JPH0348361A - Simulation system - Google Patents

Abstract

PURPOSE:To reduce the latter half time of each time step processing by using (n) processors, and when the number of processors to be simulated is reduced less than n/k, and allocating k (1<k<n) processors to the simulation of one particle. CONSTITUTION:A multiprocessor system is constituted of (n) processors 21 to 2n and a common memory 3 to be read/written from these processors 21 to 2n. In the case of simulating (m) particles by means of n(n<m) processors in each time step, one processor is allocated to each particle until n particles are left to execute simulation. When the processing of one particle can be divided into (k) processing to be mutually communicated, (k) processors are allocated to the processing of one particle after n/k particles are left to execute simulation among these processors by means of the parallel processing of fine processing units. Thus, the latter half time of each time step processing can be reduced.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a simulation system for executing a particle system simulation using a multiprocessor system.

(Prior Art) The following is known as a method for realizing particle-based simulation processing such as Monte Carlo simulation in parallel processing on a multiprocessor.

■Assign a processor with no work to particles that are not running the simulation for that time step.

■When a certain processor's work is completed, the results of the simulation are received and added to the results such as energy loss (common processing). After that, if there is a particle that has not yet been simulated for that time step, that processor is newly assigned to the simulation of that particle.

The above steps (1) and (2) are performed in one simulation process at each time step until there are no more two particles. This allows parallel processing.

(Problem to be solved by the invention) However, in such a method, since one processor is in charge of simulating each particle, there are cases where the processing time for each particle in one time step is long, or the simulation of each particle in one time step is If the processing time differs depending on the particle, after all particles are assigned to the processor in the latter half of the time step, processing cannot proceed to the next time step until the last particle has been processed. As the simulation ends, the number of idle processors increases, and the time required for the simulation increases accordingly.

The purpose of the present invention is to eliminate such conventional drawbacks and increase the utilization rate of each processor by allocating work to each processor even in the latter half of each time step.
The aim is to shorten the time required for simulation.

(Means for Solving the Problems) In the present invention, as a multiprocessor that executes simulation, the processing of one process is divided into smaller processing units, and when one process is started, multiple processors are 4. A multiprocessor system that makes it possible to execute the processing of one process using a plurality of processors by allocating the processing to each processor and distributing processing to each processor for each detailed processing unit of the process. Use.

(Function) In the present invention, m particles are n(n
When executing a simulation using <m) processors, one processor is assigned to the simulation of one particle and the simulation is executed until n particles remain.

If the processing of one particle can be divided into k processes that communicate with each other, then in this system, after the number of particles remaining, k processors are assigned to process the simulation of one particle. The simulation is executed using parallel processing in small processing units between these processors.

In general, if k processors are used to take advantage of the parallelism of detailed processing, the speed improvement will be p (1<p<k) due to the overhead of synchronization between the processors.

Therefore, by using the present invention, when using n processors, the number of processors to be simulated becomes n or less, and after the point when normally there would be a processor with no work to do, But k(1<k
By assigning <n) processors to the simulation of one particle, the effect is that it is possible to reduce the number of busy processors and reduce the time in the second half of the processing of each time step.

(Example) Next, an example of the present invention will be described using FIGS. 1 to 4.

FIG. 1 is a block diagram illustrating one example of a multiprocessor used in the present invention. In FIG. 1, a multiprocessor system 1 includes n processors 21 to 2n,
These processors constitute a shared memory 3 that is capable of reading and writing.

FIG. 2 shows the flow of processing in the simulation of one particle in the present invention. In FIG. 2, the case of k=2 of the present invention is shown to simplify the explanation. In FIG. 2(a), the processing of one particle consists of independent processing a5, processing b6, and processing 7 using the results of processing a and processing b.

When performing this particle simulation process with one machine, after performing process a5, process b6 is performed in line 1,), and then,
Processing 7 using the results of processing a and processing b is repeated.

On the other hand, when processing this particle simulation using two processors, processing a4 is performed by one processor and processing b5 is performed by the other processor, as shown in FIG. 2(b).
Processing 7 using the results of processing a and processing b is executed by one of the processors (in Fig. 2 (b), the processor that executed processing a4)
Run it with In this case, since the processor executing process b5 needs to transfer the result to the other processor after the process, it is necessary to perform synchronization and transfer 8 between Broce and Nosa. Similarly, after executing process 7 using the results of processes a and b, the results must be transferred to a processor that executes process b5. There is a need. As a result, there will be idle time9 for the processor, so the processing time when processing with two machines will not be 1/2 of the time when processing with one machine, but when it is executed with one machine. 1/p(1
<p<2).

FIG. 3 shows an algorithm for allocating each processor to particle simulation during one time step simulation in the present invention.

In Figure 3, when the total number of processors is n and the processing of one particle consists of k independent processes, it is determined whether the number of particles remaining as simulation targets is greater than or equal to sulfur ( 81 in Fig. 3), if the number is higher than that, it is determined whether there is a next free processor (82 in Fig. 3).
If the particle exists, one processor executes the simulation of one particle (FIG. 3 84), and if the particle does not exist, waits and then tries S2 again.

When the number of particles remaining as simulation targets becomes n/k or less, it waits until there are k free processors (Fig. 3, 86), and at that point, k processors Run particle simulation.

(Figure 3 87).

Figures 4 (a) and (b) each simulate 5 particles using 3 processors (n = 3), with 2 independent processes.
12 schematically shows the processing of one time step when (k=2) for the conventional processing method and the method of the present invention.

In Fig. 4, in the conventional processing method, each particle #1
~7l! Regarding No. 5, a simulation is performed by one processor, and then common processing is performed. This results in processor idle time.

On the other hand, in the method of the present invention, since n/k = 1.5, in the simulation processing of particles #1 to #4, simulation 4 is executed by one processor, and the processing of particle #5 is executed by two processors. The simulation is performed.

As a result, when comparing the processing of one time step between the conventional processing method and the method of the present invention, it can be seen that the processing time of the processing method of the present invention is shorter.

(Effects of the Invention) As explained above, in the present invention, by using n processors, after the number of processors to be simulated becomes less than 1, there are processors that have no work to do. Even after the time when k (1<k<n
) Assigning processors to the simulation of one particle has the effect of reducing the number of busy processors and reducing the time in the latter half of the processing of each time step.

[Brief explanation of drawings]

Figure 1 is a diagram of the multiprocessor system used in the present invention, Figures 2 (a) and (b) are schematic diagrams showing the flow of processing in the simulation of one particle, and Figure 3 is a diagram of the 1-time process in the present invention. Figure 4 shows an algorithm that assigns each processor to particle simulation during step simulation.
FIG. 2 is a diagram schematically showing processing of one time step when (k=2) for a conventional processing method and a method according to the present invention. 1. ．． ．． Multiprocessor system, 21~2n...
Processor, 3... shared memory.

Claims (1)

[Claims] The simulation of m (m>0) particles to be simulated is performed in time increments (time steps).
In a simulation system that is executed unit by unit, the simulation results of each particle are aggregated and the simulation is executed for the next time, and the simulation process of one particle communicates with k (1<k) each other. In a simulation system that is configured based on processing, in which multiple processes operate in parallel, the processing of one process is divided into smaller processing units, and multiple processors are activated when the process is started. Using a multiprocessor system that makes it possible to execute processing using a plurality of processors by assigning the process to each of the detailed processing units of the process, In the simulation, when the number of particles targeted for the simulation is n or more,
One processor is assigned to the simulation of one particle, the simulation continues for the specific time, and the number of particles targeted for the simulation is n/
A simulation method characterized in that, after the number of particles becomes k or less, a plurality of processors are assigned to simulation of one particle.