernel cluster send data ask
It asks, and the query caching data in other kernels of the first kernel cluster;Wherein, the kernel cluster is by multiple kernels in processor
Composition;
Confirm do not inquired in the first kernel cluster it is data cached after, pass through correspond to the first kernel cluster first node
Controller sends request of data to contents controller;Wherein, the Node Controller is configured at kernel cluster, the contents controller
It is configured at the main memory of processor;
The second kernel cluster is confirmed by contents controller;The second kernel cluster primary storage has the request of data corresponding
It is data cached;
The caching that second node controller by corresponding to the second kernel cluster is sent a query to contents controller
Data;
By contents controller by it is described inquire data cached feed back to first node controller;
By first node controller by it is described inquire data cached feed back to the first kernel.
Further, further include network configuration steps, the network configuration steps the following steps are included:
Multiple kernels of processor are grouped, shared bussing technique is then based on and connects the different kernels in every group
At kernel cluster;Wherein, data communication is carried out by shared bus between the different kernels in each kernel cluster;
The multiple kernel fasciations obtained according to grouping are at on-chip interconnection network;
For each kernel cluster configuration node controller;
For the main memory config directory controller of processor.
Further, the network configuration steps are further comprising the steps of:
For each kernel cluster configuration node memory;
It is stored by node memory data cached on each kernel in kernel cluster.
Further, the network configuration steps are further comprising the steps of:
For the main memory config directory memory of processor;
The directory information of each kernel cluster is stored by catalog memory.
Further, data communication is carried out by router between each kernel cluster in the on-chip interconnection network.
Further, the request of data sent according to the first kernel, in corresponding first cache unit of the first kernel
The step for query caching data, comprising the following steps:
First kernel issues request of data to the first cache controller;
First cache controller is according to request of data query caching data.
Further, described the step for second kernel cluster is confirmed by contents controller, comprising the following steps:
Contents controller is according to the second kernel cluster where request of data query caching data;
According to query result, the directory information of catalog memory is updated;
Request of data is sent to the second kernel cluster by contents controller.
On the other hand, the embodiment of the invention provides a kind of hybrid cache system based on layering on-chip interconnection network, packets
It includes:
Data demand module, the request of data for being sent according to the first kernel, in corresponding first caching of the first kernel
Query caching data in unit;
First sending module, for confirm the first cache unit in do not inquire it is data cached after, to the first kernel cluster
Shared bus sends request of data, and the query caching data in other kernels of the first kernel cluster;Wherein, the kernel cluster by
Multiple kernels composition in processor;
Second sending module, for confirm do not inquired in the first kernel cluster it is data cached after, by corresponding to described the
The first node controller of one kernel cluster sends request of data to contents controller;Wherein, in the Node Controller is configured at
Core cluster, the contents controller are configured at the main memory of processor;
Confirmation module, for confirming the second kernel cluster by contents controller;The second kernel cluster primary storage has described
Request of data is corresponding data cached;
Third sending module, for the second node controller by corresponding to the second kernel cluster to contents controller
What is sent a query to is data cached;
First feedback module, for by contents controller by it is described inquire data cached feed back to first node control
Device processed;
Second feedback module, for by first node controller by it is described inquire data cached feed back in first
Core.
On the other hand, the embodiment of the invention provides a kind of hybrid cache system based on layering on-chip interconnection network, packets
It includes:
At least one processor;
At least one processor, for storing at least one program;
When at least one described program is executed by least one described processor, so that at least one described processor is realized
The hybrid cache method based on layering on-chip interconnection network.
On the other hand, the embodiment of the invention provides a kind of storage mediums, wherein it is stored with the executable instruction of processor,
The instruction that the processor can be performed is when executed by the processor for executing the mixing based on layering on-chip interconnection network
Close caching method.
One or more technical solutions in the embodiments of the present invention have the advantages that the embodiment of the present invention passes through
Shared bus inquiry is data cached in different kernels in same kernel cluster;By contents controller and Node Controller not
With query caching data between kernel cluster;The present invention is mixed with bus monitoring agreement and the cache coherence association based on catalogue
View can be suitable for the processor more than number of cores few processor and number of cores, reduce storage overhead and and realization hardly possible
Degree.
Detailed description of the invention
Fig. 1 is the step flow chart of the embodiment of the present invention;
Fig. 2 is the system structure diagram of the embodiment of the present invention;
Fig. 3 is the shared bus structure schematic diagram of the embodiment of the present invention;
Fig. 4 is the mixed cache structure schematic diagram of the embodiment of the present invention;
Fig. 5 is the state transition diagram of the MEOSI agreement of the embodiment of the present invention.
Specific embodiment
The present invention is further explained and is illustrated with specific embodiment with reference to the accompanying drawings of the specification.For of the invention real
The step number in example is applied, is arranged only for the purposes of illustrating explanation, any restriction is not done to the sequence between step, is implemented
The execution sequence of each step in example can be adaptively adjusted according to the understanding of those skilled in the art.
Referring to Fig.1, the embodiment of the invention provides a kind of hybrid cache methods based on layering on-chip interconnection network, including
Following steps:
According to the request of data that the first kernel is sent, the query caching number in corresponding first cache unit of the first kernel
According to;
Confirm do not inquired in the first cache unit it is data cached after, to the shared bus of the first kernel cluster send data ask
It asks, and the query caching data in other kernels of the first kernel cluster;Wherein, the kernel cluster is by multiple kernels in processor
Composition;
Confirm do not inquired in the first kernel cluster it is data cached after, pass through correspond to the first kernel cluster first node
Controller sends request of data to contents controller;Wherein, the Node Controller is configured at kernel cluster, the contents controller
It is configured at the main memory of processor;
The second kernel cluster is confirmed by contents controller;The second kernel cluster primary storage has the request of data corresponding
It is data cached;
The caching that second node controller by corresponding to the second kernel cluster is sent a query to contents controller
Data;
By contents controller by it is described inquire data cached feed back to first node controller;
By first node controller by it is described inquire data cached feed back to the first kernel.
It is further used as preferred embodiment, further includes network configuration steps, the network configuration steps include following
Step:
Multiple kernels of processor are grouped, shared bussing technique is then based on and connects the different kernels in every group
At kernel cluster;Wherein, data communication is carried out by shared bus between the different kernels in each kernel cluster;
The multiple kernel fasciations obtained according to grouping are at on-chip interconnection network;
For each kernel cluster configuration node controller;
For the main memory config directory controller of processor.
It is further used as preferred embodiment, the network configuration steps are further comprising the steps of:
For each kernel cluster configuration node memory;
It is stored by node memory data cached on each kernel in kernel cluster.
It is further used as preferred embodiment, the network configuration steps are further comprising the steps of:
For the main memory config directory memory of processor;
The directory information of each kernel cluster is stored by catalog memory.
Be further used as preferred embodiment, in the on-chip interconnection network between each kernel cluster by router into
Row data communication.
It is further used as preferred embodiment, the request of data sent according to the first kernel is checked in first
In the first cache unit answered the step for query caching data, comprising the following steps:
First kernel issues request of data to the first cache controller;
First cache controller is according to request of data query caching data.
It is further used as preferred embodiment, described the step for second kernel cluster is confirmed by contents controller, packet
Include following steps:
Contents controller is according to the second kernel cluster where request of data query caching data;
According to query result, the directory information of catalog memory is updated;
Request of data is sent to the second kernel cluster by contents controller.
With the method for Fig. 1 to corresponding, the embodiment of the invention provides a kind of mixing based on layering on-chip interconnection network is slow
Deposit system, comprising:
Data demand module, the request of data for being sent according to the first kernel, in corresponding first caching of the first kernel
Query caching data in unit;
First sending module, for confirm the first cache unit in do not inquire it is data cached after, to the first kernel cluster
Shared bus sends request of data, and the query caching data in other kernels of the first kernel cluster;Wherein, the kernel cluster by
Multiple kernels composition in processor;
Second sending module, for confirm do not inquired in the first kernel cluster it is data cached after, by corresponding to described the
The first node controller of one kernel cluster sends request of data to contents controller;Wherein, in the Node Controller is configured at
Core cluster, the contents controller are configured at the main memory of processor;
Confirmation module, for confirming the second kernel cluster by contents controller;The second kernel cluster primary storage has described
Request of data is corresponding data cached;
Third sending module, for the second node controller by corresponding to the second kernel cluster to contents controller
What is sent a query to is data cached;
First feedback module, for by contents controller by it is described inquire data cached feed back to first node control
Device processed;
Second feedback module, for by first node controller by it is described inquire data cached feed back in first
Core.
With the method for Fig. 1 to corresponding, the embodiment of the invention provides a kind of mixing based on layering on-chip interconnection network is slow
Deposit system, comprising:
At least one processor;
At least one processor, for storing at least one program;
When at least one described program is executed by least one described processor, so that at least one described processor is realized
The hybrid cache method based on layering on-chip interconnection network.
With the method for Fig. 1 to corresponding, the embodiment of the invention provides a kind of storage mediums, wherein being stored with processor can hold
Capable instruction, the executable instruction of the processor are described based on interconnecting on layergram for executing when executed by the processor
The hybrid cache method of network.
A kind of specific implementation step of the hybrid cache method based on layering on-chip interconnection network of the present invention is described below in detail
It is rapid:
All kernels in processor are divided into several groups first by the present invention, in each group by sharing bus for each place
Reason device connects to form multiple clusters, a node of each kernel cluster as network-on-chip, all node conducts of formation
The first layer of system.
Then, data packet is forwarded to form (the i.e. piece online of Noc interference networks by router between kernel cluster and kernel cluster
Network), as the system second layer.
In first layer, all kernels in a kernel cluster are come by sharing bus interconnection using bus monitoring agreement
Safeguard the partial cache consistency in a kernel cluster;
In the second layer, interconnected by Noc, using safeguarded based on the Cache coherency protocol of catalogue all kernel clusters it
Between global buffer consistency.
The present invention by adding controller in each node, for as the communication bridge between two layer protocols, so that
Two kinds of agreements can be effectively combined, and both realize that simple, real-time is good, and preferably adapt to more massive more kernel processes
Device.
The whole system of the embodiment of the present invention is divided into double-layer structure, and first layer is used shared bus structure, supervised using bus
Agreement is listened, the second layer uses on-chip network structure, uses the Cache coherency protocol based on catalogue.Addition section in each node
Base site controller, Node Controller one end connect shared bus, and the other end connects the router of Noc, and the Node Controller is used for
Receive and forwarding consistency signal (i.e. request of data) be into intra-node or Noc network, realize first layer and the second layer it
Between data communication.
It is common, main memory and multiple processing data in each processor by one for storing all data
Kernel composition.The embodiment of the present invention adds contents controller, one end of the contents controller on the hosting node of processor
It is connected on the router of Noc, the other end connection main memory and catalog memory of the contents controller.
In first layer, the data exchange of all cachings and consistency order are all to realize to transmit by bus, each
The Node Controller moment monitors the consistency order in bus, and forwards the command on the router being attached thereto, router
It forwards consistency order into the contents controller of hosting node, realizes two kinds of protocol interactions, specifically includes the following steps:
1, the reading data between the kernel in same kernel cluster:
Specifically, when kernel is cached to it issues read data request, (the first kernel of the present invention is slow to first
Memory cell sends request of data, corresponding data cached to inquire), if in caching without the data of required reading when, will
Occur reading deletion condition.Cache controller can broadcast a reading deleted signal (data i.e. of the present invention in bus at this time
Request), it is used to request data and searches its local cache after other kernels listen to this signal in bus and determine whether to possess most
New data send the data to the caching of request data if possessing newest data, complete read operation.
2, the reading data between different kernel clusters:
Specifically, as shown in Fig. 2, Node represents a node (i.e. kernel cluster);R represents router, for forwarding data,
Realize the communication between node;Memory represents main memory;When kernel is cached to it reads data, local cache does not have required number
According to, and current inner cluster (i.e. the first kernel cluster) in other kernels caching also without this data when, first node controller will
A reading deleted signal (request of data i.e. of the present invention) is generated, and this is read into deleted signal and is sent to by router R
The contents controller of hosting node Memory.After contents controller receives deleted signal, catalog memory missing data is searched
The node (the second kernel cluster i.e. of the present invention) at place, and it is (i.e. of the present invention to node transmission reading deleted signal
Request of data), after the Node Controller (i.e. second node controller) of node where data receives reading deleted signal, by data
It transmits in Noc network, the node (i.e. of the present invention first for initiating request of data is finally sent to by router
Kernel) in, complete read operation.
As shown in Figure 3, it is preferable that in the shared bus structure of a kernel cluster, the present embodiment includes N number of kernel
CoreN, each kernel are owned by respective privately owned caching cache, remain peripheral hardware peripheral in each node;It is described outer
If sharing bus Sharebus by Peripheral Interface Interface connection;Node controller is Node Controller, is used for
Connecting node controller contents controller, Node Controller one end connect Sharebus (i.e. shared bus), other end link road
By device;Local memory is node memory, stores the data of all kernel caches of intra-node.
As shown in Figure 4, it is preferable that in the mixed cache structure of the embodiment of the present invention, first layer is shared bus structure, is made
With bus monitoring agreement;The second layer is on-chip network structure Noc, uses directory protocol, wherein main memory memory is located at Noc's
Center;Kernel cluster node is connected with memory node by the router in Noc;Directory controller represents catalogue
Controller, directory memory represent catalog memory, and the contents controller is the core of second layer directory protocol, use
In the consistency that maintenance is global.
As shown in Figure 5, it is preferable that the monitoring protocols that the first layer bus structures of the embodiment of the present invention use are MOESI association
View.Wherein, M, O, E, S, I respectively represent five kinds of states of the privately owned caching of kernel;Specifically, M modified represents modification
State;O is Owned, and representative possesses state;E is exclusive, represents exclusive state;S is share, represents shared state;I
For invalid, invalid state is represented.
State transition operation is described as follows in Fig. 5:
Probe is read: the purpose for indicating that current inner obtains data copy from other kernels is to read data;
Probe writes: the purpose for indicating that current inner obtains copy from other kernels is in order to which data are written;
It reads hit, write hit: indicating that current inner obtains data copy from local cache;
It reads to lose, write loss: indicating that current inner does not obtain data copy in local cache;
Probe reads hit, probe writes hit: indicating that current inner obtains data copy in the caching of other kernels.
During carrying out write operation for kernel, kernel is to its privately owned cache writing data, if local cache is hit, shape
State may be E, S, M or O, specifically:
1, when state is E or is M, data are directly write in privately owned caching, and the data mode of caching is changed to M;
2, when state is S, data are directly write in privately owned caching, and the data mode of caching is changed to M.Simultaneously to mesh
Record issues inquiry request, finds other kernels for saving the data copy, by the data mode of the caching of other data copies from S
Or O moves to I;
3, when state is O, data are directly write in privately owned caching, and the state of the data of caching is changed to M.While to
Catalogue issues inquiry request, other kernels for saving the data copy is found, by the data mode of the caching of other data copies
I is moved to from S;
Preferably, Node Controller (Node controller) is the Fabric Interface of two layer protocols, and function includes:
1, consistency signal in bus is monitored, and is forwarded a signal in Noc network;
2, the consistency signal of contents controller is received, and signal is broadcast in bus, is safeguarded on node memory
Data.
The function of node memory (Node memory) is the data cached of each kernel inside memory node, works as node
When controller receives the request of data of contents controller, transmit data on the node for initiating request.
The function of contents controller (Directory controller) is the consistency signal of receiving node control, maintenance
Contents controller.
Catalog memory (directory memory) is used to store the directory information in the Cache coherency protocol.
First layer of the present invention using MOESI (M:modified, O:Owned, E:exclusive, S:share, I:
Invalid) monitoring protocols, it is as follows that state shifts concrete condition:
Read hit and be M state: M state shows that the data block is unique correct data block, local cache pair in system
The data block is in exclusive state, and interior verification local cache directly reads data, and data mode is constant.
Read hit and be O state: O state shows that the data block is newest data in system, and one in other kernels
Surely there is the data block copy, interior verification local cache directly reads data, and data mode is constant.
Read hit and be E-state: E-state shows there is latest data in caching and main memory, and kernel is directly to local cache
It reads, state does not change.
Read hit and be S state: S state shows that the data block is latest data in system, and kernel is directly to local cache
It reads, state does not change.
Read hit and be I state: I state shows that the data block is invalid, and local cache is ordered to one reading of bus broadcast first
Middle invalid signals find the latest data for possessing the data, then when system other kernels listen to the signal and by comparison
It is sent in the caching for initiating request data, and data mode is changed into S.
Write hit and for M state: kernel directly writes data into caching, and data mode is constant.
Write hit and for E-state: kernel directly writes data into caching, and data block status is changed to M.
Write hit and for S state: kernel directly writes data into caching, and state is changed to M, while other kernels
The state of the data copy of preservation moves to I from S or O.
Write hit and for O state: kernel directly writes data into caching, and state is changed to M, while other kernels
The state of the data copy of preservation moves to I from S.
Write hit and for I state: needs read newest data from other cachings, and local cache writes life to bus broadcast
Middle invalid signals when other kernels listen to this signal in system, and possess effective data block, then transmit data to and ask
Caching is sought, which is changed to I, state is changed to M after requesting the caching of the data to receive data.
Generally speaking, when kernel a certain in node reads and writes data into caching, cache controller issues five seed types
Request signal: invalidation request signal reads miss request signal, reads hit but data invalid request signal, Write missing request
Signal writes hit but data invalid request signal.It is as shown in table 1 to the rewriting situation of catalogue under different request signals:
Table 1
In conclusion being provided the present invention is based on hybrid cache method, system and the storage medium of layering on-chip interconnection network
A kind of realization multi-core processor cache coherence method, the kernel for being able to solve the support of bus monitoring agreement is few, to bus bar
The problems such as width requires high and directory protocol catalogue storage overhead big;So that bus monitoring agreement and the high speed based on catalogue are delayed
Depositing consistency protocol can be effectively combined, and both realize that simple, real-time is good, and preferably adapt to more massive more kernels
Processor.
It is to be illustrated to preferable implementation of the invention, but the present invention is not limited to the embodiment above, it is ripe
Various equivalent deformation or replacement can also be made on the premise of without prejudice to spirit of the invention by knowing those skilled in the art, this
Equivalent deformation or replacement are all included in the scope defined by the claims of the present application a bit.
Claims (10)
1. the hybrid cache method based on layering on-chip interconnection network, it is characterised in that: the following steps are included:
According to the request of data that the first kernel is sent, the query caching data in corresponding first cache unit of the first kernel;
Confirm do not inquired in the first cache unit it is data cached after, to the shared bus of the first kernel cluster send request of data,
And the query caching data in other kernels of the first kernel cluster;Wherein, the kernel cluster is by multiple core groups in processor
At;
Confirm do not inquired in the first kernel cluster it is data cached after, pass through correspond to the first kernel cluster first node control
Device sends request of data to contents controller;Wherein, the Node Controller is configured at kernel cluster, the contents controller configuration
In the main memory of processor;
The second kernel cluster is confirmed by contents controller;The second kernel cluster primary storage has the corresponding caching of the request of data
Data;
Second node controller by corresponding to the second kernel cluster sends a query to data cached to contents controller;
By contents controller by it is described inquire data cached feed back to first node controller;
By first node controller by it is described inquire data cached feed back to the first kernel.
2. the hybrid cache method according to claim 1 based on layering on-chip interconnection network, it is characterised in that: further include
Network configuration steps, the network configuration steps the following steps are included:
Multiple kernels of processor are grouped, are then based in shared bussing technique connects into the different kernels in every group
Core cluster;Wherein, data communication is carried out by shared bus between the different kernels in each kernel cluster;
The multiple kernel fasciations obtained according to grouping are at on-chip interconnection network;
For each kernel cluster configuration node controller;
For the main memory config directory controller of processor.
3. the hybrid cache method according to claim 2 based on layering on-chip interconnection network, it is characterised in that: the institute
It is further comprising the steps of to state network configuration steps:
For each kernel cluster configuration node memory;
It is stored by node memory data cached on each kernel in kernel cluster.
4. the hybrid cache method according to claim 2 based on layering on-chip interconnection network, it is characterised in that: the institute
It is further comprising the steps of to state network configuration steps:
For the main memory config directory memory of processor;
The directory information of each kernel cluster is stored by catalog memory.
5. the hybrid cache method according to claim 2 based on layering on-chip interconnection network, it is characterised in that: described
Data communication is carried out by router between each kernel cluster in upper interference networks.
6. the hybrid cache method according to claim 1 based on layering on-chip interconnection network, it is characterised in that: described
According to the request of data that the first kernel is sent, in corresponding first cache unit of the first kernel the step for query caching data,
The following steps are included:
First kernel issues request of data to the first cache controller;
First cache controller is according to request of data query caching data.
7. the hybrid cache method according to claim 4 based on layering on-chip interconnection network, it is characterised in that: described logical
Cross the step for contents controller confirms the second kernel cluster, comprising the following steps:
Contents controller is according to the second kernel cluster where request of data query caching data;
According to query result, the directory information of catalog memory is updated;
Request of data is sent to the second kernel cluster by contents controller.
8. the hybrid cache system based on layering on-chip interconnection network, it is characterised in that: include:
Data demand module, the request of data for being sent according to the first kernel, in corresponding first cache unit of the first kernel
Interior query caching data;
First sending module, for confirm do not inquired in the first cache unit it is data cached after, to the shared of the first kernel cluster
Bus sends request of data, and the query caching data in other kernels of the first kernel cluster;Wherein, the kernel cluster is by handling
Multiple kernels composition in device;
Second sending module, for confirm the first kernel cluster in do not inquire it is data cached after, by correspond to described first in
The first node controller of core cluster sends request of data to contents controller;Wherein, the Node Controller is configured at kernel cluster,
The contents controller is configured at the main memory of processor;
Confirmation module, for confirming the second kernel cluster by contents controller;The second kernel cluster primary storage has the data
It requests corresponding data cached;
Third sending module is sent for the second node controller by corresponding to the second kernel cluster to contents controller
What is inquired is data cached;
First feedback module, for by contents controller by it is described inquire data cached feed back to first node control
Device;
Second feedback module, for by first node controller by it is described inquire data cached feed back to the first kernel.
9. the hybrid cache system based on layering on-chip interconnection network, it is characterised in that: include:
At least one processor;
At least one processor, for storing at least one program;
When at least one described program is executed by least one described processor, so that at least one described processor is realized as weighed
Benefit requires the hybrid cache method based on layering on-chip interconnection network described in any one of 1-7.
10. a kind of storage medium, wherein being stored with the executable instruction of processor, it is characterised in that: the processor is executable
Instruction be used to execute when executed by the processor as it is of any of claims 1-7 based on layering on-chip interconnection network
Hybrid cache method.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910198728.2A CN110049104A (en) | 2019-03-15 | 2019-03-15 | Hybrid cache method, system and storage medium based on layering on-chip interconnection network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910198728.2A CN110049104A (en) | 2019-03-15 | 2019-03-15 | Hybrid cache method, system and storage medium based on layering on-chip interconnection network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110049104A true CN110049104A (en) | 2019-07-23 |
Family
ID=67273805
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910198728.2A Pending CN110049104A (en) | 2019-03-15 | 2019-03-15 | Hybrid cache method, system and storage medium based on layering on-chip interconnection network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110049104A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111858096A (en) * | 2020-07-22 | 2020-10-30 | 中国人民解放军国防科技大学 | Method and system for monitoring reading of nearest cache based on directory |
CN113792006A (en) * | 2020-08-14 | 2021-12-14 | 阿里巴巴集团控股有限公司 | Inter-device processing system with cache coherency |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1506845A (en) * | 2002-12-10 | 2004-06-23 | 英特尔公司 | Isomeric proxy cache memory consistency and method and apparatus for limiting transmission of data |
CN1545034A (en) * | 2003-11-26 | 2004-11-10 | 中国人民解放军国防科学技术大学 | Double ring method for monitoring partial cache consistency of on-chip multiprocessors |
CN101042680A (en) * | 2006-03-23 | 2007-09-26 | 国际商业机器公司 | Data processing system, cache system and method for updating an invalid coherency state |
CN101978659A (en) * | 2008-04-02 | 2011-02-16 | 英特尔公司 | Express virtual channels in a packet switched on-chip interconnection network |
CN102929832A (en) * | 2012-09-24 | 2013-02-13 | 杭州中天微系统有限公司 | Cache-coherence multi-core processor data transmission system based on no-write allocation |
US20130254488A1 (en) * | 2012-03-20 | 2013-09-26 | Stefanos Kaxiras | System and method for simplifying cache coherence using multiple write policies |
CN103440223A (en) * | 2013-08-29 | 2013-12-11 | 西安电子科技大学 | Layering system for achieving caching consistency protocol and method thereof |
-
2019
- 2019-03-15 CN CN201910198728.2A patent/CN110049104A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1506845A (en) * | 2002-12-10 | 2004-06-23 | 英特尔公司 | Isomeric proxy cache memory consistency and method and apparatus for limiting transmission of data |
CN1545034A (en) * | 2003-11-26 | 2004-11-10 | 中国人民解放军国防科学技术大学 | Double ring method for monitoring partial cache consistency of on-chip multiprocessors |
CN101042680A (en) * | 2006-03-23 | 2007-09-26 | 国际商业机器公司 | Data processing system, cache system and method for updating an invalid coherency state |
CN101978659A (en) * | 2008-04-02 | 2011-02-16 | 英特尔公司 | Express virtual channels in a packet switched on-chip interconnection network |
US20130254488A1 (en) * | 2012-03-20 | 2013-09-26 | Stefanos Kaxiras | System and method for simplifying cache coherence using multiple write policies |
CN102929832A (en) * | 2012-09-24 | 2013-02-13 | 杭州中天微系统有限公司 | Cache-coherence multi-core processor data transmission system based on no-write allocation |
CN103440223A (en) * | 2013-08-29 | 2013-12-11 | 西安电子科技大学 | Layering system for achieving caching consistency protocol and method thereof |
Non-Patent Citations (1)
Title |
---|
汤伟等: ""基于总线监听的Cache一致性协议分析"", 《福建电脑》 * |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111858096A (en) * | 2020-07-22 | 2020-10-30 | 中国人民解放军国防科技大学 | Method and system for monitoring reading of nearest cache based on directory |
CN111858096B (en) * | 2020-07-22 | 2022-09-23 | 中国人民解放军国防科技大学 | Directory-based method and system for monitoring reading of cache at shortest distance |
CN113792006A (en) * | 2020-08-14 | 2021-12-14 | 阿里巴巴集团控股有限公司 | Inter-device processing system with cache coherency |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN103440223B (en) | A kind of hierarchical system and its method for realizing cache coherent protocol | |
US6631448B2 (en) | Cache coherence unit for interconnecting multiprocessor nodes having pipelined snoopy protocol | |
US7814279B2 (en) | Low-cost cache coherency for accelerators | |
CN104508637B (en) | Method for reciprocity Buffer forwarding | |
JP5833282B2 (en) | Multi-tier cache coherency domain system and method of configuring Share-F state in local domain of multi-tier cache coherency domain system | |
US7017011B2 (en) | Coherence controller for a multiprocessor system, module, and multiprocessor system with a multimodule architecture incorporating such a controller | |
JP3644587B2 (en) | Non-uniform memory access (NUMA) data processing system with shared intervention support | |
EP1316019B1 (en) | Managing replacement of data in a cache on a node based on caches of other nodes | |
US20040236912A1 (en) | Methods and apparatus for providing cache state information | |
US20070055826A1 (en) | Reducing probe traffic in multiprocessor systems | |
JP3686221B2 (en) | Computer system for maintaining cache coherency | |
US20030212741A1 (en) | Methods and apparatus for responding to a request cluster | |
JPH11134312A (en) | Decentralized shared memory multiprocessor system | |
CN103119568A (en) | Extending a cache coherency snoop broadcast protocol with directory information | |
KR20000052493A (en) | Non-uniform memory access (numa) data processing system having shared intervention support | |
AU2001289180A1 (en) | Managing replacement of data in a cache on a node based on caches of other nodes | |
WO2014094374A1 (en) | Method for constructing multiprocessor system with node having a plurality of cache uniformity domains | |
US6950913B2 (en) | Methods and apparatus for multiple cluster locking | |
US20040088496A1 (en) | Cache coherence directory eviction mechanisms in multiprocessor systems | |
US7308538B2 (en) | Scope-based cache coherence | |
US6925536B2 (en) | Cache coherence directory eviction mechanisms for unmodified copies of memory lines in multiprocessor systems | |
WO2014146424A1 (en) | Method for server node data caching based on limited data coherence state | |
CN110049104A (en) | Hybrid cache method, system and storage medium based on layering on-chip interconnection network | |
US7653790B2 (en) | Methods and apparatus for responding to a request cluster | |
US20050033924A1 (en) | Methods and apparatus for providing early responses from a remote data cache |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190723 |