A novel system developed by MIT researchers automatically “learns” how to schedule data-processing operations on thousands of servers — which is traditionally reserved for obscure, human-designed algorithms. Doing so will make today’s energy-hungry data centers run more efficiently.
Data centers contain tens of thousands of servers, which constantly execute data processing tasks from developers and users. Cluster scheduling algorithms allocate incoming tasks on servers to use all available computing resources in real-time and speedup jobs.
Traditionally, however, humans have worked out some basic guidelines (“policies”) and various trading scheduling algorithms. For example, they can code an algorithm to quickly complete certain tasks, or even divide resources between jobs. But workloads — mixed workgroups — come in all sizes. Therefore, it is virtually impossible for humans to optimize their scheduling algorithms for a particular workload, and as a result, they are often limited to their actual capacity.
MIT researchers instead offloaded all of the manual codings to machines. In a paper on SIGCOMM, they describe a system that leverages the “reinforcement learning” (RL), a trial-and-error machine-learning technique for making scheduling decisions for specific workloads in specific server clusters.
To do so, they built novel RL techniques that can train on complex workloads. In training, the system tries many ways to allocate the upcoming workload on the servers, eventually finding the right conversion to use compute resources and fast processing speeds. There is no need for human intervention beyond simple instructions such as “reduce the time to finish the job.”
Compared to the best-handwritten scheduling algorithms, the researcher’s system completes jobs 20 to 30 percent faster, and twice as fast during high traffic times. However, the system learns how to efficiently workload to eliminate very little waste. The results indicate that the system can enable data centers to handle a single workload at a high speed using fewer resources.
“If you have a way of trial and error using machines, they can try different ways of scheduling jobs and automatically figure out which strategy is better than others,” says Ph.D. Hongzi Mao. The student in the Department of Electrical Engineering and Computer Science (EECS). “It will automatically improve system performance. A slight improvement in consumption, even 1 percent would save millions of dollars and more energy in data centers. “
“Not everything is good enough to make scheduling decisions,” adds Mohamed Alizadeh, an EECS professor and co-author of the research at the Computer Science and Artificial Intelligence Laboratory (CSAIL). “In existing systems, these are hard-coded parameters. You have to decide in advance. Our system instead learns to fine-tune its scheduling process features based on the data center and workload.”
Joining Mao and Alizadeh on paper: Postdocs Malte Schwarzkopf and Shailesh Bozza Venkatakrishnan, and Graduate Research Assistant Jili Meng, all CSAIL.
RL for schedule.
Typically, data processing jobs fall into data centers referred to as “nodes” and “edge” graphs. Each node represents some computational task, where the larger the node, the more computing power required. The edges connecting the nodes connect the connected tasks. Scheduling algorithms allocate nodes to servers based on different mechanisms.
Traditional RL systems are not accustomed to processing such dynamic graphs. These systems use a software “agent” that makes decisions and rewards the feedback signal. Essentially, it seeks to maximize its rewards for any action to learn ideal behavior in a particular context. For example, they help robots learn how to take objects by interacting with the environment while processing video or images through simple set of pixels.
To create their RL-based scheduler called Decima, researchers need to develop a model that can process graph-structured jobs and scale it to a large number of jobs and servers. The “agent” of their system is a scheduling algorithm that affects the graph neural network, commonly used to process graph-structured data. To come up with a graph neural network suitable for scheduling, they implemented a custom component that integrates information along the paths of the graph — quickly assessing how much computation is needed to complete a given portion of the graph. This is important for job scheduling because “child” (bottom) nodes cannot begin to execute until their “parent” (top) nodes are completed, so it is important to attenuate future work in different ways on the graph to make future scheduling decisions.
To train their RL system, the researchers simulated several different graph sequences that simulate the workload that comes into the data centers. The agent then makes decisions about how to assign each node along the graph to each server. For each decision, a component calculates the reward based on how well it has performed in a particular job — such as reducing the average time taken to process a single job. Until the agent receives as much reward as possible, it will improve its decisions.
Baselining Workload
One concern, however, is that some workload sequences are more difficult to process than others because they involve large tasks or complex structures. They always take longer to process — and, therefore, the reward signal is always shorter — than the simpler ones. This does not mean that the system performs poorly: it can make a good time on challenging workloads but is slower than lightweight workloads. That variation in difficulty makes it difficult to determine which measures are good for the model.
To address this, researchers have adopted a technique called “baselining” in this case. This technique takes averages of scenarios with a large number of variables and uses those averages as a baseline to compare future outcomes. During training, they calculated the baseline for each input sequence. Then, they allow the scheduler to train multiple times for each workload. Subsequently, the system took on average performance in all decisions taken for a single input workload. That average is the baseline by which its future decisions can be compared to determine whether its decisions are good or bad. They refer to this new method as “input-based baselining”.
Researchers say that innovation applies to many different computer systems. “This is a simple way to practice reinforcement in an environment where there is this input process that affects the environment, and you want every training event to consider a model of that input process,” he said. “Almost all computer systems deal with constantly changing environments.”
Aditya Akella, a professor of computer science at the University of Wisconsin in Madison, whose team has designed several high-performance schedulers, found that the MIT system helps them improve their own processes. “Decima can take it a step further and find opportunities for [scheduling] optimization, which are not realized through manual design/tuning processes,” Akella says. “The schedulers we designed have made significant improvements to the methods used in production in terms of application performance and cluster efficiency, but there is still a gap with the ideal improvements we can achieve. , which is very surprising Spilled. ”
Currently, their model is trained on simulations that attempt to re-create incoming online traffic in real-time. Later, researchers hope to train the model on real-time traffic, which can crash servers. So, they are currently developing a “safety net” that will turn off their system when it causes a crash. “We think of it as training cycles,” says Alizadeh. “We want this system to be constantly trained, but it also has some training cycles. If it goes too far, it can keep falling.”
We, humans, are pretty good at sugar coating, but what if we only built an algorithm or bot brand or company for the purpose of marketing?
It does the most amazing work!In the early 2000s, finding a product would be a nightmare if we searched online stores without knowing the exact name of the product.
But now when we search for an item in any e-commerce store, we get all the results of that item.
Eva can gather knowledge from thousands of sources and provide simple answers in less than 0.4 seconds.Artificial Intelligence Applications - AI in BankingAI in bankingUsing AI to prevent fraud is not a new concept.
This has saved millions of dollars.
Issues such as climate change, population growth, and food security have pushed the industry to seek more innovative approaches to improving crop yields.Early-based agricultural tech called Peat has developed an app called Start-Up Planets, which identifies potential deficiencies and nutrient deficiencies in the soil through images.
Resolving the problem of bias and variation is only about coping with over-fitting and under-fitting.
Bias is minimized, and the variance concerning the complexity of the model is increased.Read Full Article: https://bit.ly/2NsIaF4
Artificial intelligence matches human performance by learning, deriving conclusions on its own, and executing a clearly defined task.
Machine Learning is an artificial intelligence technology that uses data to automatically discover patterns and trends, resulting in deriving more accurate predictions of future events.Many organizations are leveraging ML and AI to advance their product development, sales, and overall customer experience.
Industries like manufacturing, construction, energy etc.
Most companies understand the value of a safety culture and have ideologies about what a hyper-vigilant safety culture looks like — where employees are on high alert of their surroundings, perpetually evaluating risks for the task at hand and taking precautions accordingly.
Algorithms can identify objects, edges, and velocity because of the advancements of Machine Vision.Workplace Safety Solutions through AI Computer Vision:Process and Production Safety - Majority of the workplace facilities have a CCTV system installed on the premises.
Vision-based object detection and tracking use feature such as shape, color, motion etc.
IntroductionArtificial Intelligence and data wisdom are good fields of operations, systems, and further that end at replicating mortal intelligence through machines.
AI represents action-planned feedback of perception.Data Science uses different corridors of this pattern or circle to interrupt specific problems.
Perception, data scientists attempt to identify patterns with the assistance of the info.
planning, there are two aspects Artificial Intelligence Artificial Intelligence (AI) is that the reproduction of human knowledge measures by machines, particularly PC frameworks.
Explicit utilizations of AI contain master frameworks, normal language preparing, discourse acknowledgment, and machine vision.Artificial Intelligence is critical because it can give endeavors experiences into their activities that they could not have known about already and since, now and again, AI can execute assignments better than people.
Especially with regards to dull, thorough assignments like examining enormous quantities of authoritative archives to ensure significant fields are filled in appropriately, AI apparatuses regularly complete positions rapidly and with somewhat a few blunders.Data scienceData science, analytics, and machine learning demand are growing at an astronomical rate and top companies are now trying to find professionals in Data science who can sift through the goldmine of knowledge and help them to require business decisions efficiently.Data science may be a concept wont to tackle big data and includes data cleansing, preparation, and analysis.
The process of data annotation, for speech, image, video, audio, or text, is a highly specialized task that needs expertise.
When we talk about high quality data, it means that the labels are both accurate and consistent.
By outsourcing data annotation services, companies save time and effort freeing up their in-house data scientists to focus on areas they are experienced in instead of annotating data.Here are a few of data annotation techniques used for specific AI initiatives.Audio Transcription for Speech Recognition Audio transcription is used to train speech recognition models.
High-accuracy audio transcription enables interactions to accurately happen between human speech and AI-based models like smart TVs, phones, virtual assistants, watches, computers, or other in-home or on-the-road technology.Human speech must be accurately recognized to understand not only what words are being spoken but also what they mean.
Text annotation detects and labels words depending on a predetermined set of categories required by an AI company.
Text annotation for Natural Language Processing or NLP helps machines predict and understand the human language more easily.
