CN1204494C - Load balancing design method based space-time transform used in parallel drawing system of computer cluster - Google Patents

Abstract

The present invention discloses a load balance design method based on space-time transformation in computer cluster parallel rendering systems. The present invention comprises the procedures: 1. a client side sends geometric transformation matrices and a screen area A<ik> to a rendering node Si, No. K frame rendering is started, and a load balance module is informed; 2. the rendering node S#-[i] finishes the rendering of A<ik> pixels, and the client side is informed; 3. the main program of the client side informs the load balance module, and the load balance module records rendering time t<ik>; 4. when all Si renderings are finished, images are spliced and output; 5. the load balance module carries out calculation, and a screen decomposition mode of A<ik+1> is obtained and is used for the rendering task allocation of the next frame. The load balance degree of systems can be effectively improved by applying the load balance method, an LB value is increased, the rendering speed of systems can be improved, and the present invention has advantages of low calculation cost and good practicability.

Description

In the computer cluster parallel drawing system based on the load balance method for designing of space-time transformation

Technical field

The present invention relates in a kind of computer cluster parallel drawing system load balance method for designing based on space-time transformation.

Background technology

Computer cluster parallel drawing system is based upon on many workstations or PC that connects with high performance network, realizes the parallel processing of figure by software.There are high-performance, low cost, characteristics that extensibility is good in cluster parallel drawing system, and its application comprises three aspects:

1) utilizes its low-cost constructive simulation system, new hardware architecture is studied, as Pomegranate;

2) in field of human-computer interaction, drive many projector with cluster, construct large scale, high-resolution projection wall, as Display Wall system;

3) by parallel processing, provide quick drafting, be applied to fields such as immersion VR, visualization in scientific computing, as WireGL and AnyGL to the scene that comprises mass data and complex texture to graphics tasks.

The distribution splicing was synthesized before the parallel drawing system can be divided into, middle distribution splicing is synthetic and the full images degree of depth is synthesized three types, the group system major part is preceding distribution splicing synthesis type, its characteristics are: the screen subdivision is the series of rectangular zone, each processing node is responsible in one or several zone the drafting of the geometric element of (comprise in the zone and with the zone boundary and intersecting), and the subimage of each node output is spliced into final output image.The distribution of geometric element on screen is normally uneven, therefore draw the imbalance that occurs operating load between node easily: the drafting of a large amount of intensive dough sheets of some node processing and be in busy condition, other nodes have been assigned to a small amount of sparse dough sheet and have been in the free time simultaneously.Because the speed that the frame of total system is drawn is subjected to the restriction of the slowest node, laod unbalance will make the overall performance of system descend.

The existing balancing method of loads of preceding distribution joint synthesis system is divided into static and dynamic two classes, static method is split into some fritter rectangles regularly with screen, with load heavier (being positioned at central authorities) and load light (being positioned at the edge) the rectangle collocation distribute to the drafting node.Dynamic approach is according to the ruuning situation real time altering screen subdivision of graphic package, and the median-cut method of Whelan is with the barycenter of the pel unit as computational load, and constantly subdivision is made on the long limit in each zone equals to draw the node number until number of regions; The top-down decomposition method of Whitman is made comparisons the outer bounding box of each pel and a meticulous grid, if a grid is covered by a pel, then its weight is added 1, on the pel grid, make up the Huffman tree afterwards, be 10 times of processing unit number from tree root visit downwards until the number of regions of process again, the regional dynamic assignment that these loads are equated substantially is to each processing unit then; MAHD (mesh-based adaptive hierarchicaldecomposition) uses a meticulous grid equally, but when a pel covers N grid, the weight of each grid increases 1/N, grid is organized into zone and table (summed area table) afterwards, makes dividing region again on zone and table.

These methods can be referred to as the method for " based on the geometric data analysis ", and their common characteristic are with the screen subdivision of geometric datas such as vertex position, geometric transformation as input computational load balance the best.Because considered the geological information of scene global, above-mentioned each method all has very accurate counterbalance effect, differ such as the dough sheet number of the subdivision line both sides of Whelan method and to be no more than 1.But carrying out these methods needs bigger time overhead, and the source of its calculated amount is: one, the transmission of a large amount of geometric datas or traversal; Two, operation consuming time, as the ordering in the Whelan method, Haffman tree structure, the zone of MAHD method and table structure and division etc. in the Whitman method.

A balancing method of loads will be used for reality, must satisfy following condition: the expense of itself is less than the resulting benefit of this method of use.The communication and the computing cost of above balancing method of loads itself are bigger, its practicality is greatly affected,, in any one system, is not applied though widely quote as the median-cut method of Whelan, because concerning real system, its computing cost is unacceptable.The present invention has obtained state key fundamental research development plan (2002CB312100) and has subsidized.

Summary of the invention

The purpose of this invention is to provide in a kind of computer cluster parallel drawing system load balance method for designing based on space-time transformation.

The steps include:

1) client sends geometric transformation and screen area A _IkTo drawing node S _i, start the k frame and draw, and the notification payload balance module;

2) S _iFinish A _IkThe drafting of pixel, the notice client;

3) client master routine notification payload balance module, the latter writes down drafting time t _Ik

4) all Si draw and finish, stitching image and output;

5) the load balance module is calculated, and draws screen partition patterns { A _Ik+1, be used as next frame and draw Task Distribution.

Said load balance module is to next frame screen partition patterns { A _Ik+1Calculating, specifically have vertical segmentation, level to cut apart and cut apart three kinds of methods with y-bend; The vertical segmentation method is: be provided with n and draw node, each drafting time of drawing a certain frame of node is { t ₁, t ₂, t ₃... .t _n, the vertical segmentation peak width is { w ₁, w ₂, w ₃... .w _n, order $\stackrel{\&OverBar;}{t}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{t}_i,({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_0,{\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_1,...p_n),$ Wherein $p_i=\underset{j=1}{\overset{i}{\mathrm{\&Sigma;}}}{t}_j,{\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_0=0,$ I t ∈ [p _a, p _A+1], (a ∈ [0, n-1]), w ₀'=0, h ₀'=0, the width that new screen is cut apart is: ${w_i}^{\&prime;}=(i\stackrel{\&OverBar;}{t}-p_a)\frac{w_{a+1}}{t_{a+1}}+\underset{j=1}{\overset{a}{\mathrm{\&Sigma;}}}{w}_j-\underset{k=1}{\overset{i-1}{\mathrm{\&Sigma;}}}{w_{i-1}}^{\&prime;};$ Horizontal dividing method is: be provided with n and draw node, each drafting time of drawing a certain frame of node is { t ₁, t ₂, t ₃... .t _n, horizontal cut zone height is { w ₁, w ₂, w ₃... .w _n, order $\stackrel{\&OverBar;}{t}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{t}_i,({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_0,{\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_1,...p_n),$ Wherein $p_i=\underset{j=1}{\overset{i}{\mathrm{\&Sigma;}}}{t}_j,{\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_0=0,$ I t ∈ [p _a, p _A+1], (a ∈ [0, n-1]), w ₀'=0, h ₀'=0, the height that new screen is cut apart is: ${w_i}^{\&prime;}=(i\stackrel{\&OverBar;}{t}-p_a)\frac{w_{a+1}}{t_{a+1}}+\underset{j=1}{\overset{a}{\mathrm{\&Sigma;}}}{w}_j-\underset{k=1}{\overset{i-1}{\mathrm{\&Sigma;}}}{w_{i-1}}^{\&prime;};$ The y-bend dividing method is: the basic operation of the space-time conversion method that y-bend is cut apart is the horizontal or vertical space-time conversion of cutting apart of n=2, and client has the time binary tree, and the leaf of tree is time value { t _i, the power of father node be child node power and, by the preorder traversal binary tree, top-downly alternately carry out vertical and equal the node number cutting apart of level until number of regions to the zone.

The present invention defines the index LB=T of load balance degree _First/ T _Render, T _FirstBe the time of drafting one frame of the node of at first finishing drafting, T _RenderDraw total time of a frame for total system.Fig. 3 (a)-(d) has shown some three-dimensional model continuous multiple frames vertical segmentation, the time curve of on 8PC cluster machine, drawing, (e) for Zhejiang University's new school district roaming program is drawn on 8PC cluster machine, use the time curve of y-bend subdivision, Fig. 4 is and the corresponding LB value of Fig. 3 curve that table 1 has been listed the mean value of drafting time and LB.

The average of 1 drafting time of form and LB

	Bunny	horse	hand	dragon	Campus
						Non-loaded balance frame is drawn time s/fram e	0.298	0.344	1.158	1.700	1.796
The frame of working load balance is drawn time s/frame	0.252	0.282	0.909	1.192	1.061
						Load balance calculates s/frame service time	0.017	0.023	0.024	0.028	0.046
Non-loaded balance LB	0.304	0.284	0.299	0.265	0.721
						Load balance LB	0.650	0.679	0.911	0.911	0.950

From above chart and data as can be seen:

Use the load balance degree that this balancing method of loads can improve system effectively, the LB value is raise, can improve render speed, and computing cost is very little, have good practical value.

Description of drawings

Fig. 1 is the structural representation of computer cluster drawing system;

Fig. 2 (a) (b) (c) is the ultimate principle figure of vertical segmentation space-time transformation;

Fig. 3 is the schematic diagram data that the present invention can accelerate render speed;

Fig. 4 is the schematic diagram data that the present invention can improve the load balance degree;

Fig. 5 is this method implementation system structural representation;

Fig. 6 is to use one section continuous pictures diagrammatic sketch of the sense of reality scene walkthrough of this method;

One section continuous scene diagrammatic sketch that Fig. 7 is to use the geometric model of this method to show.

Embodiment

Principle based on the balancing method of loads of space-time transformation is: with working time of rendering server tolerance as its load, the method by suitable is converted to spatial value with time value, again with the Task Distribution of spatial value control to rendering server.Owing to abandoned huge geometric data, so this method is a kind of method of " light weight " relatively, test result proves that this method is very practical and effective.

Fig. 1 has shown the structure of computer cluster drawing system.The order of client modules response application program, application programs drive many rendering servers (also claiming to draw node) pellucidly and carry out parallel drawing, and rendering server output pixel outputs to display device through splicing.Flow process based on the balancing method of loads of space-time transformation is:

1) client sends geometric transformation and screen area A _IkTo drawing node S _i, start the k frame and draw, and the notification payload balance module;

2) S _iFinish A _IkThe drafting of pixel, the notice client;

3) client master routine notification payload balance module, the latter writes down drafting time t _Ik

4) all Si draw and finish, stitching image and output;

5) the load balance module is calculated, and draws screen partition patterns { A _Ik+1, be used as next frame and draw Task Distribution.

Main points based on space-time conversion ground balancing method of loads are:

1) with the index of drafting time as the drafting node load;

2) time value is converted into spatial value;

3) carry out cutting apart and recombinating of screen area.Its ultimate principle can be reduced to shown in Figure 2, suppose initial subdivision state such as Fig. 2 (a), three drafting times of drawing node were respectively 1 second, 2 seconds, 3 seconds, think that the ratio of their loads is 1: 2: 3, the zone done cut apart as Fig. 2 (b), recombinating by the principle of task mean allocation at last obtains Fig. 2 (c).

The way that screen is cut apart has three kinds: vertical segmentation, level are cut apart with y-bend and are cut apart, and are provided with n and draw node, and each drafting time of drawing a certain frame of node is { t ₁, t ₂, t ₃... .t _n, the vertical segmentation peak width is { w ₁, w ₂, w ₃... .w _n, order $\stackrel{\&OverBar;}{t}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{t}_i,({\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_0,{\mathrm{<figure-callout\; id="1"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_1,...p_n),$ Wherein $p_i=\underset{j=1}{\overset{i}{\mathrm{\&Sigma;}}}{t}_j,{\mathrm{<figure-callout\; id="0"\; label="p"\; filenames="C0311621600091.png"\; state="\{\{state\}\}">p</figure-callout>}}_0=0,$ I t ∈ [p _a, p _A+1], (a ∈ [0, n-1]), w ₀'=0, h ₀'=0, the width that new screen is cut apart is:

${w_i}^{\&prime;}=(i\stackrel{\&OverBar;}{t}-p_a)\frac{w_{a+1}}{t_{a+1}}+\underset{j=1}{\overset{a}{\mathrm{\&Sigma;}}}{w}_j-\underset{k=1}{\overset{i-1}{\mathrm{\&Sigma;}}}{w_{i-1}}^{\&prime;}.$

It is identical with vertical segmentation that level is cut apart way.The basic operation of the space-time conversion method that y-bend is cut apart is the horizontal or vertical space-time conversion of cutting apart of n=2, and client has the time binary tree, and the leaf of tree is time value { t _i, the power of father node (time value) be child node power and, by the preorder traversal binary tree, top-downly alternately the zone is carried out equaling the node number until number of regions cutting apart of vertical and level and gets final product.

Fig. 5 implements synoptic diagram for this method, and load balance works in client, controls drawing node by computer network.Fig. 6 is one section continuous pictures of sense of reality scene walkthrough, and regional A1, A2, A3, A4 draw node by 4 respectively and generate, and table 2 has been listed the drafting time of each node.Fig. 7 is one section continuous scene that geometric model shows, the pixel of 4 rectangular areas is drawn node by 4 respectively and generated, and has also shown the drafting time (unit is second) of each this frame of node on the figure.From Fig. 6,7 and table 2 can be clear that how this method constantly make adjusting to load balance according to the drafting time.

Each node of each frame of form 2 Fig. 5 is drawn the time

	(a)	(b)	(c)	(d)	(e)	(f)
							A1 (second)	0.583	0.715	1.116	1.427	2.955	1.662
A2 (second)	0.902	0.972	0.878	1.847	2.167	1.865
							A3 (second)	0.936	0.999	1.142	1.472	2.191	1.889
A4 (second)	0.967	1.228	1.169	1.900	2.614	1.913

Claims (3)

1. In the computer cluster parallel drawing system based on the load balance method for designing of space-time transformation, it is characterized in that: the steps include:

   1) client sends geometric transformation and screen area A _IkTo drawing node S _i, start the k frame and draw, and the notification payload balance module;

   2) S _iFinish A _IkThe drafting of pixel, the notice client;

   3) client master routine notification payload balance module, the latter writes down drafting time t _Ik

   4) all Si draw and finish, stitching image and output;

   5) the load balance module is calculated, and draws screen partition patterns { A _Ik+1, be used as next frame and draw Task Distribution;

   Said load balance module is to next frame screen partition patterns { A _Ik+1Calculating, specifically have vertical segmentation, level to cut apart and cut apart three kinds of methods with y-bend; The vertical segmentation method is: be provided with n and draw node, each drafting time of drawing a certain frame of node is { t ₁, t ₂, t ₃... .t _n, the vertical segmentation peak width is { w ₁, w ₂, w ₃... .w _n, order $\stackrel{\&OverBar;}{t}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{t}_i,$ (p ₀, p ₁... p _n), wherein $p_i=\underset{i=1}{\overset{i}{\mathrm{\&Sigma;}}}{t}_j,p_0=0,$ I t ∈ [p _a, p _A+1], (a ∈ [0, n-1]), w ₀'=0, h ₀'=0, the width that new screen is cut apart is: ${w_i}^{\&prime;}=(i\stackrel{\&OverBar;}{t}-p_a)\frac{w_{a+1}}{t_{a+1}}+\underset{j=1}{\overset{a}{\mathrm{\&Sigma;}}}{w}_j-\underset{k=1}{\overset{i-1}{\mathrm{\&Sigma;}}}{w_{i-1}}^{\&prime;};$ Horizontal dividing method is: be provided with n and draw node, each drafting time of drawing a certain frame of node is { t ₁, t ₂, t ₃... .t _n, horizontal cut zone height is { w ₁, w ₂, w ₃... .w _n, order $\stackrel{\&OverBar;}{t}=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{t}_i,$ (p ₀, p ₁... p _n), wherein $p_i=\underset{i=1}{\overset{i}{\mathrm{\&Sigma;}}}{t}_j,p_0=0,$ I t ∈ [p _a, p _A+1], (a ∈ [0, n-1]), w ₀'=0, h ₀'=0, the height that new screen is cut apart is: ${w_i}^{\&prime;}=(i\stackrel{\&OverBar;}{t}-p_a)\frac{w_{a+1}}{t_{a+1}}+\underset{j=1}{\overset{a}{\mathrm{\&Sigma;}}}{w}_j-\underset{k=1}{\overset{i-1}{\mathrm{\&Sigma;}}}{w_{i-1}}^{\&prime;};$ The y-bend dividing method is: the basic operation of the space-time conversion method that y-bend is cut apart is the horizontal or vertical space-time conversion of cutting apart of n=2, and client has the time binary tree, and the leaf of tree is time value { t _i, the power of father node be child node power and, by the preorder traversal binary tree, top-downly alternately carry out vertical and equal the node number cutting apart of level until number of regions to the zone.
2. 2. in a kind of computer cluster parallel drawing according to claim 1 system based on the load balance method for designing of space-time transformation, it is characterized in that described step 1) also comprises: after the load balance module receives and starts the notice of drawing, open clock, opening entry is drawn the drafting time of node.
3. 3. based on the load balance method for designing of space-time transformation, it is characterized in that described step 3) also comprises: the load balance module stops clock, obtains to draw the drafting time of node in a kind of computer cluster parallel drawing according to claim 1 system.

Images