CN105183693A - Multicast transmission method based on three-dimensional network on chip - Google Patents
Multicast transmission method based on three-dimensional network on chip Download PDFInfo
- Publication number
- CN105183693A CN105183693A CN201510272323.0A CN201510272323A CN105183693A CN 105183693 A CN105183693 A CN 105183693A CN 201510272323 A CN201510272323 A CN 201510272323A CN 105183693 A CN105183693 A CN 105183693A
- Authority
- CN
- China
- Prior art keywords
- subregion
- node
- child partition
- network
- label
- 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.)
- Granted
Links
Landscapes
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention discloses a multicast transmission method based on a three-dimensional network on chip. According to the multicast transmission method, two delay performance parameters of starting delay and network delay are considered according to the size of a transmitted data packet and the size of a network structure; optical sub-area structures are divided around source nodes according to the positions of the source nodes in the network; and routing transmission is performed on information in each sub-area based on completing of sub-area dividing. The multicast transmission method has functions of overcoming a defect of a total delay in network transmission, selecting an optimal area dividing method according to the current network state, and finishing routing transmission of each sub-area according to the area dividing method, thereby equalizing the path length of each sub-area in the network, reducing the total delay in multicast transmission as possible, and optimizing a parallel communication performance. Along with an increase of the data packet size of transmitted information and a network topology structure, a delay reduction performance of the network in which the multicast transmission method is applied is improved, thereby improving communication efficiency of the continuously expanded three-dimensional network on chip.
Description
Technical field
The present invention relates to integrated circuit fields, particularly relate to a kind of multicast transmission method based on network on three-dimensional chip.
Background technology
Along with the development of integrated circuit technique, and the increase of SOC (system on a chip) (SoC) scale and complexity, on-chip interconnect has become the bottleneck that system performance promotes.Network on three-dimensional chip (3DNoC) is the Perfected process solving on-chip interconnect design bottleneck.In a network, the performance of message transmission mode on global system has larger impact, multicast transmission can make full use of multiprocessor systems on chips (Multiprocessorsystem-on-chips, MPSoCs) concurrency provided carrys out transmission information efficiently, therefore applies very extensive in 3DNoC communication.
Before the present invention makes, the transmission total delay in system comprises start delay and transmission delay two aspects, and in order to reduce to transmit total delay, improve network performance, network partition is essential based on the application in the multicast in path.The quality of partition method directly affects the system performance of network-on-chip.At present, partition method in multicast transmission is mainly obtained by fixed partition number and partitioned mode, as two subregion (Two-blockpartitioning, TBP), multi partition (Multi-blockpartitioning, MBP), vertical partitioning (Vertical-blockpartitioning, VBP) etc., the communication requirement that the performances such as the resource overhead of such partition method, transmission delay can not be adapted to the MPSoCs of expanding day is applied.Particularly in transmission delay, major defect shows that the partition method in existing transmission mode only considers the one-side impact of transmission delay, and have ignored the impact of start delay aspect on transmission total delay, namely have ignored in network and transmit data package size and information source node location to the impact of network partition result, secure subregion number, thus dynamically can not change subregion number and partitioned mode to reduce transmission total delay according to real network parameter status, cause obtaining sub-optimal result.There is following Railway Project in the multicast transmission method applying such partition method: (a) is when in the partition method selected, subregion number is fixed less, the transmission path of each subregion is longer, network delay is comparatively large, thus transmission total delay is increased, and transfer efficiency is lower.B (), when in the partition method selected, number of partitions is fixed larger, the injection competition of source node is comparatively large, and overlapping trees is shorter, and start delay increases, thus make to transmit total delay increase, also can cause the wasting of resources while causing transfer efficiency lower.C (), when source node is positioned at network edge locations, each partition network Imbalance in weight, thus increase path-length, causes network delay to increase, and then transmission total delay is increased, and have impact on the balance of transfer efficiency and network.
Therefore, in order to improve information transfer efficiency in network, research is applicable to the multicast transmission method of heterogeneous networks state, enable its network partition from the angle of total delay, according to the length of packet and the balance start delay of network size and these two aspects of network delay, dynamically select the optimum partition number of current network, and divide optimum partitioned organization according to source node location, information on the basis of this partition method in route each sub point, thus to reduce Internet Transmission total delay, it is a kind of important means optimizing and improve network parallel communication performance.
Summary of the invention
Object of the present invention is just to overcome above-mentioned defect, develops a kind of multicast transmission method based on network on three-dimensional chip.
The invention discloses a kind of multicast transmission method based on network on three-dimensional chip, the packet of the method according to transmission and the size of network structure, the optimum partition number of current network is selected on balance start delay and these two delay performance dynamic state of parameters ground of network delay, and according to source node position in a network, optimum partitioned organization is divided centered by source node, route is carried out to the information during each height divides in the basis completing subregion, thus reduce the total delay of multicast transmission as much as possible, improve network transmission efficiency, optimize parallel communications performance.Along with the data package size of transmission information and the expansion of network topology structure, the performance of network in reduction delay applying this transmission method also can improve thereupon, is conducive to the communication efficiency of the network on three-dimensional chip improving constantly expansion.
Realize technical scheme of the present invention as follows:
Based on a multicast transmission method for network on three-dimensional chip, comprise the steps:
A) network layer at source node place is found out, in four frontier points of source node layer, determine arbitrarily that one for initial point, specify initial X, Y, Z axis direction according to this initial point, and according to source node position by corresponding label distribution method be this layer of all peer distribution label, from 0 label;
B) on the basis of distributing labels, apply two zone method and global network is divided into two parts centered by source node: high subregion and low subregion, and find out larger subregion and less subregion in these two subregions by the node number compared in subregion;
C) according to the data packet length of transmission, try to achieve the optimum child partition number compared with bigdos compared with the network size of bigdos and network flow line structure level and judge that comparatively bigdos is the need of subregion, and the optimal maximum node number of child partition in this subregion and minimum node number is obtained;
D) judge whether less subregion should be divided, and try to achieve the optimal maximum node number of child partition and minimum node number in the optimum child partition number of less subregion and this subregion according to the data packet length of transmission, the network size of less subregion and network flow line structure level;
E) in the subregion needing division subregion, find the initial row of division and the stripe sequence of source node straton subregion respectively;
F) respectively need divide subregion subregion in the current child partition of initial division;
G) judge the columns of current child partition on the last one deck of this subregion, make node number in child partition as much as possible close to ideal value, and divide and arrange to this child partition accordingly, complete whole divisions of current child partition;
H) repeat e), f), operation g) until in this subregion penultimate child partition subregion complete, by the node division of all for this subregion remainders in last child partition of this subregion.So far, divide Division complete.
I) on the basis that network partition completes, route is carried out to the information in each child partition, complete the transmission of information in network.
In order to efficiently divide network easily, prevent the transmission path between each subregion from intersecting with deadlock prevention, all peer distribution labels only to source node layer in the method simultaneously.But, when source node is in longitudinal axis border, i.e. y
s=0 (x
s≠ x
max) or y
s=y
max(x
s≠ x
max) time, other layer cannot provide logical path make the information of source node be sent to each subregion.Therefore, first should according to the method for source node present position adjustment distributing labels, make source node not be in longitudinal axis boundary position, its step is as follows:
A1) find out the network layer at source node place, in four frontier points of source node layer, determine arbitrarily that one for initial point, specifies initial X, Y, Z axis direction according to this initial point;
A2) if the y-axis coordinate of source node equals 0 or y
max, x-axis coordinate is not equal to x
max, and be not initial point (0,0), all x coordinates are the node label of even number is L (x, y)=b × x+y, after label distribution, exchanges X, Y-axis;
A3) if the y-axis coordinate of source node equals 0 or y
max, x-axis coordinate is not equal to x
max, and be not initial point (0,0), all x coordinates are the node label of odd number is L (x, y)=b × x+b-y-1, after label distribution, exchanges X, Y-axis;
A4) if the y-axis coordinate of source node is not equal to 0 or y
max; Or y-axis coordinate equals 0 or y
maxand x-axis coordinate equals x
max, all y coordinates are the node label of even number is L (x, y)=a × y+x;
A5) if the y-axis coordinate of source node is not equal to 0 or y
max; Or y-axis coordinate equals 0 or y
maxand x-axis coordinate equals x
max, all y coordinates are the node label of odd number is L (x, y)=a × y+a-x-1;
A6) if source node is initial point (0,0), all y-axis coordinates are the node label of odd number is L (x, y)=a × y+x, exchanges X-direction after label distribution;
A7) if source node is initial point (0,0), all y-axis coordinates are the node label of even number is L (x, y)=a × y+a-x-1, exchanges X-direction after label distribution.
Two zone method (TBP) by the node division on the same direction of source node in a region, can make transmission path not easily produce of deadlock and the competition between child partition.In most of the cases, the most longer transmission path in network is all produce from larger subregion, therefore needs finding out compared with bigdos and less subregion in the high subregion obtained and low subregion.These process concrete steps are as follows:
B1) label being greater than the node division that the node of source node label and all z-axis coordinates be greater than source node z-axis coordinate is high subregion.
B2) label being less than the node division that the node of source node label and all z-axis coordinates be less than source node z-axis coordinate is low subregion.
B3) nodes in higher subregion and low subregion, finds out the more subregion of nodes as comparatively bigdos, finds out the less subregion of nodes as less subregion.
In multiprocessor systems on chips, the transmission total delay of network-on-chip middle ideal can use equation
represent.Wherein, n and N represents the total nodes in subregion number and network respectively.L represents the size of packet, and m represents that network is m stage pipeline structure.N × L is start delay,
for network delay.To this expression formula minimizing, optimum subregion number can be obtained.Thus, try to achieve compared with subregion number optimum in bigdos, and the concrete steps of the minimum and maximum node number of each child partition middle ideal are as follows:
C1) the comparatively total number N of bigdos interior joint is calculated
1, to desired transmission total delay formula minimizing, obtain compared with subregion number optimum in bigdos be
n
1get the most contiguous integer, n
1be greater than 1 needs subregion, otherwise comparatively bigdos does not need subregion.
C2) if the subregion number of optimum is greater than total columns, so optimum partition number n is adjusted
1for total columns x
max+ 1.
C3) be calculated as compared with the minimum and maximum node number of child partition middle ideal each in bigdos
After trying to achieve the minimum and maximum node number compared with bigdos middle ideal, just can to obtain in this subregion the longest path of last node in from source node to child partition, this path is path the longest in network, therefore, the subregion number in less subregion needs to judge according to this parameter.Concrete steps are as follows:
D1) total path length in less subregion is compared with compared with the longest path length in bigdos, the total node number N namely in less subregion
2with
size.If N
2be not more than compared with the longest path path length in bigdos, then less subregion does not need further subregion, i.e. the optimum partition number n of less subregion
2be 1.
D2) if N
2be greater than compared with the longest path path length in bigdos, then the optimum partition number of this subregion is calculated as
D3) in less subregion, the minimum and maximum node number of each child partition middle ideal is calculated as
In order to deadlock prevention, child partition should divide row to ensure that the transmission path between each child partition mutually disjoints by the order of specifying, and the order of different points of Division row is also different.Concrete steps are as follows:
E1) if the subregion being about to divide is high subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be ascending order, namely the node label of last column of the initial row of the first child partition is minimum in this row node label;
E2) if the subregion being about to divide is low subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be descending, namely the node label of the first row of the initial row of the first child partition is maximum in this row node label;
The stripe sequence specified by place subregion respectively needs the subregion of Further Division to carry out the initial division of current child partition to each.Concrete steps are as follows:
F1) by the child partition that needs in the subregion of Further Division by appointment stripe sequence divided by column, from source node layer, divide row;
Whether the row f2) judging in this subregion, adjacent layer to have divided with current layer last row x-axis of row identical are divided;
F3) if these row are divided, then by stripe sequence standardized row again on current layer.Then f2 is returned) step judges;
F4) if these row are not divided, then divide in these these row of subregion adjacent layer to current child partition, and using adjacent layer as current layer;
F5) judging in this subregion that whether every layer is divided all, if also have layer not to be divided in this subregion, then returns f2) step continues to divide.Otherwise terminate the initial division of this child partition;
Initialization divides the fundamental sequence can only determining current child partition each layer in subregion, crucial determinant partition size be columns in subregion on last one deck, need according to the optimum partition number of child partition in each tried to achieve subregion, and desirable minimum and maximum node number, judge the optimum columns of current child partition in this subregion on last one deck, make node number in child partition as much as possible close to ideal value, and division arranges accordingly to this child partition, completes whole divisions of current child partition.Concrete steps are as follows:
G1) calculate the number that this child partition current has node, comprise the nodes of the row that the last one deck of this subregion has divided;
G2) minimum child partition node number desirable to the number of existing node and this subregion is compared, if existing node number is less than desirable minimum child partition node number, then balance increases columns on last one deck to node number closest to desirable child partition nodes;
G3) if existing node number is greater than desirable minimum child partition node number, then the columns on last one deck remains 1 constant;
G4) divide respective column in this child partition according to the columns on this child partition judging to obtain in the end one deck by the order that its place subregion specifies, complete whole divisions of this child partition.
Need in the subregion of Further Division at each, then repeat e) by stripe sequence after current child partition has divided, f), g) operation until in this subregion penultimate child partition subregion complete, then by the node division of all for this subregion remainders in last child partition of this subregion.So far the child partition in subregion divides complete.
On the basis that network partition completes, route is carried out to the information in each child partition, complete the transmission of information in network.For there is deadlock in prevention transmitting procedure, in the child partition of this network partition structure, the information routing aspects of different layers is also different.Concrete steps are as follows:
I1) on source node layer, information uses Hamilton path to transmit in each child partition, and namely carry out information transmission according to the number order of source node node layer, if child partition belongs to high child partition, then information is transmitted by label ascending order at source node layer.
I2) if child partition belongs to low child partition, then information is transmitted by label descending at source node layer.
I3) in all the other network layers, information is transmitted in each child partition by YXZ path, and namely the transmission priority of information when transmitting in Y direction is higher than the priority of X-direction, and priority in X-direction is higher than the priority of Z-direction.
The described multicast transmission method based on network on three-dimensional chip, namely according to the data package size of network structure size and transmission, dynamically select optimum partition number, and divide best child partition structure according to source node position in a network, route is carried out to the information during each height divides in the basis completing subregion.
Advantage of the present invention and effect are that its partitioned mode dynamically can select the number of subregion according to current network state, and divide each subregion according to source node position equilibrium in a network, thus the longest path length of each subregion in equalising network, reduce the total delay of multicast transmission as much as possible, then improve network transmission efficiency, optimize parallel communications performance.The method can when the packet of transmission information and network topology structure larger improve the performance reducing to postpone, be conducive to improving the communication efficiency of the network on three-dimensional chip of constantly expansion.
Other concrete effects of the present invention will go on to say below.
Accompanying drawing explanation
Fig. 1---main working process schematic diagram of the present invention.
The workflow diagram of Fig. 2---distributing labels.
The example schematic of Fig. 3---distributing labels.
Fig. 4---the workflow diagram of two zone method (TBP).
Fig. 5---the example schematic of two zone method (TBP).
Fig. 6---judge the workflow diagram of child partition number in less subregion.
Fig. 7---the workflow diagram of initial division child partition.
Fig. 8---judge the child partition in the end workflow diagram of optimum columns on one deck.
Fig. 9---division result example schematic.
Figure 10---multicast information transmission example schematic.
Embodiment
Technical thought of the present invention sums up and is:
According to the packet of transmission and the size of network structure, the subregion number of balance start delay and these two delay performance Selecting parameter optimums of network delay, and according to source node position in a network, optimum partitioned organization is divided centered by source node, route is carried out to the information during each height divides in the basis completing subregion, thus reduces the total delay of multicast transmission.Dynamic ground of the present invention is according to packet and the size of network structure and start delay is weighed in the position of source node and network delay divides optimum partition structure and carries out information transmission, multicast total delay is reduced as much as possible, along with the data package size of transmission information and the expansion of network topology structure, the performance of network in reduction delay of application the method also can improve thereupon, is conducive to the communication efficiency of the network on three-dimensional chip improving constantly expansion.
Illustrate the present invention below.
Based on the data packet length of multicast transmission method according to transmission and the size of network structure of network on three-dimensional chip, the subregion number of balance start delay and these two delay performance Selecting parameter optimums of network delay, and centered by the position of source node, divide the subregion of corresponding optimum, route is carried out to the information during each height divides in the basis completing subregion, finally reaches the object reducing multicast transmission total delay.Its main working process figure is as shown in Fig. 1 in accompanying drawing, and embodiment mode is as follows.
In order to efficiently divide network easily, prevent the transmission path between each subregion from intersecting with deadlock prevention simultaneously, first the corresponding label of all peer distribution for source node layer is needed, for ensureing that the information that source node is sent can be sent to each subregion by logical path, the method for this label distribution needs to determine according to the network site at source node place.After label distribution, two zone method (TBP) are applied on the basis of distributing labels global network is divided into two parts centered by source node: high subregion and low subregion, and found out in these two subregions compared with bigdos and less subregion by the node number compared in subregion.In network partition process, the overall situation the longest transmission path is normally by producing compared with bigdos, therefore in order to the path-length in each child partition of equilibrium and network partition size, first need to try to achieve optimum partition number n compared with bigdos according to the transmission total delay equation of network-on-chip middle ideal
1, and the minimum and maximum node number of its child partition middle ideal.Then, according to the subregion number n determined compared with the longest path of child partition in bigdos in less subregion
2, and try to achieve the minimum and maximum node number of its child partition middle ideal.So far, the preliminary work of subregion completes all, next namely starts the division two subregions being carried out to child partition.
First divide wherein any one subregion, the optimum partition number comparing this subregion judges that it is the need of division child partition, if its optimum partition number is greater than 1, this subregion needs to carry out Further Division, otherwise skips division link.In order to deadlock prevention, child partition should divide row to ensure that the transmission path between each child partition mutually disjoints by the order of specifying, and the order of different points of Division row is also different.Therefore need to judge this zoned logic type, if this subregion is high subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be ascending order, namely the node label of last column of the initial row of the first child partition is minimum in this row node label; If this subregion is low subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be descending, namely the node label of last column of the initial row of the first child partition is maximum in this row node label.After determining the stripe sequence of subregion, the stripe sequence specified by its place subregion carries out the initial division of current child partition.Initial division determines the fundamental sequence of current child partition each layer in subregion, and crucial determinant partition size is columns in subregion on last one deck, therefore, need according to the optimum partition number of child partition in each tried to achieve subregion, and desirable minimum and maximum node number, judge the optimum columns of current child partition in this subregion on last one deck, and division arranges accordingly to this child partition, completes whole divisions of current child partition.Optimum child partition number in each subregion is determined by the node number in subregion, therefore, needs to judge the size type of this subregion before this subregion of judgement divides the need of end, if this subregion is comparatively bigdos, then when n-th
1terminate child partition partiting step after-1 child partition divides, and all row remaining in this subregion are drawn to n-th
1in individual child partition, complete and divide compared with the child partition of bigdos.If this subregion is less subregion, then when n-th
2child partition partiting step is terminated after-1 child partition divides.And all row remaining in this subregion are drawn to n-th
2in individual child partition, the child partition completing less subregion divides.After a point of Division terminates, judge the whether all processed end of two subregions obtained by TBP method.If only processed a subregion, then iteron subregion partiting step has divided another subregion, otherwise terminates the subregion work of the overall situation, completes the division of network partition.On the network foundation completing subregion, route is carried out to the information in each child partition, complete the parallel transmission of multicast information in network.
In conjunction with instantiation, its specific implementation step is as follows respectively:
Now suppose that given network on three-dimensional chip topological structure (a × b × c) is 4 × 4 × 3,3 stage pipeline structure, the former coordinate of source node is (2,0,0), and the data package size transmitted in network is 15flits.The main performing step this network being carried out to subregion is as follows:
(1) find out the network layer at source node place, pressing corresponding label distribution method according to source node position is this layer of all peer distribution label, from 0 label.The workflow diagram of distributing labels is as shown in Fig. 2 in accompanying drawing.
In four boundary intersection of source node layer, determine arbitrarily that one for initial point, specifies initial XY, Z-direction according to this initial point.From initial point, choose a node (x, y, z) carry out label distribution, if the Y-axis coordinate y of source node in an initial condition
sbe not 0 or y
maxnamely source node is not positioned at the longitudinal axis border of network two dimensional surface, then judge whether this node Y-axis coordinate y carrying out label distribution is even number, if y is even number, then the label of this node is L (x, y)=a × y+x, if y is odd number, then the label of this node is L (x, y)=a × y+a-x-1.If the Y-axis coordinate y of source node in an initial condition
sbe 0 or y
max, namely source node is positioned at the longitudinal axis border of network two dimensional surface, then continue to judge the transverse axis position of source node at network two dimensional surface.
When source node is positioned at the longitudinal axis border of network two dimensional surface, if the X-axis coordinate x of source node
sfor x
max, then the label distribution mode of carrying out this node of label distribution is identical with the node label allocation scheme that source node is not positioned at the longitudinal axis border of network two dimensional surface.If the X-axis coordinate x of source node
sbe not x
max, then continue to judge whether source node is initial point.
The longitudinal axis border of network two dimensional surface is positioned at and the X-axis coordinate x of source node at source node
sbe not x
maxwhen, if source node X, Y-axis coordinate x
s, y
sbe not 0, namely source node is not initial point, then judge whether this nodes X axial coordinate x carrying out label distribution is even number, if x is even number, then the label of this node is L (x, y)=b × x+y, if x is odd number, then the label of this node is L (x, y)=b × x+b-y-1.If all node labels are assigned, then exchange the x of all nodes, y coordinate figure, namely exchanges X, Y-axis.
If the X of source node, Y-axis coordinate x
s, y
sbe all 0, namely source node is initial point, then judge whether this node Y-axis coordinate y carrying out label distribution is even number, if y is even number, then the label of this node is L (x, y)=a × y+a-x-1, if x is odd number, then the label of this node is L (x, y)=a × y+x.If all node labels are assigned, then change the X-axis coordinate figure of all nodes into x
max-x, namely turns X-direction.
Judge whether all nodes of source node layer distribute all, if be all assigned with end, then complete label distribution procedure, otherwise continue to carry out label distribution to unappropriated node.
In instances, the coordinate of source node is (2,0,0), its y
s=0, x
s≠ x
maxand x
s≠ y
s≠ 0, be therefore the node of even number for transverse axis coordinate, its label is L (x, y)=b × x+y, wherein b=4.Be the node of odd number for transverse axis coordinate, its label is L (x, y)=b × x+b-y-1, wherein b=4.Need after a node label is assigned to exchange this node x, the coordinate figure of y, after all node labels (a × b altogether, a=4, b=4) of source node layer are assigned, exchange the x of source node, y coordinate figure, the i.e. X of exchange network, Y-axis.As illustrated in figure 3 of the drawings, wherein the label of source node is 8 to the example schematic of label distribution result.
(2) on the basis of distributing labels, apply two zone method (TBP) and global network is divided into two parts centered by source node: high subregion and low subregion.And found out in these two subregions compared with bigdos and less subregion by the node number compared in subregion.The workflow diagram of two zone method (TBP) is as shown in Fig. 4 in accompanying drawing.
All nodes in network are divided, from initial point, inputs a node, if present node has label, then its label value and source node label value are compared, if its label value L (x, y) is greater than the label value L (x of source node
s, y
s), then this node is divided to high subregion, otherwise this node is divided to low subregion.
If present node does not have label, then by the z-axis coordinate z of layer and the z-axis coordinate figure z of source node residing for its its
scompare, if the z-axis coordinate z of present node is greater than the z-axis coordinate z of source node
s, then this node is divided to high subregion, otherwise this node is divided to low subregion.
Judge whether all nodes in network are divided all, if be all divided end, then complete two zone method, otherwise continue to divide unallocated node.
In instances, the source node coordinate after label distribution is (0,2,0), and label value is 8, then in network all label values be greater than 8 node and the z-axis coordinate node that is greater than 0 be all divided in high subregion.In network all label values be less than 8 node and the z-axis coordinate node that is less than 0 be all divided in low subregion.Divide by this principle until all nodes (a × b × c altogether, a=4, b=4, c=3) divide and terminate in network.As illustrated in figure 5 of the drawings, wherein high subregion is comparatively bigdos, totally 39 (N to the sample result schematic diagram of two zone method (TBP)
1) individual node.Low subregion is less subregion, totally 8 (N
2) individual node.
(3) calculate and compared with subregion number optimum in bigdos be
its value is less than maximum number of column 4, i.e. x
max+ 1, therefore do not need adjusted value again.Its optimum partition number is greater than 1, needs to divide child partition further to it.In addition, be calculated as compared with the minimum and maximum node number of child partition middle ideal each in bigdos
(4) according to the subregion number n determined compared with the longest path of child partition in bigdos in less subregion
2, and try to achieve the minimum and maximum node number of its child partition middle ideal.Particular flow sheet is as shown in Fig. 6 in accompanying drawing
First, the total path length in less subregion with compared with the longest path length in bigdos, the total node number N namely in less subregion
2with
size.If N
2be not more than compared with the longest path path length in bigdos, then less subregion does not need further subregion, i.e. its optimum child partition number n
2be 1.If N
2be greater than compared with the longest path path length in bigdos, then less subregion needs point child partition further, its optimum child partition number
then, calculated the minimum and maximum node number of each child partition middle ideal in less subregion by the optimum partition number of the less subregion obtained, be respectively
In instances, the total nodes in less subregion is 8, and its value is less than compared with the longest path length 16 (16=13+3) in bigdos, and therefore less subregion does not need further subregion, i.e. optimum child partition number n
2be 1.In less subregion, the minimum and maximum node number of each child partition middle ideal is calculated as
(5) be deadlock prevention in transmitting procedure, child partition should divide row to ensure that the transmission path between each child partition mutually disjoints by the order of specifying, and the order of different points of Division row is also different.If the subregion being about to divide is high subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be ascending order, namely the node label of last column of the initial row of the first child partition is minimum in this row node label; If the subregion being about to divide is low subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be descending, namely the node label of the first row of the initial row of the first child partition is maximum in this row node label;
In instances, the subregion of Further Division child partition is needed to only have high subregion.As shown in figure 9, it is ascending order at source node layer by be classified as unit be the stripe sequence of benchmark with node label, and therefore, the division of first high child partition is initial is classified as x=3, divides namely minimum from last column node label 12 column.After a child partition has divided, the label ascending order order of all the other child partitions in this subregion by regulation in the remaining row of source node layer has divided successively.
(6), after having specified stripe sequence, the stripe sequence specified by place subregion respectively needs the subregion of Further Division to carry out the initial division of current child partition to each.The workflow diagram of initial division child partition is as shown in Fig. 7 in accompanying drawing.
First from source node layer, on source node layer, divide row to current child partition by regulation order, then judge on the adjacent layer that is not divided in this subregion, whether be divided with the row that current layer has divided last row x-axis of row identical; If these row on adjacent layer are divided, then divide row in order successively again to this child partition at current layer, current layer often increases by row, then judge on the adjacent layer that is not once divided in this subregion, whether be divided with the row that current layer has divided last row x-axis of row identical, be not divided until judged result is these row on adjacent layer, then divide on adjacent layer in these row to current child partition, and this adjacent layer is considered as current layer.
Judge the zoned logic type of current division, if the subregion of current division is high subregion, then except source node layer, have z in this subregion
max-z
slayer, if the number of times that in child partition, current layer and adjacent layer convert is more than or equal to z
max-z
s, namely in this high subregion, all layers are divided all, then the initialization terminating current child partition in high subregion divides.Otherwise continue to divide on the adjacent layer be not divided in respective column to current child partition, until all layers are divided all in this high subregion.
If the subregion of current division is low subregion, then except source node layer, have z in this subregion
slayer, if the number of times that in child partition, current layer and adjacent layer convert is more than or equal to z
s, namely in this low subregion, all layers are divided all, then the initialization terminating current child partition in low subregion divides.Otherwise continue to divide on the adjacent layer be not divided in respective column to current child partition, until all layers are divided all in this low subregion.
In instances, the subregion of Further Division child partition is needed to only have high subregion.As shown in figure 9, first, source node layer is considered as current layer and divides from this layer, namely divide in these row to the first of x=3 high child partition by the order of label ascending order from the 1st layer of high subregion, judge whether upper these row of x=3 of adjacent layer (the 2nd layer) be not divided in high subregion are divided, if be not divided, then divide in the 1st layer of upper these row to the first of x=3 high child partition and this adjacent layer (the 2nd layer) is considered as current layer simultaneously, and whether upper these row of x=3 of the adjacent layer (the 3rd layer) continuing to judge not to be divided in this subregion are divided, if be not divided, then divide in the 3rd layer of upper these row to the first of x=3 high child partition and this adjacent layer (the 3rd layer) is considered as current layer simultaneously, now be divided on last one deck, then terminate the initial division of first high child partition.After the initial division of this child partition terminates, judge that this child partition optimum columns on last layer of high subregion as 2, then increases successively and draws in respective column to a first high child partition, complete the division of first high child partition.
The division of second high child partition is first same to be started to be considered as current layer by source node layer, the order of namely pressing label ascending order in the 1st layer of remaining columns divides in these row to the second of x=2 high child partition, judge whether be not divided by these row of adjacent layer (the 2nd layer) x=2 that this child partition divides in high subregion, if be not divided, then divide in the 2nd layer of upper these row to the second of x=2 high child partition and this adjacent layer is considered as current layer simultaneously, and continue to judge whether be not divided by these row of adjacent layer (the 3rd layer) x=2 that this child partition divides in this subregion, in this example, the 3rd layer of upper this row of x=2 are divided by first high child partition, then need increase again successively in order on the 2nd layer and draw 1 row, until the row of the 3rd layer of upper identical x-axis are not divided, namely the division scope of the 2nd layer of this child partition upper is increased to these row of x=1, the 3rd layer of this row of x=1 are continued to be divided in second high child partition, now be divided on last one deck, terminate the initial division of second high child partition.After the initial division of this child partition terminates, judge this child partition in the end on one deck optimum columns as 1, therefore do not need in the end one deck to increase to draw row again, complete the division of second high child partition.High subregion optimum partition number in the example penultimate child partition that to be 3, second high child partition be in this subregion.Therefore initialization completes to second high child partition step by step and namely terminates to divide.
(7) initial division determines the fundamental sequence of current child partition each layer in subregion, and crucial determinant partition size is columns in subregion on last one deck, therefore, need according to the optimum partition number of child partition in each tried to achieve subregion, and desirable minimum and maximum node number, judge the optimum columns of current child partition in this subregion on last one deck, and division arranges accordingly to this child partition, completes whole divisions of current child partition.Judge child partition in the end on one deck the workflow diagram of optimum columns as shown in Fig. 8 in accompanying drawing.
First, calculate the existing node number of this child partition current after initial division, comprise the nodes of the row that the last one deck of this subregion has divided; By node number C existing after initial division and the desirable minimum child partition node number β of this subregion
s(β
s=N
small-average (N1)or N
small-average (N2)) compare, if existing node number C is not more than desirable minimum child partition node number β after initial division
s, increase by a row nodes in this case, if current existing node number C is less than desirable minimum child partition node number β
sand nodes C+y current after increasing by row
max+ 1 is still less than the desirable minimum child partition node number β of this subregion
s, then increase current columns α, often increase by row and then judge once, repeat this step until result of determination is C+y
max+ 1 is not less than β
l(β
l=N
large-average (N1)or N
large-average (N2)).
If increase the nodes C+y after row
max+ 1 is not less than the desirable maximum child partition node number β of this subregion
l, then further to the absolute difference of nodes when not increasing these row and minimum child partition node number | C-β
s| with the absolute difference of existing nodes and maximum child partition node number | C+y
max-β
l| compare, if | C-β
s| be less than | C+y
max-β
l|, then optimum columns α
optimalfor current columns α is constant, otherwise optimum columns α
optimalfor α+1.
If the existing node number C after initialization divides is greater than desirable minimum child partition node number β
s, then optimum columns α
optimalfor current columns α is constant.
In instances, as shown in figure 9, the fundamental sequence of first high child partition and second high child partition is obtained by initial division, therefore needs to judge the columns on last one deck.The number of first high child partition existing node after initial division is 10.Minimum child partition node number (13) desirable to the number (10) of current existing node and this subregion is compared, in example, current existing node number is less than desirable minimum child partition node number, the present node number (14) increased after row is greater than desirable maximum child partition node number (13), and difference (1) is between the two less than the difference (3) of nodes and desirable minimum child partition node number when not increasing, therefore the row on last one deck are judged to increase by 1 row, namely on last one deck, optimum columns is 2, now the total number of this child partition node is closest to desirable child partition nodes.The number of second high child partition existing node after initial division is 14, and its value is greater than desirable minimum child partition node number 13, and therefore the optimum columns of this child partition in the end on one deck is that current columns 1 is constant.
(8) the optimum child partition number in each subregion is determined by the node number in subregion, therefore, needs to judge the size type of this subregion before this subregion of judgement divides the need of end, if this subregion is comparatively bigdos, then when n-th
1terminate child partition partiting step after-1 child partition divides, and all row remaining in this subregion are drawn to n-th
1in individual child partition, complete and divide compared with the child partition of bigdos.If this subregion is less subregion, then when n-th
2child partition partiting step is terminated after-1 child partition divides.And all row remaining in this subregion are drawn to n-th
2in individual child partition, the child partition completing less subregion divides.
In example, 2 (n after completing above partitioning step in example comparatively in bigdos (i.e. high subregion)
1-1) child partition is divided end, and therefore, remain all row be not divided in this subregion and be all classified as the 3rd (n-th
1individual) in high child partition, namely comprise the x=0 of the 0th layer, these two row of x=1, the 1st layer of x=0 row and the 2nd layer of x=0 arrange.All child partitions of first subregion so far inputted have divided, as shown in figure 9.
After (9) point Divisions terminate, judge the whether all processed end of two subregions obtained by TBP method.If only processed a subregion, then iteron subregion partiting step has divided another subregion, otherwise terminates the subregion work of the overall situation.
What first analyze in example is high subregion, then also need to analyze another subregion (low subregion).Optimum partition number due to subregion low in example is 1, and therefore this subregion does not need to be divided.So far this network completes partitioning step, as shown in figure 9.
(10) on the basis completing partition network, source node can produce the packet number corresponding with child partition number and it be delivered to respectively in each child partition, then, needs the multicast information in child partition to carry out route.When information is on source node layer, information uses Hamilton path to transmit in each child partition, and namely carry out information transmission according to the number order of source node node layer, if child partition belongs to high child partition, then information is transmitted by label ascending order at source node layer.If child partition belongs to low child partition, then information is transmitted by label descending at source node layer.
After completing the transmission of source node layer, information is transferred in the network layer of vertical direction by TSV pipe.In this kind of network layer, nodal information transmits in each child partition by YXZ path, and namely the transmission priority of information when transmitting in Y direction is higher than the priority of X-direction, and priority in X-direction is higher than the priority of Z-direction.The parallel transmission of information in network on three-dimensional chip is completed in this multicast transmission mode.
In example, the information transmission routed path in each child partition is as shown in arrow in Figure 10 in accompanying drawing.
Apply the multicast transmission total delay that above-mentioned multicast transmission method can reduce in network, improve network transmission efficiency, optimize parallel communications performance, when the packet of transmission information and network topology structure larger improve the performance reducing to postpone, be conducive to the communication efficiency of the network on three-dimensional chip improving constantly expansion.
When transmission data packet length is 5flits, 4 × 4 × 3, the network average transmission total delay of applying the method in the topological structure of 8 × 8 × 3 and 16 × 16 × 3 corresponds to 63cycles, 132cycles and 273cycles respectively.With the existing best multicast transmission method based on network on three-dimensional chip---compared with the transmission method of recursive partitioning, under these three kinds of topological structures, the network average transmission total delay of this method is less by 1.6% respectively than the network average transmission total delay of recursive partitioning transmission method, 3.8%, 4.1%.When transmission data packet length is 15flits, 4 × 4 × 3, the network average transmission total delay of applying the method in the topological structure of 8 × 8 × 3 and 16 × 16 × 3 corresponds to 110cycles, 210cycles and 426cycles respectively.Compared with the transmission method of recursive partitioning, under these three kinds of topological structures, the network average transmission total delay of this method is less by 5.2% respectively than the network average transmission total delay of recursive partitioning transmission method, and 12.9%, 18.3%.Find out thus, this method is more superior than existing best recursive partitioning transmission method in reduction transmission total delay, and along with the data package size of transmission information and the expansion of network topology structure, the performance of network in reduction delay applying this transmission method also can improve thereupon.
This multicast transmission method mainly has benefited from the design of its partition method reducing the advantage in Internet Transmission total delay, this partition method has considered the parameters of impact transmission total delay in current network, can according to current network state, from angle balance start delay and these two aspects of network delay of total delay, dynamically select the optimum partition number being applicable to current network, and mark off optimum partitioned organization according to source node location.On the basis completing subregion, according to graded properties, route is carried out to the information in each child partition, form complete multicast transmission method, realize the parallel transmission of multicast information in network.Apply the total delay that above-mentioned multicast transmission method can reduce multicast transmission as much as possible, improve network transmission efficiency, optimize parallel communications performance.
Claims (9)
1., based on a multicast transmission method for network on three-dimensional chip, it is characterized in that comprising the steps:
A) network layer at source node place is found out, in four frontier points of source node layer, determine arbitrarily that one for initial point, specify initial X, Y, Z axis direction according to this initial point, and according to source node position by corresponding label distribution method be this layer of all peer distribution label, from 0 label;
B) on the basis of distributing labels, apply two zone method and global network is divided into two parts centered by source node: high subregion and low subregion, and find out larger subregion and less subregion in these two subregions by the node number compared in subregion;
C) according to the data packet length of transmission, try to achieve the optimum child partition number compared with bigdos compared with the network size of bigdos and network flow line structure level and judge that comparatively bigdos is the need of subregion, and the optimal maximum node number of child partition in this subregion and minimum node number is obtained;
D) judge whether less subregion should be divided, and try to achieve the optimal maximum node number of child partition and minimum node number in the optimum child partition number of less subregion and this subregion according to the data packet length of transmission, the network size of less subregion and network flow line structure level;
E) in the subregion needing division subregion, find the initial row of division and the stripe sequence of source node straton subregion respectively;
F) respectively need divide subregion subregion in the current child partition of initial division;
G) judge the columns of current child partition on the last one deck of this subregion, make node number in child partition as much as possible close to ideal value, and divide and arrange to this child partition accordingly, complete whole divisions of current child partition;
H) repeat e), f), operation g) until in this subregion penultimate child partition subregion complete, by the node division of all for this subregion remainders in last child partition of this subregion.So far, divide Division complete.
I) on the basis that network partition completes, route is carried out to the information in each child partition, complete the transmission of information in network.
2. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described step is a) further comprising the steps:
A1) find out the network layer at source node place, in four frontier points of source node layer, determine arbitrarily that one for initial point, specifies initial X, Y, Z axis direction according to this initial point;
A2) if the y-axis coordinate of source node equals 0 or y
max, x-axis coordinate is not equal to x
mx, and be not initial point (0,0), all x coordinates are the node label of even number is L (x, y)=b × x+y, after label distribution, exchanges X, Y-axis;
A3) if the y-axis coordinate of source node equals 0 or y
max, x-axis coordinate is not equal to x
max, and be not initial point (0,0), all x coordinates are the node label of odd number is L (x, y)=b × x+b-y-1, after label distribution, exchanges X, Y-axis;
A4) if the y-axis coordinate of source node is not equal to 0 or y
max; Or y-axis coordinate equals 0 or y
maxand x-axis coordinate equals x
max, all y coordinates are the node label of even number is L (x, y)=a × y+x;
A5) if the y-axis coordinate of source node is not equal to 0 or y
max; Or y-axis coordinate equals 0 or y
maxand x-axis coordinate equals x
max, all y coordinates are the node label of odd number is L (x, y)=a × y+a-x-1;
A6) if source node is initial point (0,0), all y-axis coordinates are the node label of odd number is L (x, y)=a × y+x, exchanges X-direction after label distribution;
A7) if source node is initial point (0,0), all y-axis coordinates are the node label of even number is L (x, y)=a × y+a-x-1, exchanges X-direction after label distribution.
3. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described step b) further comprising the steps:
B1) label being greater than the node division that the node of source node label and all z-axis coordinates be greater than source node z-axis coordinate is high subregion;
B2) label being less than the node division that the node of source node label and all z-axis coordinates be less than source node z-axis coordinate is low subregion;
B3) nodes in higher subregion and low subregion, finds out the more subregion of nodes as comparatively bigdos, finds out the less subregion of nodes as less subregion.
4. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described step c) further comprising the steps:
C1) the comparatively total number N of bigdos interior joint is calculated
1, to desired transmission total delay formula minimizing, obtain compared with subregion number optimum in bigdos be
n
1get the most contiguous integer, n
1be greater than 1 needs subregion, otherwise comparatively bigdos does not need subregion.
C2) if the subregion number of optimum is greater than total columns, so optimum partition number n is adjusted
1for total columns x
max+ 1.
C3) be calculated as compared with the minimum and maximum node number of child partition middle ideal each in bigdos
5. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described steps d) further comprising the steps:
D1) total path length in less subregion is compared with compared with the longest path length in bigdos, the total node number N namely in less subregion
2with
size.If N
2be not more than compared with the longest path path length in bigdos, then less subregion does not need further subregion, optimum partition number n
2be 1;
D2) if N
2be greater than compared with the longest path path length in bigdos, then less subregion needs further subregion, and the optimum partition number of this subregion is calculated as
D3) in less subregion, the minimum and maximum node number of each child partition middle ideal is calculated as
.
6. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described step e) further comprising the steps:
E1) if the subregion being about to divide is high subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be ascending order, namely the node label of last column of the initial row of the first child partition is minimum in this row node label;
E2) if the subregion being about to divide is low subregion, at source node layer by being classified as unit, take node label as the stripe sequence of benchmark be descending, namely the node label of the first row of the initial row of the first child partition is maximum in this row node label.
7. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described step f) further comprising the steps:
F1) by the child partition that needs in the subregion of Further Division by appointment stripe sequence divided by column, from source node layer, divide row;
Whether the row f2) judging in this subregion, adjacent layer to have divided with current layer last row x-axis of row identical are divided;
F3) if these row are divided, then by stripe sequence standardized row again on current layer.Then f2 is returned) step judges;
F4) if these row are not divided, then divide in these these row of subregion adjacent layer to current child partition, and using adjacent layer as current layer;
F5) judging in this subregion that whether every layer is divided all, if also have layer not to be divided in this subregion, then returns f2) step continues to divide.Otherwise terminate the initial division of this child partition.
8. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described step g) further comprising the steps:
G1) calculate the number that this child partition current has node, comprise the nodes of the row that the last one deck of this subregion has divided;
G2) minimum child partition node number desirable to the number of existing node and this subregion is compared, if existing node number is less than desirable minimum child partition node number, then balance increases columns on last one deck to node number closest to desirable child partition nodes;
G3) if existing node number is greater than desirable minimum child partition node number, then the columns on last one deck remains 1 constant;
G4) divide respective column in this child partition according to the columns on this child partition judging to obtain in the end one deck by the order that its place subregion specifies, complete whole divisions of this child partition.
9. a kind of multicast transmission method based on network on three-dimensional chip according to claim 1, is characterized in that: described step I) further comprising the steps:
I1) on source node layer, information uses Hamilton path to transmit in each child partition, and namely carry out information transmission according to the number order of source node node layer, if child partition belongs to high child partition, then information is transmitted by label ascending order at source node layer.
I2) if child partition belongs to low child partition, then information is transmitted by label descending at source node layer.
I3) in all the other network layers, information is transmitted in each child partition by YXZ path, and namely the transmission priority of information when transmitting in Y direction is higher than the priority of X-direction, and priority in X-direction is higher than the priority of Z-direction.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510272323.0A CN105183693B (en) | 2015-05-26 | 2015-05-26 | A kind of multicast transmission method based on network on three-dimensional chip |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510272323.0A CN105183693B (en) | 2015-05-26 | 2015-05-26 | A kind of multicast transmission method based on network on three-dimensional chip |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105183693A true CN105183693A (en) | 2015-12-23 |
| CN105183693B CN105183693B (en) | 2019-06-14 |
Family
ID=54905783
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510272323.0A Active CN105183693B (en) | 2015-05-26 | 2015-05-26 | A kind of multicast transmission method based on network on three-dimensional chip |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105183693B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110413562A (en) * | 2019-06-26 | 2019-11-05 | 北京全路通信信号研究设计院集团有限公司 | Synchronization system and method with self-adaptive function |
| CN112039786A (en) * | 2019-06-04 | 2020-12-04 | 清华大学 | Torus network-based broadcasting method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101483614A (en) * | 2008-10-20 | 2009-07-15 | 电子科技大学 | Method for constructing network on three-dimensional chip |
| CN102065019A (en) * | 2011-01-21 | 2011-05-18 | 西安电子科技大学 | IP (Internet Protocol) core fast mapping method for network on chip based on region division |
| CN104049200A (en) * | 2014-06-23 | 2014-09-17 | 合肥工业大学 | Conflict-free test scheduling method based on link distribution in NoC |
-
2015
- 2015-05-26 CN CN201510272323.0A patent/CN105183693B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101483614A (en) * | 2008-10-20 | 2009-07-15 | 电子科技大学 | Method for constructing network on three-dimensional chip |
| CN102065019A (en) * | 2011-01-21 | 2011-05-18 | 西安电子科技大学 | IP (Internet Protocol) core fast mapping method for network on chip based on region division |
| CN104049200A (en) * | 2014-06-23 | 2014-09-17 | 合肥工业大学 | Conflict-free test scheduling method based on link distribution in NoC |
Non-Patent Citations (2)
| Title |
|---|
| MASOUMEH EBRAHIMI等: "Partitioning Methods for Unicast/Multicast Traffic in 3D NoC Architecture", 《13TH IEEE SYMPOSIUM ON DESIGN AND DIAGNOSTICS OF ELECTRONIC CIRCUITS AND SYSTEMS》 * |
| ZHOU LEI等: "3-D SPIDERGON：3-D TOPOLOGY 0F DELAY 0PTIMIZATION FOR NETWORKS-ON-CHIP", 《TRANSACTIONS OF NANJING UNIVERSITY OF AERONAUTICS＆ASTRONAUTICS》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112039786A (en) * | 2019-06-04 | 2020-12-04 | 清华大学 | Torus network-based broadcasting method |
| CN110413562A (en) * | 2019-06-26 | 2019-11-05 | 北京全路通信信号研究设计院集团有限公司 | Synchronization system and method with self-adaptive function |
| CN110413562B (en) * | 2019-06-26 | 2021-09-14 | 北京全路通信信号研究设计院集团有限公司 | Synchronization system and method with self-adaptive function |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105183693B (en) | 2019-06-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103650435B (en) | Routing traffic method of adjustment, device and controller | |
| CN102223281B (en) | Method for remapping resource demand dynamic change based on virtual network | |
| CN108260169A (en) | A kind of service function chain dynamic deployment method ensured based on QoS | |
| CN105205205B (en) | FPGA coarse grain parallelism wiring methods based on netlist location information optimal dividing | |
| CN104683488B (en) | Streaming computing system and its dispatching method and device | |
| CN106961343A (en) | A kind of virtual map method and device | |
| CN105652833A (en) | Bi-directional intelligent search-based manufacturing enterprise shop scheduling optimization method | |
| CN103763205A (en) | Three-dimensional on-chip network delay upper bound optimization method with TSV loads balanced overall | |
| US9817934B2 (en) | Multi-FPGA prototyping of an ASIC circuit | |
| CN109379281A (en) | A kind of traffic scheduling method and system based on time window | |
| CN109286528A (en) | A kind of SDN network multi-controller dispositions method based on time delay | |
| CN102420797B (en) | Topology mapping method and system | |
| CN105183693A (en) | Multicast transmission method based on three-dimensional network on chip | |
| US10157076B2 (en) | Virtual machine migration device and method | |
| CN108256182A (en) | A kind of layout method of dynamic reconfigurable FPGA | |
| CN113836846A (en) | Integrated circuit pessimism-free path analysis method for GPU accelerated computing | |
| CN110958192B (en) | Virtual data center resource allocation system and method based on virtual switch | |
| CN108848027B (en) | Minimum fault link sensing mapping method for network on chip | |
| CN109038543B (en) | State estimation calculation method based on CPU + GPU mixed isomerism | |
| CN111723545A (en) | Parallel layer distribution method based on through hole perception under super-large scale integrated circuit | |
| CN106716935A (en) | Cross-layer service configuration method and controller | |
| CN104270308A (en) | On-radio-frequency-piece network application mapping method facing unbalanced communication feature | |
| CN109710314B (en) | A method of based on graph structure distributed parallel mode construction figure | |
| CN102867095A (en) | Bus wiring method | |
| CN112615736A (en) | Delay optimal distributed NNs collaborative optimization method facing linear edge network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |