Challenge 10’s 1143913 power algorithm combination: This is not a matter!

Time:2021-3-5

Abstract:In the GECCO 2020 International Conference, the scheduling algorithm team of Huawei cloud Optimus architecture won the championship of OCP and USCP.

We started the pre research in 2012, built the hardware and software collaborative system in 2014, and applied it to Huawei cloud in 2017. We can find out how strong the Huawei cloud Optimus architecture, which took eight years to build, is.

In the just concluded GECCO 2020 International Conference, the scheduling algorithm team of Huawei cloud Optimus architecture won the championship of OCP and USCP.

Challenge 10's 1143913 power algorithm combination: This is not a matter!

The top meeting beyond imagination

GECCO conference was held in 1999, which is one of the most important events in the field of evolutionary computing. The competition attracted world-renowned research institutions and top scholars from the UK and France, such as artelys, a French optimization solution provider (roadef / Euro challenge 2018 champion in the field of industrial optimization), and roadef / Euro challenge, a British university 2016 champion), University of Grenoble in France, University College London in the UK, etc., are full of experts.

After understanding the general situation of the meeting, let’s take a look at the problems that need to be solved in the winning track of this competition.

“Evolutionary computing” in computer science refers to a series of “global optimization algorithms inspired by biological evolution”, and the sub fields of artificial intelligence to study such algorithms, which are mainly used to solve optimization problems. OCP and USCP, as classical discrete optimization problems, are proven NP hard problems. USCP is one of the 21 NP complete problems proposed by Karp, which is of great significance in the study of computational complexity theory, and is widely used in practical industrial scenarios such as edge site location, software fuzzy testing, etc.

The OCP problem can be simply described as follows: assuming that a city needs to deploy a group of cameras for full coverage, and the location (4 million optional locations), angle and coverage area of each camera are not the same, how to use the least cameras to achieve full coverage of city monitoring. The USCP problem is described in a more abstract mathematical set form, and they are essentially the same.

Challenge 10's 1143913 power algorithm combination: This is not a matter!

OCP problem diagram

Collision between cloud practice and algorithm theory

This competition provides a data set converted from the actual urban monitoring layout, in which the largest data contains more than 3.8 million monitoring candidate locations. In order to select the optimal layout scheme from 3.8 million candidate positions, the search space is as high as 2 ^ (3.8 million) ≈ [10] ^ (1143913), which is far more than the total number of all atoms in the universe. Even if the world’s computing power is used, the advantages and disadvantages of each scheme can not be verified one by one in limited time.

Challenge 10's 1143913 power algorithm combination: This is not a matter!

Huge search space, greatly enhance the difficulty of the competition

The weight based parallel local search (WPLS) algorithm submitted by Huawei cloud Optimus architecture scheduling algorithm team combines the skills of machine learning and operational research optimization, uses tabu list strategy in the process of local search, and adjusts the evaluation function to jump out of the local optimum by self-learning. In terms of implementation, the algorithm not only takes advantage of the unique hardware advantages of Huawei cloud Kunpeng and shengteng, but also finds a solution close to the theoretical optimal solution in a very short time.

For the problem of how to select the optimal layout from 3.8 million candidate locations, the core problem is to select the optimal locationHow to select a large number of edge sites, plan the capacity of each site, and achieve the optimal global business access experience through intelligent global schedulingIts essence can also be abstracted as a series of optimization problems with set covering as the core. Huawei cloud team put forward the “cloud site location” scheme. It is planned to deploy a large number of sites nationwide, calculate the limited coverage area caused by delay, quality of service, actual environment and other constraints, calculate the coverage area of each site and the corresponding construction cost, and finally put forward the optimal site deployment scheme to achieve full coverage.

It has to be said that the solution benefits from the current cloud technology, which has become an important thrust of the development of the times. With the development of the industry, the industry has put forward higher requirements for massive computing power and extreme delay experience. As the core production tool in the era of digital economy, cloud computing is gradually extending to the edge to meet the demand of surging computing power at any time, anywhere and on-demand, and realize the nearest access of business.

Full stack technology investment of Huawei cloud Optimus architecture for the future

After eight years of technology accumulation, Huawei cloud Optimus architecture provides cloud services with hard core performance, extreme stability, excellent performance and cloud edge collaboration through full stack technology investment in minimalist data center, dedicated hardware, virtualization and cloud operating system, providing consistent experience and consistent ecology for Huawei cloud, Huawei cloud stack and Huawei cloud edge.

“Smart cloud brain”, as the management and control surface of Huawei cloud Optimus architecture, is a distributed cloud operating system for the cloud, AI and 5g era, which achieves excellent global resource supply and simple use of diverse computing power. Among them, the global resource scheduling capability can support the complex scheduling collaboration between the future 100000 level distributed sites, complete the intelligent on-demand scheduling between the center and the edge, and between the edge and the edge, match the optimal node according to the business demands, and realize the nearest access. For tenants, smart cloud brain realizes intelligent recommendation of computing power through resource portrait and prediction algorithm, so that the application load runs on the most appropriate computing power.

In the future, Huawei cloud will continue to give full play to its full stack technology innovation ability, and work with partners to enable thousands of businesses to help government and enterprises realize digital transformation and intelligent upgrading.

Click follow to learn about Huawei’s new cloud technology for the first time~