More-efficient recovery from failures during large-ML-model training

Novel “checkpointing” scheme that uses CPU memory reduces the time wasted on failure recovery by more than 92%.

Today’s large machine learning models — such as generative language models or vision-language models — are so big that the process of training them is typically divided up across thousands or even tens of thousands of GPUs. Even with all that parallelism, training still frequently takes months.

With such a massive deployment of resources, hardware and software failures are common, often occurring multiple times a day. To reduce wasted work when resources fail, the large-model training procedure involves checkpointing, or regularly copying the model states to storage servers on the network. That way, if a resource fails, its most recent checkpoint can be retrieved and either reloaded or copied to a new machine, and training can continue.

Related content
Contiguous parameter management and prefetched activation offloading expand the MiCS tool kit.

Because the models are so large, checkpointing to remote storage can take a while — maybe 30 or 40 minutes. So it’s done sparingly, usually around every three hours. If a resource fails, and the training has to back up to the last checkpoint, it could mean the loss of several hours’ work. And on top of that, it can take 10 to 20 minutes just to retrieve checkpoints from storage. If failures happen several times a day, they can seriously slow down training.

In a paper my colleagues and I are presenting at this year’s Symposium on Operating System Principles (SOSP), we describe a checkpointing procedure that, instead of relying on remote storage, stores checkpoints in the CPU memory of the machines involved in model training. This makes both checkpointing and retrieval much more efficient, to the point that we can checkpoint after every training step, so that failures don’t set training back as far. In our experiments, this approach reduces the training time lost to hardware or software failures by about 92%.

In our paper, we explain how we address two major challenges to our approach: optimal checkpoint placement on machines and optimal traffic scheduling to accommodate both checkpointing and training.

GPU training

A typical GPU machine includes CPUs for general processing tasks — including allocating work to the GPUs — and eight or so GPUs, which have a special-purpose architecture optimized for massively parallel tasks such as model training. Each GPU has its own memory, but the CPU memory is much larger.

Related content
In tests, new approach is 15 to 18 times as fast as predecessors.

Training a large machine learning (ML) model — or foundation model — requires clusters of thousands of such GPU machines. Communication between machines in a cluster is much higher bandwidth than communication with remote storage servers, which is one of the reasons that CPU checkpointing is so efficient.

Optimal checkpoint placement

In our approach, which we call Gemini, each machine checkpoints to an onboard “RAM drive” — that is, a dedicated portion of its own CPU memory. This is sufficient for recovery from software failures, which typically don’t compromise the content of RAM drives. To recover from hardware failures, each machine also checkpoints to the CPU memory of at least one other machine in the cluster.

The person training the model can specify how many copies of each checkpoint should be stored on the network. Typically, that number will be two or three, but let’s call it M. Gemini divides the training cluster into groups of M machines each, and each machine checkpoints to the CPU memories of the other machines in its group.

In our paper, we prove that if the number of machines is evenly divisible by M, this checkpoint placement is optimal. If the number of machines is not evenly divisible by M, we create as many M-machine groups as possible without creating a one-machine group (which can result in one group with M + 1 machines).

Checkpointing placement.png
A sampling of checkpoint placement strategies. When the number of machines on the network is evenly divisible by the number of replicas of each checkpoint, our mixed-placement strategy reduces to the group strategy, which is provably optimal.

Gemini stores checkpoints for failure recovery in CPU memory, while storing checkpoints for other purposes, such as transfer learning and model debugging, in remote storage. This procedure is tiered, so that if the checkpoint is not in local CPU memory, Gemini attempts to retrieve it from the CPU memory of adjacent machines; if it is still unavailable, Gemini looks for it in remote storage.

Interleaved communication

During large-model training, GPUs will share model weights for computation. Checkpointing to CPU memory uses the same communication network that training traffic does. We need to make sure that the two uses don’t get in each other’s way.

Our approach includes a system profiler that learns the lengths of the idle time spans between training traffic and schedules checkpoint traffic for those time spans.

Gemini-interleaving scheme.png
A comparison of the existing communication scheme for large-model training (a), a naïve “blocking” approach to CPU checkpointing (b), and Gemini’s interleaving scheme (c).

This approach poses some difficulties, though. A GPU receiving part of a checkpoint transmission must store it locally before copying it to CPU memory, but GPU memory is limited. We allocate a small amount of each GPU’s memory to checkpointing and send checkpoints in small enough chunks that they won’t overflow those allocations.

Related content
In tests, a new way to allocate virtual machines across servers outperforms baselines by 10%.

That means, however, that before the GPU can receive the next checkpoint transmission, it needs to free up its memory allocation by copying the contents to CPU memory. If we wait for that copying to complete before sending another checkpoint transmission, we waste valuable time.

So we further subdivide each GPU memory allocation into two halves and pipeline the transfer of data to CPU memory, constantly refilling one half of the allocation while emptying the other. This optimizes our use of the precious idle time between bursts of training traffic for checkpoint traffic.

Gamini overview.jpeg
To avoid overflowing GPU memory (b), Gemini transmits checkpoints in chunks sized to a buffer of reserved GPU memory. To avoid wasted time while the contents of the buffer are copied to CPU memory (c), both the checkpoint chunks and the GPU buffers are split in half to enable pipelining (d).

To evaluate Gemini, we used it for checkpointing during the training of three popular large language models, and as baselines, we trained the same models using two prior checkpointing procedures. In our evaluation, Gemini could checkpoint model states for every iteration, and as a consequence, it reduced the training time lost because of hardware or software failures by more than 92% relative to the best-performing baseline.

Wasted-time results.png
Training time wasted due to failure recovery under three checkpointing schemes: a naïve implementation of a remote-storage scheme (blue); a remote-storage scheme optimized to maximize the use of network bandwidth (orange); and Gemini (green).

Acknowledgments: Zhen Zhang, Xinwei Fu, Yida Wang

Research areas

Related content

US, WA, Seattle
Are you a brilliant mind seeking to push the boundaries of what's possible with intelligent robotics? Join our elite team of researchers and engineers - led by Pieter Abeel, Rocky Duan, and Peter Chen - at the forefront of applied science, where we're harnessing the latest advancements in large language models (LLMs) and generative AI to reshape the world of robotics and unlock new realms of innovation. As an Applied Science Intern, you'll have the unique opportunity to work alongside world-renowned experts, gaining invaluable hands-on experience with cutting-edge robotics technologies. You'll dive deep into exciting research projects at the intersection of AI and robotics. This internship is not just about executing tasks – it's about being a driving force behind groundbreaking discoveries. You'll collaborate with cross-functional teams, leveraging your expertise in areas such as deep learning, reinforcement learning, computer vision, and motion planning to tackle real-world problems and deliver impactful solutions. Throughout your journey, you'll have access to unparalleled resources, including state-of-the-art computing infrastructure, cutting-edge research papers, and mentorship from industry luminaries. This immersive experience will not only sharpen your technical skills but also cultivate your ability to think critically, communicate effectively, and thrive in a fast-paced, innovative environment where bold ideas are celebrated. Join us at the forefront of applied robotics and AI, where your contributions will shape the future of intelligent systems and propel humanity forward. Seize this extraordinary opportunity to learn, grow, and leave an indelible mark on the world of technology. Amazon has positions available in San Francisco, CA and Seattle, WA. The ideal candidate should possess: - Strong background in machine learning, deep learning, and/or robotics - Publication record at science conferences such as NeurIPS, CVPR, ICRA, RSS, CoRL, and ICLR. - Experience in areas such as multimodal LLMs, world models, image/video tokenization, real2Sim/Sim2real transfer, bimanual manipulation, open-vocabulary panoptic scene understanding, scaling up multi-modal LLMs, and end-to-end vision-language-action models. - Proficiency in Python, Experience with PyTorch or JAX - Excellent problem-solving skills, attention to detail, and the ability to work collaboratively in a team Join us at the forefront of applied robotics and AI, and be a part of the team that's reshaping the future of intelligent systems. Apply now and embark on an extraordinary journey of discovery and innovation! Key job responsibilities - Develop novel, scalable algorithms and modeling techniques that advance the state-of-the-art in areas at the intersection of LLMs and generative AI for robotics - Tackle challenging, groundbreaking research problems on production-scale data, with a focus on robotic perception, manipulation, and control - Collaborate with cross-functional teams to solve complex business problems, leveraging your expertise in areas such as deep learning, reinforcement learning, computer vision, and motion planning - Demonstrate the ability to work independently, thrive in a fast-paced, ever-changing environment, and communicate effectively with diverse stakeholders
US, WA, Seattle
Join the next revolution in robotics at Amazon's Frontier AI & Robotics team, where you'll work alongside world-renowned AI pioneers like Pieter Abbeel, Rocky Duan, and Peter Chen to lead key initiatives in robotic intelligence. As a Senior Applied Scientist, you'll spearhead the development of breakthrough foundation models that enable robots to perceive, understand, and interact with the world in unprecedented ways. You'll drive technical excellence in areas such as perception, manipulation, scence understanding, sim2real transfer, multi-modal foundation models, and multi-task learning, designing novel algorithms that bridge the gap between cutting-edge research and real-world deployment at Amazon scale. In this role, you'll combine hands-on technical work with scientific leadership, ensuring your team delivers robust solutions for dynamic real-world environments. You'll leverage Amazon's vast computational resources to tackle ambitious problems in areas like very large multi-modal robotic foundation models and efficient, promptable model architectures that can scale across diverse robotic applications. Key job responsibilities - Lead technical initiatives in robotics foundation models, driving breakthrough approaches through hands-on research and development in areas like open-vocabulary panoptic scene understanding, scaling up multi-modal LLMs, sim2real/real2sim techniques, end-to-end vision-language-action models, efficient model inference, video tokenization - Design and implement novel deep learning architectures that push the boundaries of what robots can understand and accomplish - Guide technical direction for specific research initiatives, ensuring robust performance in production environments - Mentor fellow scientists while maintaining strong individual technical contributions - Collaborate with engineering teams to optimize and scale models for real-world applications - Influence technical decisions and implementation strategies within your area of focus A day in the life - Develop and implement novel foundation model architectures, working hands-on with our extensive compute infrastructure - Guide fellow scientists in solving complex technical challenges, from sim2real transfer to efficient multi-task learning - Lead focused technical initiatives from conception through deployment, ensuring successful integration with production systems - Drive technical discussions within your team and with key stakeholders - Conduct experiments and prototype new ideas using our massive compute cluster - Mentor team members while maintaining significant hands-on contribution to technical solutions Amazon offers a full range of benefits that support you and eligible family members, including domestic partners and their children. Benefits can vary by location, the number of regularly scheduled hours you work, length of employment, and job status such as seasonal or temporary employment. The benefits that generally apply to regular, full-time employees include: 1. Medical, Dental, and Vision Coverage 2. Maternity and Parental Leave Options 3. Paid Time Off (PTO) 4. 401(k) Plan If you are not sure that every qualification on the list above describes you exactly, we'd still love to hear from you! At Amazon, we value people with unique backgrounds, experiences, and skillsets. If you’re passionate about this role and want to make an impact on a global scale, please apply! About the team At Frontier AI & Robotics, we're not just advancing robotics – we're reimagining it from the ground up. Our team, led by pioneering AI researchers Pieter Abbeel, Rocky Duan, and Peter Chen, is building the future of intelligent robotics through groundbreaking foundation models and end-to-end learned systems. We tackle some of the most challenging problems in AI and robotics, from developing sophisticated perception systems to creating adaptive manipulation strategies that work in complex, real-world scenarios. What sets us apart is our unique combination of ambitious research vision and practical impact. We leverage Amazon's massive computational infrastructure and rich real-world datasets to train and deploy state-of-the-art foundation models. Our work spans the full spectrum of robotics intelligence – from multimodal perception using images, videos, and sensor data, to sophisticated manipulation strategies that can handle diverse real-world scenarios. We're building systems that don't just work in the lab, but scale to meet the demands of Amazon's global operations. Join us if you're excited about pushing the boundaries of what's possible in robotics, working with world-class researchers, and seeing your innovations deployed at unprecedented scale.
IL, Haifa
We’re looking for a Principal Applied Scientist in the Personalization team with experience in generative AI and large models. You will be responsible for developing and disseminating customer-facing personalized recommendation models. This is a hands-on role with global impact working with a team of world-class engineers and scientists across the wider organization. You will lead the design of machine learning models that scale to very large quantities of data, and serve high-scale low-latency recommendations to all customers worldwide. You will embody scientific rigor, designing and executing experiments to demonstrate the technical efficacy and business value of your methods. You will work alongside a science team to delight customers by aiding in recommendations relevancy, and raise the profile of Amazon as a global leader in machine learning and personalization. Successful candidates will have strong technical ability, focus on customers by applying a customer-first approach, excellent teamwork and communication skills, and a motivation to achieve results in a fast-paced environment. Our position offers exceptional opportunities for every candidate to grow their technical and non-technical skills. If you are selected, you have the opportunity to make a difference to our business by designing and building state of the art machine learning systems on big data, leveraging Amazon’s vast computing resources (AWS), working on exciting and challenging projects, and delivering meaningful results to customers world-wide. Key job responsibilities Develop machine learning algorithms for high-scale recommendations problem Rapidly design, prototype and test many possible hypotheses in a high-ambiguity environment, making use of both quantitative analysis and business judgement. Collaborate with software engineers to integrate successful experimental results into large-scale, highly complex Amazon production systems capable of handling 100,000s of transactions per second at low latency. Report results in a manner which is both statistically rigorous and compellingly relevant, exemplifying good scientific practice in a business environment.
US, WA, Seattle
The Private Brands Discovery team designs innovative machine learning solutions to drive customer awareness for Amazon’s own brands and help customers discover products they love. Private Brands Discovery is an interdisciplinary team of Scientists and Engineers, who incubate and build disruptive solutions using cutting-edge technology to solve some of the toughest science problems at Amazon. To this end, the team employs methods from Natural Language Processing, Deep learning, multi-armed bandits and reinforcement learning, Bayesian Optimization, causal and statistical inference, and econometrics to drive discovery across the customer journey. Our solutions are crucial for the success of Amazon’s own brands and serve as a beacon for discovery solutions across Amazon. This is a high visibility opportunity for someone who wants to have business impact, dive deep into large-scale problems, enable measurable actions on the consumer economy, and work closely with scientists and engineers. As a scientist, you bring business and industry context to science and technology decisions. You set the standard for scientific excellence and make decisions that affect the way we build and integrate algorithms. Your solutions are exemplary in terms of algorithm design, clarity, model structure, efficiency, and extensibility. You tackle intrinsically hard problems, acquiring expertise as needed. You decompose complex problems into straightforward solutions.. With a focus on bias for action, this individual will be able to work equally well with Science, Engineering, Economics and business teams. Key job responsibilities - 5+ yrs of relevant, broad research experience after PhD degree or equivalent. - Advanced expertise and knowledge of applying observational causal interference methods - Strong background in statistics methodology, applications to business problems, and/or big data. - Ability to work in a fast-paced business environment. - Strong research track record. - Effective verbal and written communications skills with both economists and non-economist audiences.
DE, Aachen
The Artificial General Intelligence (AGI) team is looking for a passionate, talented, and inventive Applied Scientist with a strong deep learning background, to build industry-leading Generative Artificial Intelligence (GenAI) technology with Large Language Models (LLMs) and multimodal systems. Key job responsibilities As an Applied Scientist with the AGI team, you will work with talented peers to lead the development of novel algorithms and modeling techniques, to advance the state of the art with LLMs. Your work will directly impact our customers in the form of products and services that make use of speech and language technology. You will leverage Amazon’s heterogeneous data sources and large-scale computing resources to accelerate advances in spoken language understanding. About the team The AGI team has a mission to push the envelope in GenAI with LLMs and multimodal systems, in order to provide the best-possible experience for our customers.
US, WA, Seattle
The AWS Marketplace & Partner Services Science team is hiring an Applied Scientist to develop science products that support AWS initiatives to grow AWS Partners. The team is seeking candidates with strong background in machine learning and engineering, creativity, curiosity, and great business judgment. As an applied scientist on the team, you will work on targeting and lead prioritization related AI/ML products, recommendation systems, and deliver them into the production ecosystem. You are comfortable with ambiguity and have a deep understanding of ML algorithms and an analytical mindset. You are capable of summarizing complex data and models through clear visual and written explanations. You thrive in a collaborative environment and are passionate about learning. Key job responsibilities - Work with scientists, product managers and engineers to deliver high-quality science products - Experiment with large amounts of data to deliver the best possible science solutions - Design, build, and deploy innovative ML solutions to impact AWS Co-Sell initiatives About the team The AWS Marketplace & Partner Services team is the center of Analytics, Insights, and Science supporting the AWS Specialist Partner Organization on its mission to provide customers with an outstanding experience while working with AWS partners. The Science team supports science models and recommendation systems that are deployed directly to AWS Customers, AWS partners, and internal AWS Sellers.
US, CA, Palo Alto
We’re working to improve shopping on Amazon using the conversational capabilities of large language models, and are searching for pioneers who are passionate about technology, innovation, and customer experience, and are ready to make a lasting impact on the industry. You'll be working with talented scientists and engineers to innovate on behalf of our customers. If you're fired up about being part of a dynamic, driven team, then this is your moment to join us on this exciting journey! Key job responsibilities We seek strong Applied Scientists with domain expertise in machine learning and deep learning, transformers, generative models, large language models, computer vision and multimodal models. You will devise innovative solutions at scale, pushing the technological and science boundaries. You will guide the design, modeling, and architectural choices of state-of-the-art large language models and multimodal models. You will devise and implement new algorithms and new learning strategies and paradigms. You will be technically hands-on and drive the execution from ideation to productionization. You will work in collaborative environment with other technical and business leaders, to innovate on behalf of the customer.
US, CA, Palo Alto
We’re working to improve shopping on Amazon using the conversational capabilities of large language models, and are searching for pioneers who are passionate about technology, innovation, and customer experience, and are ready to make a lasting impact on the industry. You’ll be working with talented scientists, engineers, and technical program managers (TPM) to innovate on behalf of our customers. If you’re fired up about being part of a dynamic, driven team, then this is your moment to join us on this exciting journey!
US, CA, East Palo Alto
The Customer Engagement Technology team leads AI/LLM-driven customer experience transformation using task-oriented dialogue systems. We develop multi-modal, multi-turn, goal-oriented dialog systems that can handle customer issues at Amazon scale across multiple languages. These systems are designed to adapt to changing company policies and invoke correct APIs to automate solutions to customer problems. Additionally, we enhance associate productivity through response/action recommendation, summarization to capture conversation context succinctly, retrieving precise information from documents to provide useful information to the agent, and machine translation to facilitate smoother conversations when the customer and agent speak different languages. Key job responsibilities Research and development of LLM-based chatbots and conversational AI systems for customer service applications. Design and implement state-of-the-art NLP and ML models for tasks such as language understanding, dialogue management, and response generation. Collaborate with cross-functional teams, including data scientists, software engineers, and product managers, to integrate LLM-based solutions into Amazon's customer service platforms. 4. Develop and implement strategies for data collection, annotation, and model training to ensure high-quality and robust performance of the chatbots. Conduct experiments and evaluations to measure the performance of the developed models and systems, and identify areas for improvement. Stay up-to-date with the latest advancements in NLP, LLMs, and conversational AI, and explore opportunities to incorporate new techniques and technologies into Amazon's customer service solutions. Collaborate with internal and external research communities, participate in conferences and publications, and contribute to the advancement of the field. A day in the life Amazon offers a full range of benefits that support you and eligible family members, including domestic partners and their children. Benefits can vary by location, the number of regularly scheduled hours you work, length of employment, and job status such as seasonal or temporary employment. The benefits that generally apply to regular, full-time employees include: 1. Medical, Dental, and Vision Coverage 2. Maternity and Parental Leave Options 3. Paid Time Off (PTO) 4. 401(k) Plan If you are not sure that every qualification on the list above describes you exactly, we'd still love to hear from you! At Amazon, we value people with unique backgrounds, experiences, and skillsets. If you’re passionate about this role and want to make an impact on a global scale, please apply!
US, MA, Boston
The Amazon Dash Cart team is seeking a highly motivated Research Scientist (Level 5) to join our team that is focused on building new technologies for grocery stores. We are a team of scientists invent new algorithms (especially artificial intelligence, computer vision and sensor fusion) to improve customer experiences in grocery shopping. The Amazon Dash Cart is a smart shopping cart that uses sensors to keep track of what a shopper has added. Once done, they can bypass the checkout lane and just walk out. The cart comes with convenience features like a store map, a basket that can weigh produce, and product recommendations. Amazon Dash Cart’s are available at Amazon Fresh, Whole Foods. Learn more about the Dash Cart at https://www.amazon.com/b?ie=UTF8&node=21289116011. Key job responsibilities As a research scientist, you will help solve a variety of technical challenges and mentor other engineers. You will play an active role in translating business and functional requirements into concrete deliverables and build quick prototypes or proofs of concept in partnership with other technology leaders within the team. You will tackle challenging, novel situations every day and given the size of this initiative, you’ll have the opportunity to work with multiple technical teams at Amazon in different locations. You should be comfortable with a degree of ambiguity that’s higher than most projects and relish the idea of solving problems that, frankly, haven’t been solved before - anywhere. Along the way, we guarantee that you’ll learn a ton, have fun and make a positive impact on millions of people. About the team Amazon Dash cart allows shoppers to checkout without lines — you just place the items in the cart and the cart will take care of the rest. When you’re done shopping, you leave the store through a designated dash lane. We charge the payment method in your Amazon account as you walk through the dash lane and send you a receipt. Check it out at https://www.amazon.com/b?ie=UTF8&node=21289116011. Designed and custom-built by Amazonians, our Dash cart uses a variety of technologies including computer vision, sensor fusion, and advanced machine learning.