Contoh Banner

Better Information Centers Through Auto Learning

Better Information Centers Through Auto Learning
Better Information Centers Through Auto Learning
It’s no hugger-mugger that we’re obsessed amongst saving energy. For over a decade we’ve been designing in addition to edifice information centers that purpose one-half the liberate energy of a typical information center, in addition to we’re ever looking for ways to trim down our liberate energy purpose fifty-fifty further. In our pursuit of extreme efficiency, we’ve hitting upon a novel tool: machine learning. Today we’re releasing a white paper (PDF) on how we’re using neural networks to optimize information oculus operations in addition to displace our liberate energy purpose to novel lows.

It all started every bit a twenty portion project, a Google tradition of carving out fourth dimension for operate that falls exterior of one’s official task description. Jim Gao, an engineer on our information oculus team, is well-acquainted amongst the operational information nosotros assemble daily inwards the course of pedagogy of running our information centers. We calculate PUE, a stair out of liberate energy efficiency, every thirty seconds, in addition to we’re constantly tracking things similar full information technology charge (the amount of liberate energy our servers in addition to networking equipment are using at whatever time), exterior air temperature (which affects how our cooling towers work) in addition to the levels at which nosotros ready our mechanical in addition to cooling equipment. Being a smart guy—our affectionate nickname for him is “Boy Genius”—Jim realized that nosotros could last doing to a greater extent than amongst this data. He studied upward on machine learning in addition to started edifice models to predict—and improve—data oculus performance.
The mechanical found at our facility inwards The Dalles, Ore. The information oculus squad is constantly tracking the performance of the oestrus exchangers in addition to other mechanical equipment pictured here.

What Jim designed industrial plant a lot similar other examples of machine learning, similar speech communication recognition: a reckoner analyzes large amounts of information to recognize patterns in addition to “learn” from them. In a dynamic surroundings similar a information center, it tin forcefulness out last hard for humans to run across how all of the variables—IT load, exterior air temperature, etc.—interact amongst each other. One affair computers are proficient at is seeing the underlying story inwards the data, hence Jim took the information nosotros assemble inwards the course of pedagogy of our daily operations in addition to ran it through a model to manage brand feel of complex interactions that his team—being mere mortals—may non otherwise compass got noticed.
A simplified version of what the models do: compass got a bunch of data, detect the hidden interactions, in addition to then render recommendations that optimize for liberate energy efficiency.

After some lawsuit in addition to error, Jim’s models are straight off 99.6 portion accurate inwards predicting PUE. This agency he tin forcefulness out purpose the models to come upward up amongst novel ways to crush to a greater extent than efficiency out of our operations. For example, a duo months agone nosotros had to compass got some servers offline for a few days—which would ordinarily brand that information oculus less liberate energy efficient. But nosotros were able to purpose Jim’s models to alter our cooling setup temporarily—reducing the impact of the alter on our PUE for that fourth dimension period. Small tweaks similar this, on an ongoing basis, add together upward to pregnant savings inwards both liberate energy in addition to money.

The models tin forcefulness out predict PUE amongst 99.6 portion accuracy.

By pushing the boundaries of information oculus operations, Jim in addition to his squad compass got opened upward a novel dry reason of opportunities to better information oculus performance in addition to trim down liberate energy consumption. He lays out his approach inwards the white paper, hence other information oculus operators that dabble inwards machine learning (or who compass got a resident genius around who wants to figure it out) tin forcefulness out laissez passer it a endeavour every bit well.