JPH0243849A - Inter-node relay system - Google Patents

Abstract

PURPOSE:To attain high speed data transmission without causing dead lock by applying half sequencing to numbers of nodes and selecting relay nodes so that the sign of the 2nd-order differentiation coefficients of the node numbers subject to half sequencing is constant. CONSTITUTION:Half sequencing is applied to numbers of nodes 1 in a network in a multi-processor system comprising the networks and a node whose sign of the 2nd-order differentiation coefficients of the node numbers subject to half sequencing is constant is selected while keeping the half sequencing and the transmission route is decided so as to reach a destination node. Thus, the relay of nodes of hiper cube by the hardware is implemented economically without causing dead lock.

Description

[Detailed Description of the Invention] [Table of Contents] Overview Industrial Application Fields Conventional Technology and Problems to be Solved by the Invention Means for Solving the Problems Actions Examples Effects of the Invention [Summary] For example, hypercube connection network Regarding the relay method between nodes in a multiprocessor system composed of The configuration is such that the relay nodes are selected such that the signs of the second-order differential coefficients of the semi-ordered node numbers are constant.

[Industrial application field]

The present invention relates to a relay method between nodes in a multiprocessor system configured with, for example, a hypercube connection network.

Due to the technological limitations of the logic elements used to construct modern computer systems, there are limits to the machine clocks used in the computer systems' processors (CPUs), for example. It has become.

Therefore, since there is a limit to the performance of a single processor, the processing power of the entire computer system is improved by connecting multiple processors to a network and having the processors operate in parallel. As a result, communication between processors within the network has become necessary.

There are various types of networks for constructing the multiprocessor system, and one of them is a so-called hypercube connection network.

This hypercube connection network has the characteristic that it can have a large number of nodes compared to the number of links between nodes.

For example, in this network, 27 processors can be connected to 0 links of one node, so a multiprocessor system using the hypercube is often constructed.

However, simple communication between processing elements (PEs) connected to each node can cause deadlocks, so since control is performed using software, the overhead is large and high-speed transfer is difficult. Because of this problem, there has been a long-awaited need for a relay method that can execute the relay of the hypercube connection network using Hatoware without causing deadlock.

[Prior art and problems to be solved by the invention] FIG. 3 is a diagram explaining a conventional hypercube coupling network relay system, in which (a) shows an example of the configuration of a node, and (b)
) shows an example of the configuration of a two-dimensional hypercube, (c),
(d) are each three-dimensional.

An example of the configuration of a four-dimensional hypercube is shown.

Typically, each node of the hypercube connection network 1
(a) As shown in the figure, consists of a repeater (NPE) 11 with 0 links (for example, 4 in this example) and a processing element (PH) 10, and each link connects to another node 1. The processing element (PE)-10 is composed of a central processing unit (CPU), a main memory, an arithmetic unit, etc. In this case, it is called an n-dimensional hypercube. (See figures (c) and (d).) Regarding the concept of the hypercube connection network, for example, "Divided memory hypercube, E, D
, Prusokus ■, Parallel Computing 6.1988゜pages 235-245, ("The 5hared m
emoryhypercube'', E, D, BROOK
Sm, Parallel Computing6 (19
8B) 235-245) J, but as mentioned above, the number of nodes can be increased to, for example, 2'' compared to the number of links between nodes, and recent multiprocessor systems can be constructed. It is a network that is actively used in

In (a) of the figure, relay information from another node (for example, a binary destination number expressed as 2'').

(consisting of transferred data, etc.) is read by the repeater (NPE) 11, and if it is recognized that the data is for its own node, the processing NPE of that node is read.
10, but if it is recognized that it is for another node, it selects a link to the other node in a specific order and sends the relay information to the link.

If, in the two-dimensional hypercube shown in the figure (b), processing element O (hereinafter referred to as PIEO, the same shall apply hereinafter) ===> PE3. PEI # PE2
．． When data communication of PE3 = OPEO and PE2 groan PEI occurs at the same timing, node (Bl) 1 acts as a relay point for PEO = 6 PE3.
is used, node (B3) l is used as a relay point for PEI = OPE2, node (B2) 1 is used as a relay point for PE3 = OPEO, and PE2 = O
Node (BO) 1 is used as a PEI relay point.

Therefore, if you try to perform such a relay, it is necessary to send the data from one node (Bi) 1 to the next node, but because it is looped as described above, Each node (Bi) 1 repeater (NP
E O~3) Buffer at 11 (BUFQ~BtIF,
) is full and data cannot be transferred.
A so-called deadlock will occur.

Therefore, in the past, relaying had to be done using software.

That is, each node (Bi) 1 repeater (Nr'
Ei) Operating system (O3) in 11
, the note page Bi) 1 repeater (NPIEi
) All links and nodes connected to 11 (Bj)
The buffer (BUF) in the repeater (NPEj) 11
j), and selected a route that would not cause the deadlock mentioned above by collectively managing the usage status.

For this reason, there is a problem in that the overhead on the software becomes large due to the creation of data packets to be transferred, the establishment of transfer routes within the network, interrupt processing for notification of transfer completion, etc., making high-speed transfer difficult. Ta.

Therefore, it is possible to use Hatoware to implement the integrated management using this software, but in the case of a hypercube-connected network where each node has 0 links, for example, , as mentioned above, there is a problem in that it is necessary to collectively manage management information from 2'' nodes), which is impractical.

In view of the above-mentioned conventional drawbacks, the present invention provides an economical method for relaying between hypercube nodes using hardware without causing deadlock, in a relay method between nodes in a multiprocessor system configured with a hypercube connection network, for example. The purpose of this is to provide an inter-node relay system that can perform

[Means to solve the problem]

FIG. 1 is a diagram showing the principle of the inter-node relay system of the present invention.

The above problem is solved by the inter-note relay system configured as follows.

(1) In a multiprocessor system configured with a network, semi-ordering is performed for each node 10 number of the network, and the nodes 1 Select and configure to relay.

(2) The number of dimensions of the above network is n, the number of relay node 1 is expressed in binary, A = (ao + a + + - + an-1), the number of destination node 10 is expressed in binary, B = (bo, b+, -,
b,, -+), find Ci -ai■b, and if all 01, 0, the above own node 1
Processing Element (PE) connected to
10, and if all C2-〇 are not found, calculate d,・a,b, and if there is one that satisfies SS d i・1, each
Represented by DiL・(0,0,0,−・,LO,・−0),
(D;t)=a+■D;t (i=o ~(n-1)
1, find the number of the node (1) whose Manhattan distance is "1", select the node (1) that has a free buffer, and connect to that node (1). processing element (PE) (1
0), and if there is no d,・1, find ei=aitli, and if there is one that satisfies e8-1, convert each to Eil
・Represented by (0, 0, 0, −・, LO, −0), (E=,
)=ai■ EIL (i=o ~(n-1)
) and find the node (
1), select the node (1) that has a free buffer, and select the processing element (PE) (10) connected to that node (1).
In each of the above cases, if there is no free buffer, the configuration is configured to wait until a free buffer becomes available.

[Effect]

That is, according to the present invention, for example, in a relay method between nodes in a multiprocessor system configured with a hypercube connection network, arbitrary two node numbers X and Y that configure a hypercube are expressed as X = (XO, K1.X21 −+ Xn−1)Y
= (yo, yr, yz+ −・−, yn−+)
and introduce a partial order between the two nodes.

One of them is relay node A (ao, al, az
an-1), and the other one is the forwarding destination node B (b,,
b.

h2. -・,b,,-,), select a node from the relay node A where the sign of the second-order differential coefficient of the semi-ordered node number is constant without breaking the above partial ordering. , if the transfer route is determined so as to reach the transfer destination node B, high-speed transfer can be performed without causing deadlock.

Specifically, first, to detect whether or not A=B, c;
, al■b□, and when it is recognized that all of them are not 0.20, paying attention to the fact that a, ■bH = a, Ig + aHJ, first calculate dH, aHJ, and dl, 1
After finding the node number DiL, (Dot)・
a, ■DiL is calculated, and among the nodes whose Manhattan distance from the relay node A is 1, a node whose buffer is ``empty'' is selected.

If there is no node d,81 in the above node selection process, calculate eH=a;b, find the node number E1 that gives el,1, and then calculate (Etj)=a4■EiL. After calculating, the distance between the above relay node A and Manhattan is “
Among the nodes that are 1, select one whose buffer is empty.

In the route selected in this way, the numbers between the nodes are di, 1 . Or, since only one specific bit, C8,1, has changed, the partial order is maintained in the order,
This is a data transfer route that does not cause deadlock.

In the following, it will be shown that deadlock does not occur in data transfer between nodes where semi-order is maintained.

First, let the arbitrary two node numbers X and Y that constitute the hypercube be expressed as X = (xo, XI+ X2+ -..., Xn-+
)Y ””(yo, yr, Vz, ”-'+y11-
1), and if a partial order is introduced between the two nodes, then for all i, X, ≧V i +, that is, x=+yt,
When L, it is defined as X≧Y.

Similarly, for all i, Xi = yi +, that is, ×
When iyi +xiyi-1, define X=Y.

Also, X≧Y and not X=Y, that is, X for all 1, +yi=1, and xiY for some i;
When +X, y==O, define x>y.

The method for determining a forwarding destination node in the present invention is to send a data packet sent from node X via relay node Y,
When determining the transfer destination node Z, X −> y < z
Between, X>Y &Y>Z −・−・−・・・−(i)X
＞ Y & Y < Z −−−−−−−(ii
)X < Y & Y < Z ------(iii
)X < Y , & Y > Z −−−−−
---In the order shown in (iv), (i) to (iii)
) and prohibits the selection of (iv).

As mentioned above, the deadlock condition is that, for node,×,,,X,is,×,. = Used as a relay point of OX2, ×2 is
, = 6x3, where X is Xl-1-
It is used as a relay point for OXi+1, and Xn-1 is
-2-Used as a relay point for OXo, ×. is Xn-1
” When used as a relay point for XI, a deadlock occurs.

This is because, in order to use x8 as a relay point,

I have to send data to,X,. The conditions for data to be sent to node 1 are that it must be sent to the node beyond it - these conditions depend on loops.

X = Y is used as a relay point of X->Z
If we write OY < Z, the above condition becomes
As X, ≠XJ (however, i≠j), XO<XI<
XZ −−−−−−−−−−−−−−−(11X,
−> XZ-OXI −−−−−−−−−−−−f
2) Xz −Xz =6X4−−−−−−−−−−
(31L−z −Xn−+ =OXo −−1−
−−−−−−(n−2)X, l−+ → ×. -C
I
This means that (1) to (n-1) shown in (1) to (n-1) hold true.

In order to show that deadlock does not occur in the inter-node relaying method of the present invention, the above-mentioned condition (iv) must be prohibited and (
In cases where only i) to (iii) are allowed, the above (
It is only necessary to show that n relay conditions 1) to (n-1) are not satisfied simultaneously.

For that purpose, import) ~ (iii), (1) ~ (n-
It is sufficient to show that (n-1) is contradictory under the condition 2).

First, xo ≠xI, and X. From the condition that and X are adjacent, XO〈XI+ or X. ＞X.

It is.

C1) When XO<XI: From (1), X<XZ holds true. (Above, (iv
) in the same way (from the condition of (iii) excluding),
From (2), XZ<X holds true. Hereinafter, by repeatedly applying the same process up to equation (n-3), we obtain XO<XI<XZ<X3<--<X,...
-z〈Xl! -1 is obtained.

On the other hand, from the above formula (n-2) and the above Xn-2<Xn-1, Xn-1<X. It turns out that this is inconsistent with the above magnitude relationship. Therefore, equation (n-1) does not hold either.

(2) When XO> XI: (a) Xl-1> X positive 3 and X, < X,,
, (however, if ■≦i≦n-2) exists: X=”z
=OXt-+-6×, the condition (i-2) and the above X
From 1-1〉Xi, Xl-2〉Xl-t % or less, by repeatedly applying (i-3), (i-4), ---, Xo> L > Xz > X3 >...- ＞Xt
-----m---(a) Similarly, X, =81 X
,. ,=OXL. Condition 2 (i) and the above X, <
X. From 1, L-+ < L-z, hereafter, (i+1)
, (i+2), ---, (n-2), X; < X4+1 < Xi+z <--< X
n-1<X6-(b) From the above formulas (a) and (b), the condition for Xo > L + Therefore, X n-1+=: I Xo< ×,
does not hold true.

That is, condition (n-1) does not hold.

(a) When X, -, >X, (1≦i≦n-1) always holds: Xo>xi>X2>X3>->Xn-
z>Xn-1-', Xo>χ1. And Xn-1<
X.

Since this condition is the prohibition condition mentioned above, Xri-1=O
Xo −L does not hold.

That is, condition (n-1) does not hold.

Therefore, by determining the transfer destination node within the conditions of formulas (i) to (iii) above, it is possible to select a route that does not cause deadlock.

The conditions of equations (i) to (iii) above mean that the sign of the second-order differential coefficient of the number of each node forming the hypercube is constant (positive).

Therefore, in (i) to (iv) above, it is clear that (ii) can be set as a prohibition condition, and in this case, it means that the sign of the second-order differential coefficient is reversed (constant negative). do.

In this way, the present invention semi-orders the numbers of each node constituting a hypercube, and selects relay nodes such that the sign of the second-order differential coefficient of the semi-ordered node numbers is constant. This has the effect of allowing high-speed data transfer through a relay route that does not cause deadlock.

〔Example〕

Embodiments of the present invention will be described in detail below with reference to the drawings.

The above-mentioned FIG. 1 is the 31st note of the present invention! FIG. 3 is a diagram showing the principle of the method, in which (a) shows a conventional route selection method, and (b) shows a route selection method according to the present invention;
FIG. 2 is an operational flow showing an embodiment of the present invention,
As shown in Figure 1(b), communication is partially ordered for each node (B, ~B3) 1 of the hypercube network, and the sign of the second derivative of the partially ordered node number is constant. Means for selecting and relaying the node 1 such that the node 1 is selected is a means necessary for carrying out the present invention. still,
Like numbers refer to like objects throughout the design.

Hereinafter, the inter-node relay system of the present invention will be explained with reference to FIGS. 1 and 2.

First, in the conventional system shown in FIG. 1(a), as a data duplex transmission route,
,, B, -OB , =OB2. to B5-0B2)
Since there are four routes taken (to B) to Bo) to B1, as mentioned above, when data communication taking the above transfer route occurs at the same timing, each node's buffer is used (the diagonal line indicates ), resulting in a deadlock.

Therefore, in the present invention, as shown in FIG.
0-CI B , mourning B 3 , B 1 =6B o
-B 2, B s rumor B2 mourning B,, to B2) Bo
→ B1, the partially ordered node number 2
If the data transfer route is selected so that the sign of the order differential coefficient is constant (positive in this example), for example, node (B3) 1 becomes "vacant" and the occurrence of deadlock can be resolved. can.

Hereinafter, the specific data transfer route selection means will be explained with reference to the operation flow shown in FIG.

First, let the number of dimensions of the above network be n, and the number of relay node 1 is expressed in binary number, A-(ao, at, "-, an-+), and the number of destination node 1 is expressed in binary number, B = ( b,, b,, ・-", bn-+), C" (co, c++ "'-+ 1-n-+), C4-ai■b, (i=o ~ (n-1 )).

If all C8.0 are present, the data is transferred to the processing element (PE) 10 connected to the own node 1, which is the relay node. (See steps 100 and 104 in FIG. 2) If all of the C8-0s are not true, dz-a, b are determined.

When there are L items satisfying d,・1, dil・
Let di:l.-'=dik.1 (L pieces) (il=ik are all different, and k indicates the bit position of d,.1).

Therefore, set DiL=(0,0,0,-・, 1, (1-0), ↑ (Dip)=aL■ Dot (+=O~(n-1)
) to find the number of node 1 whose Manhattan distance is "1", select the node that has a free buffer among them, and select the processing element (PE) connected to node 1. ) Transfer to 10. (Refer to Steps 101, 102, and 103 in Figure 2) If the above d,・1 does not exist, find eH=a4J, and when there are L pieces of the above e,・1, e;+ -ei:
+・””′=el=1 (L pieces) (il to ik are all different, and k indicates the bit position of d,・1).

Therefore, set Ett・(0,0, sail −, 1,0−0), ↑ (E, I)=a;■EHt (i=0~(n−1)
) to find the number of node 1 whose Manhattan distance is "l", select the node that has an "free buffer" among them, and select the processing element (PE ) Transfer to 10.
(See steps 101, 105, and 106 in Figure 2.) In each of the above cases, if there is no free buffer, wait until a free buffer becomes available ('IDLE'). (Steps in Figure 2) 10
(See 7.108) When selecting such a forwarding node, the above (D; t
)=az■Dzt (i=o ~(n-1)) or (E;t)=at■E=t (i□o ~(n-1)
1 is calculated by inverting only the bit whose k-th value is "1" to obtain the transfer destination node number, so one of the partial order conditions mentioned above, al ≧ El
i=a=”Eli=1.

Therefore, deadlock can be avoided by transferring to the node selected by the above means.

Next, the operation when selecting a transfer route by the above means will be specifically explained using the source node number (01001010) and transfer destination node number (10111000) as an example.

■ Processing element (PE). +001°1o
10 to repeater (NPE) o+oo+o+o 11.

■ Operation of repeater (NPE) o+oo+o+o 11:
8. (01001010) B = (10111000) Therefore, C = A ■ eye (11110010) and D = aH
Since b+ = (01000010), dil・
di6.1 (L.2) is obtained.

Therefore, D,,=(01000000)Di6・(00
000010) (Il+)=^■D,,=(00001010)(ni
b)・A■Di6・(01001000) Similarly, E=aHb,= (10110000), so e
io=e;z=ei:+□1 (L=3) is obtained.

Therefore, EiO=(11001010)E, 2=(0
1101010) E, 3・(01011010) (Eio)=A■E,,=(11001010)(E
;z)・A■Ei2・(01101010)(Ei,
)=A■EH3=(01011010) Therefore, since C1.0 does not hold for all i, the transfer to self-relay node 1 is not performed.

Then, the Manhattan distance from the relay node 1 is °1”
Above, (Dil)=
A(E)Di, = (00001010) (D=6)・A
■Di&... (01001000) Since there are two nodes, data is transferred to the node with an empty buffer.

For example, if the Dil side is vacant, the repeater (NP
E) Transfer to o+oo+ooo 11.

■ Repeater (NPE). 1°. ,. . . Operation in 11:
Similarly, A・(01001000), B・(1
0111000).

C='(11110000), D=(01000
000).

E・(10110000) And (Dt+)・A■o, ,=(0000100
0), so it is a repeater (NPE). . . . 1°. . 1
Wait until 1 becomes free, then transfer the data.

■ Repeater (NPE). . . . 1°. . Operation in 11:
Similarly, A= (00001000), B=
(10111000)C= (10110000),
D. (00000000).

E= (10110000) Therefore, C and O do not hold for all i, but Di
=O holds true, so in this case, (Ei.)・A■
E4゜・(10001000)(Eiz)=A■Ei
Z・(00101000)(E;z)・A■Ei3・(
00011000), one vacant node is selected from among the three candidate nodes.

For example, (Ei□)・A■Ei2・(00101000
) is selected, the repeater (NPE). 01010
゜. Transfer the data to 11. Above repeater (NPE)
. . . . 1°. .

The Manhattan distance from 11 is '1'.

■ Repeater (NPE). . 1゜1゜. . Operation in 11:
In this case, A・(00101000), B=(101
11000) C=(10010000), D=(000
00000)E・(10010000) Therefore, the candidate node is (Ei.)=A■E1゜−(10101000)(E
ii) ·A■E, , = (00111000) is obtained, so one of them, for example, (Hio) ·A■E.

・Select (10101000) and connect the repeater (NPE)
+o+o+o. . Transfer to 11.

■ Repeater (NPE) + o + o + ooo 11 operations:
In this case, A・(10101000), B・(1
0111000)C=(00010000),D
= (00000000).

E= (00010000) Therefore, the candidate node is (Eii)・A■Ei3・(10111000) Therefore, the repeater (NPE) l. ,,. . . Transfer to 11.

■ Operation of repeater (NPE) + o + zoooo 11: In this case, A・(10111000), B・(1011
1000)C・(00000000), D=(0
0000000) E= (00000000) Therefore, for all i, it becomes C8.0, so its own processing element (PE) +o+++oo. 10
Transfer to.

In the present invention, node selection is performed in this way, and from the source (01001010) to the forwarding destination (10111).
000) to a processing element (PE) 10.

If the data is transferred along this route, deadlock will not occur because the second-order differential coefficient of each node number described above is constant (positive in this example).

As described above, in a multiprocessor system composed of a network of hypercube connections, the numbers of each node are semi-ordered, and the second-order differential coefficient of each number of the semi-ordered nodes is constant. (Correct,
or negative). in particular,
Find candidates whose Manhattan distance is °l'' from the selected relay node, and select the relay (NPE) of each node among them.
) is selected and transferred to that node, which is sequentially repeated until the destination node is reached.

〔Effect of the invention〕

As described above in detail, the node relay method of the present invention is, for example, a relay method between nodes in a multiprocessor system configured with a hypercube connected network, in which the number of each node of the network is semi-ordered. , and the relay nodes are selected so that the sign of the second-order differential coefficient of the semi-ordered node numbers is constant. Therefore, data can be transferred at high speed through a relay route that does not cause deadlock. There is an effect that can be done.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the principle of the inter-node relay system of the present invention. FIG. 2 is an operational flow diagram illustrating an embodiment of the present invention. FIG. 3 is a diagram illustrating a conventional hypercube connection network relay system. It is. In the drawing, are nodes (X, Y, Z ~ B3 BO ~ B3), and 10 is a processing element (PE). 11 is a repeater (NPE). ~10B indicate processing steps, respectively. (lo)'';,451'l(n/-tomon Shizukui (gahe's 1 principle) The cape of J\ is buried Σ and the second side is shown. Figure 1 (Part 1: The present invention is stagnantly shown. Akiyoshi Ekioro map! 3rd eye (I no 19 (lo) Slightly poor Heinomerkyuf wear 7th time, ・7th towa eye cliff? jte\゛5.say January 7th ] No. 3 [D ('zono 2
) the law of nature

Claims (2)

[Claims] (1) In a multiprocessor system configured with a network, the numbers of each node (1) of the network are semi-ordered, and the signs of the second-order differential coefficients of the semi-ordered node numbers are kept constant. An inter-node relay method characterized in that node (1) is selected and relayed. (2) Let the number of dimensions of the above network be n, and the number of relay node (1) is expressed in binary notation, A = (a_0, a_1, ..., a_n_-_1)
The number of the transfer destination node (1) is expressed in binary as B=(b_
0, b_1, ..., b_n_-_1), find c_i=a_i■b_i{i=0~(n-1)}, and if all c_i=0, the above Transfer to the processing element (PE) (10) connected to the own node (1), and if all c_i = 0, d_i = a_i@b_
Find i@, and if there is something that satisfies the d_i=1, each is D_
i_L=(0, 0, 0,..., 1, 0,...
・0), (D_i_L)=a_i■D_i_L{
i=0~(n-1)}, find the number of the node (1) whose Manhattan distance is '1', and select the node (1) that has an 'free buffer' among them. and transfers it to the processing element (PE) (10) connected to that node (1), and if there is no one with d_i=1 above, e_i=@a_i@
Find b_i, and if there is one that satisfies e_i=1, each of them can be expressed as E_
i_L=(0, 0, 0,..., 1, 0,...
・0), (E_i_L)=a_i■E_i_L{
i=0~(n-1)}, find the number of the node (1) whose Manhattan distance is '1', and select the node (1) that has an 'free buffer' among them. and transfers it to the processing element (PE) (10) connected to that node (1), and in each of the above cases, if there is no free buffer, it waits until a free buffer becomes available. The inter-node relay method described in item 1.