Chapter 364: A simulation modeling tool that transcends the times!
After this sudden episode, Chang Haonan turned over the blackboard next to him and continued to explain along the same route:
“It is precisely because of the various problems faced by traditional methods. , In our program, a parallel constrained tetrahedral mesh generation method based on Delaunay-AFT will be used to generate separation layers around the split plane, thereby decomposing the mesh coverage area. The dividing plane is only used as a guide when separating layers, thereby avoiding the introduction of geometric constraints. At the same time, the efficiency of parallel grid generation is ensured through a multi-level regional decomposition strategy; the use of computer memory is saved through a communication mechanism based on disk cache... ..."
"This process is mainly divided into the following steps. The first is the assessment of the grid size..."
"The second is the introduction of the dividing plane. The dividing plane ∏ can be determined with the help of the main inertial axis with the smallest moment of inertia of the center of gravity of M. The direction vector n of this axis can be determined by the eigenvector corresponding to the smallest eigenvalue of the inertia matrix of M To determine, the calculation method is as follows..."
"..."
Only the "most basic" parallel surface triangle/quadrilateral mesh generation method mentioned by Chang Haonan, he has talked about nearly two hours of time.
This is just the theoretical part.
“During the preparation of the laboratory, I briefly verified the mesh generation method just mentioned. The test objects were common gear hob models and cylinder head models, with 348 and 348 respectively. 2736 surfaces, including a variety of mesh thin surfaces that are difficult to generate using traditional methods. I put the final results here for your reference. ”
Chang Haonan clicked the mouse and finally released the last picture of the first PPT——
Because he had so much to talk about, it was still quite early in the previous testing process. The office software often freezes, and I have to split it into several files to proceed.
"Here, I define the unit mass Q as the ratio of the radius Ri of the inscribed sphere of a unit and the radius Rc of the circumscribed sphere. In order to change the unit mass into an integer as much as possible, the result is multiplied by three , so the final definition of Q=3*(Ri/Rc)"
"It can be seen that compared with the high-precision mesh generated by the traditional method, the mesh generated by the new method even though the shape is extremely complex There are no garbage meshes with quality 0.1-0.2 in the dirty area. The total proportion of meshes with quality 0.6-1.0 is more than 91.7%. The weighted average quality is about 0.2% lower than the former. This difference is It will have almost no impact in actual work. ”
“In terms of final time consumption, although the parallel strategy adds three steps of regional decomposition, grid refinement and node merging compared to the traditional method, it is greatly improved. The grid generation efficiency is improved, so for the two test models, the total time is only 750 seconds and 2100 seconds, respectively, only 5.9% and 4.7% of the serial method..."
When Chang Haonan posted the program running log and told the results, the originally silent conference room exploded almost instantly.
The former specifically did not maintain order immediately.
After people suddenly receive information that is far beyond their cognitive level, it always takes some time to stabilize their emotions.
After all, that is 20 times the original efficiency.
In the field of numerical computing, quantitative changes in time are themselves qualitative changes.
The gear hob and the cylinder head are just two very simple models. It may not make much difference whether the mesh generation takes 20 minutes or 6 hours.
But for some very large engineering projects, there is a huge difference between half a month and a year...
What's more, there are more careful people It can be seen that the new method developed by Chang Haonan works better for more complex models.
That is to say, in practical applications, this advantage is likely to be more than 20 times.
Li Jie, who had just been sitting back and taking notes calmly, raised his head almost subconsciously after hearing the two data that Chang Haonan said.
He first raised his head slightly and looked at the curtain hanging on the wall in front of the conference room along the lower edge of the lenses of his glasses.
According to the enlarged display of the running log above, I heard it right.
That’s 20 times.
"This..."
Li Jie turned his head and looked at Song Jian, director of the Science and Technology Commission, who was sitting on his left.
Then he found that the other party was also looking towards him at this time.
As absolute giants in the industry, both of them obviously saw the potential of this software. Although what Chang Haonan has talked about so far is only about 10% of the entire project, it is enough to give a glimpse of the superiority of the final product.
Absolutely an epoch-making product!
Since my country's first supercomputer Galaxy No. 1 was put into use in 1983, the main problem that has plagued China's computer field has been the inability to increase its computing power.
It's not even our turn to discuss component performance issues.
But even if others export the same chips and memory to you, the performance of your final finished product will still be very different.
So much so that the United States has even relaxed its export controls on electronic components in recent years. Instead, it has very strict control over platform architecture.
In short...
I don’t think you can pose any threat.
The actual situation is indeed similar.
The latest Sugon 1000A supercomputer developed by the Institute of Computing for the Liaohe Oilfield has a peak operating speed of only about 6 billion operations per second. The Sugon 2000-1, which is currently under construction, has a target of only 20 billion operations.
The speed of ASCIRed, which was just put into use on the other side of the ocean last month, has reached 1 trillion times per second.
As for mass-produced commercial models like those purchased by major universities, the performance gap is even more obvious.
Such a gap is also the main reason for the delay in the development of domestic computer commercial applications.
After all, a limited number of topics can fully occupy the computing power of supercomputers. As a result, only the most important national projects can be allocated resources, and most researchers have no experience at all.
Although this situation has been alleviated to a certain extent by the supercomputing boom led by Beijing Airlines, these commercial servers with average performance often appear powerless when facing complex problems.
If you want to change people's old ideas, you have to make people feel the benefits.
At this time, Li Jie recalled in his mind the time when he met Chang Haonan a few days ago. The latter also said that if supercomputing or server businesses cannot be sold, forget it. Especially in China, where related industries are still underdeveloped, it is necessary to provide continuous services and promote supercomputer applications.
At that time, he just thought that this idea was very good and would definitely improve some user stickiness. However, the Institute of Computing Technology currently did not have the ability to do this work, so he entrusted it to Chang Haonan.
In exchange for the priority use of the future Sugon 2000 computer and the opening of the investment license to Sugon Group.
But he didn’t expect that before he even had time to convene the board of directors, the other party had already produced results.
Although there are thousands of fields that require supercomputers, numerical simulation and simulation are definitely one of the fields with the widest range of applications.
A direct 20-fold improvement in efficiency is equivalent to narrowing the computing power gap between China and the United States from three orders of magnitude to one order of magnitude.
Although it is not a direct overtake, at least you can see the opponent's back.
Moreover, almost all engineering disciplines can benefit from it.
"Comrade Chang Haonan, I have a question to ask."
A calm voice cut through the noise in the conference room.
Almost everyone turned back and looked towards the end of the room.
(End of this chapter)