"Mat. Wall Street model ”or an attempt to optimize the cost of cloud IT infrastructure

MIT engineers have developed a mathematical model that can increase the performance of IaaS provider networks. It was based on some approaches used by professional investors. We will tell you more about this under the cut.

Photos - Chris Li - Unsplash

Power issue

Data centers consume almost 5% of all electricity generated on the planet. And this figure is only increasing every year. Among the reasons, experts highlight the growth in the volume of processed data and the inefficient use of network resources.

Many links between modern data centers are only partially loaded or in general are in reserve. This approach allows you to cope with sudden jumps in traffic if a break occurs somewhere in the network. But there is a drawback - as a result, part of the network equipment is idle and consumes electricity when idle.

The problem was solved at MIT and Microsoft. They developed a new mathematical model, TEAVAR, which optimizes traffic transmission in wide area networks. They applied the approaches that investors use on Wall Street ( p . 3 ).

And here is Wall Street

Engineers drew inspiration from the theory of financial risks. It helps investors maximize their returns. In economics, there is such a term as “conditional value of risk” ( VaR ) - it defines the amount that an investor risks with a given probability. For example, if the risk value with a confidence level of 99% is $ 50, then the investor will lose this money in 1% of cases. Similarly: the amount of data broadcast over the network for a fixed period of time was considered an investment, and network equipment with a different probability of failure was considered an asset. Based on this, the idea was born - to design a system that maximizes the transmission of traffic through channels with a confidence level of 99.9%.

To bring it to life, experts collected data for three years on the stability of links in Microsoft networks, and then estimated the probability of failure for each of them.

How was the model built

First, IT companies presented the network topology as a graph. The nodes were cities, and the ribs were data channels connecting them. Further, the status of the links highlighted in the diagram was checked every 15 minutes. If the signal quality fell below the set mark (for example, a large number of errors occurred or the delay exceeded the permissible limits), then the channel was considered to be out of order. Having collected all the necessary information, the engineers calculated the average operating time of each channel and estimated the probability of its failure for a fifteen-minute period. As a result, it became possible to predict the moment when one or another data line fails.

Based on the mathematical model, a load balancer was introduced that distributes traffic in the network, taking into account the "reliability" of data transmission channels. The balancer is transferred to open source - the code is on GitHub , there is also a demo on the MIT website.


According to the developers, TEAVAR effectively avoids “weak” channels and sends all the data via reliable links, increasing their effective workload. Compared with similar systems, this solution allows you to transfer three times as much data over the network. Tests were conducted on the IT infrastructure of companies such as Google, IBM and AT&T.

Photos - Victorgrigas - CC BY-SA

TEAVAR is expected to help reduce data center power consumption. Optimization of traffic will allow cloud providers to abandon part of the server and network equipment. The data center will reduce the number of cars that are idling, which will reduce the average size of a receipt for electricity.

The representative of TIBCO, a software company for organizations that works with MIT, said that customers of IaaS providers will also feel the effect of the savings. The latter will be able to direct the released funds to the development of new services and improving the quality of service.

What we write about in our blogs and social networks:

Open source: graphics card developer revealed driver documentation for Linux
How to evaluate Linux server performance: open benchmarking tools

How to protect a virtual server on the Internet
File backup: how to be safe from data loss

Backlog Grooming: Save Time on Makeover
MVP and RAT: how development approaches evolve

We at 1cloud.ru offer the Virtual Server service. We provide customers with flexible configuration management and free testing.

Source: https://habr.com/ru/post/464499/

All Articles