System performance
We develop the methodology that informs the roadmap, architecture and design of all computer systems deployed in Google data centers, enabling efficient utilization of our software and hardware infrastructure.
We develop the methodology that informs the roadmap, architecture and design of all computer systems deployed in Google data centers, enabling efficient utilization of our software and hardware infrastructure.
About the team
Our team guides the roadmap, architecture and design of Google’s global computer infrastructure. We bring together experts in computer architecture, machine learning, software systems, compilers and operating systems to define and build the next generation of technology that powers Google.
Our research encompasses the entire system stack, from distributed software and runtime systems to microarchitecture and circuits. We seek to propose new computing substrates and accelerators, build and optimize large-scale real-world systems, research techniques to maximize code efficiency and define new machine-learning-based systems and paradigms. Research and open-ended exploration are key aspects of our work and we seek to share this work externally with the broader research community. We publish at a wide array of conferences, including ISCA, ASPLOS, MICRO, NeurIPS, ICML and ICLR.
Team focus summaries
The combination of the end of Moore’s law and exponential increases in demand for computing and data has created an opportunity to redefine many of the layers that power computing. We architect state-of-the-art hardware accelerators, define new microarchitectures, and drive hardware and software co-design for Google-scale workloads.
Using machine learning to improve computing systems enables us to replace many traditional heuristics within Google’s large-scale systems in the short-term, and a longer-term focus to automate the processes that we use to architect computer systems. We research, propose, and prototype ML-based techniques and then seek to deploy those techniques at scale across Google.
Google’s data centers operate on a global scale. We seek to understand how to optimize a wide range of workloads and computing resources to ensure that Google’s workloads operate at peak performance and efficiency. Research into runtime systems at Google exposes us to the scale and complexity of warehouse computing.
To optimize Google’s workloads, we must understand how they execute at the datacenter scale, which requires cutting-edge research focused on code efficiency, new profiling techniques and co-design across layers of the stack, including operating systems and compilers.
Featured publications
Warehouse-Scale Video Acceleration: Co-design and Deployment in the Wild
Danner Stodolsky
Jeff Calow
Jeremy Dorfman
Clint Smullen
Aki Kuusela
Aaron James Laursen
Alex Ramirez
Alvin Adrian Wijaya
Amir Salek
Anna Cheung
Ben Gelb
Brian Fosco
Cho Mon Kyaw
Dake He
David Alexander Munday
David Wickeraad
Devin Persaud
Don Stark
Drew Walton
Elisha Indupalli
Fong Lou
Hon Kwan Wu
In Suk Chong
Indira Jayaram
Jia Feng
JP Maaninen
Maire Mahony
Mark Steven Wachsler
Mercedes Tan
Niranjani Dasharathi
Poonacha Kongetira
Prakash Chauhan
Raghuraman Balasubramanian
Ramon Macias
Richard Ho
Rob Springer
Roy W Huffman
Sandeep Bhatia
Sarah J. Gwin
Sathish K Sekar
Srikanth Muroor
Ville-Mikko Rautio
Yolanda Ripley
Yoshiaki Hase
Yuan Li
Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Association for Computing Machinery, New York, NY, USA (2021), pp. 600-615
Preview abstract
Video sharing (e.g., YouTube, Vimeo, Facebook, TikTok) accounts for the majority of internet traffic, and video processing is also foundational to several other key workloads (video conferencing, virtual/augmented reality, cloud gaming, video in Internet-of-Things devices, etc.). The importance of these workloads motivates larger video processing infrastructures and – with the slowing of Moore’s law – specialized hardware accelerators to deliver more computing at higher efficiencies. This paper describes the design and deployment, at scale, of a new accelerator targeted at warehouse-scale video transcoding. We present our hardware design including a new accelerator building block – the video coding unit (VCU) – and discuss key design trade-offs for balanced systems at data center scale and co-designing accelerators with large-scale distributed software systems. We evaluate these accelerators “in the wild" serving live data center jobs, demonstrating 20-33x improved efficiency over our prior well-tuned non-accelerated baseline. Our design also enables effective adaptation to changing bottlenecks and improved failure management, and new workload capabilities not otherwise possible with prior systems. To the best of our knowledge, this is the first work to discuss video acceleration at scale in large warehouse-scale environments.
View details
Learning Memory Access Patterns
Jamie Alexander Smith
Heiner Litz
Christos Kozyrakis
ICML (2018)
Preview abstract
The explosion in workload complexity and the recent slow-down in Moore's law scaling call for new approaches towards efficient computing. Researchers are now beginning to use recent advances in machine learning in software optimizations, augmenting or replacing traditional heuristics and data structures. However, the space of machine learning for computer hardware architecture is only lightly explored. In this paper, we demonstrate the potential of deep learning to address the von Neumann bottleneck of memory performance. We focus on the critical problem of learning memory access patterns, with the goal of constructing accurate and efficient memory prefetchers. We relate contemporary prefetching strategies to n-gram models in natural language processing, and show how recurrent neural networks can serve as a drop-in replacement. On a suite of challenging benchmark datasets, we find that neural networks consistently demonstrate superior performance in terms of precision and recall. This work represents the first step towards practical neural-network based prefetching, and opens a wide range of exciting directions for machine learning in computer architecture research.
View details
Software-defined far memory in warehouse-scale computers
Andres Lagar-Cavilla
Suleiman Souhlal
Neha Agarwal
Radoslaw Burny
Shakeel Butt
Junaid Shahid
Greg Thelen
Kamil Adam Yurtsever
Yu Zhao
International Conference on Architectural Support for Programming Languages and Operating Systems (2019)
Preview abstract
Increasing memory demand and slowdown in technology scaling pose important challenges to total cost of ownership (TCO) of warehouse-scale computers (WSCs). One promising idea to reduce the memory TCO is to add a cheaper, but slower, "far memory" tier and use it to store infrequently accessed (or cold) data. However, introducing a far memory tier brings new challenges around dynamically responding to workload diversity and churn, minimizing stranding of capacity, and addressing brownfield (legacy) deployments.
We present a novel software-defined approach to far memory that proactively compresses cold memory pages to effectively create a far memory tier in software. Our end-to-end system design encompasses new methods to define performance service-level objectives (SLOs), a mechanism to identify cold memory pages while meeting the SLO, and our implementation in the OS kernel and node agent. Additionally, we design learning-based autotuning to periodically adapt our design to fleet-wide changes without a human in the loop. Our system has been successfully deployed across Google's WSC since 2016, serving thousands of production services. Our software-defined far memory is significantly cheaper (67% or higher memory cost reduction) at relatively good access speeds (6 us) and allows us to store a significant fraction of infrequently accessed data (on average, 20%), translating to significant TCO savings at warehouse scale.
View details
Preview abstract
Cloud applications are increasingly shifting from large monolithic services to complex graphs of loosely-coupled microservices. Despite the advantages of modularity and elasticity microservices offer, they also complicate cluster management and performance debugging, as dependencies between tiers introduce backpressure and cascading QoS violations.
We present Sage, a machine learning-driven root cause analysis system for interactive cloud microservices. Sage leverages unsupervised ML models to circumvent the overhead of trace labeling, captures the impact of dependencies between microservices to determine the root cause of unpredictable performance online, and applies corrective actions to recover a cloud service’s QoS. In experiments on both dedicated local clusters and large clusters on Google Compute Engine we show that Sage consistently achieves over 93% accuracy in correctly identifying the root cause of QoS violations, and improves performance predictability.
View details
A Hierarchical Neural Model of Data Prefetching
Zhan Shi
Akanksha Jain
Calvin Lin
Architectural Support for Programming Languages and Operating Systems (ASPLOS) (2021)
Preview abstract
This paper presents Voyager, a novel neural network for data prefetching. Unlike previous neural models for prefetching, which are limited to learning delta correlations, our model can also learn address correlations, which are important for prefetching irregular sequences of memory accesses. The key to our solution is its hierarchical structure that separates addresses into pages and offsets and that introduces a mechanism for learning important relations among pages and offsets.
Voyager provides significant prediction benefits over current data prefetchers. For a set of irregular programs from the SPEC 2006 and GAP benchmark suites, Voyager sees an average IPC improvement of 41.6% over a system with no prefetcher, compared with 21.7% and 28.2%, respectively, for idealized Domino and ISB prefetchers. We also find that for two commercial workloads for which current data prefetchers see very little benefit, Voyager dramatically improves both accuracy and coverage.
At present, slow training and prediction preclude neural models from being practically used in hardware, but Voyager’s overheads are significantly lower—in every dimension—than those of previous neural models. For example, computation cost is reduced by 15-20×, and storage overhead is reduced by 110-200×. Thus, Voyager represents a significant step towards a practical neural prefetcher.
View details
Oops I Took A Gradient: Scalable Sampling for Discrete Distributions
Christopher Joseph Maddison
David Duvenaud
Will Grathwohl
ICML (2021)
Preview abstract
We propose a general and scalable approximate sampling strategy for probabilistic models with discrete variables. Our approach uses gradients of the likelihood function with respect to its discrete inputs to propose updates in a Metropolis-Hastings sampler. We show empirically that this approach outperforms generic samplers in a number of difficult settings including Ising models, Potts models, restricted Boltzmann machines, and factorial hidden Markov models. We also demonstrate the use of our improved sampler for training deep energy-based models on high dimensional discrete data. This approach outperforms variational auto-encoders and existing energy-based models. Finally, we give bounds showing that our approach is near-optimal in the class of samplers which propose local updates.
View details
Searching for Fast Models on Datacenter Accelerators
Ruoming Pang
Andrew Li
Norm Jouppi
Conference on Computer Vision and Pattern Recognition (2021)
Preview abstract
Neural Architecture Search (NAS), together with model scaling, has shown remarkable progress in designing high accuracy and fast convolutional architecture families. However, as neither NAS nor model scaling considers sufficient hardware architecture details, they do not take full advantage of the emerging datacenter (DC) accelerators. In this paper, we search for fast and accurate CNN model families
for efficient inference on DC accelerators. We first analyze DC accelerators and find that existing CNNs suffer from insufficient operational intensity, parallelism, and execution efficiency and exhibit FLOPs-latency nonproportionality. These insights let us create a DC-accelerator-optimized search space, with space-to-depth, space-to-batch, hybrid fused convolution structures with vanilla and depthwise convolutions, and block-wise activation functions. We further propose a latency-aware compound scaling (LACS), the first multi-objective compound scaling method optimizing both accuracy and latency. Our LACS discovers that network depth should grow much faster than image size and network width, which is quite different from the observations from previous compound scaling. With the new search space
and LACS, our search and scaling on datacenter accelerators results in a new model series named EfficientNet-X. EfficientNet-X is up to more than 2X faster than EfficientNet (a model series with state-of-the-art trade-off on FLOPs and accuracy) on TPUv3 and GPUv100, with comparable accuracy. EfficientNet-X is also up to 7X faster than recent RegNet and ResNeSt on TPUv3 and GPUv100. Source code is at https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet/tpu
View details
Preview abstract
As the performance of computer systems stagnates due to the end of Moore’s Law, there is a need for new models that can understand and optimize the execution of general purpose code. While there is a growing body of work on using Graph Neural Networks (GNNs) to learn static representations of source code, these representations do not understand how code executes at runtime. In this work, we propose a new approach using GNNs to learn fused representations of general source code and its execution. Our approach defines a multi-task GNN over low-level representations of source code and program state (i.e., assembly code and dynamic memory states), converting complex source code constructs and data structures into a simpler, more uniform format. We show that this leads to improved performance over similar methods that do not use execution and it opens the door to applying GNN models to new tasks that would not be feasible from static code alone. As an illustration of this, we apply the new model to challenging dynamic tasks (branch prediction and prefetching) from the SPEC CPU benchmark suite, outperforming the state-of-the-art by 26% and 45% respectively. Moreover, we use the learned fused graph embeddings to demonstrate transfer learning with high performance on an indirectly related algorithm classification task.
View details
Thunderbolt: Throughput-Optimized, Quality-of-Service-Aware Power Capping at Scale
Shaohong Li
Sreekumar Kodakara
14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20), {USENIX} Association (2020), pp. 1241-1255
Preview abstract
As the demand for data center capacity continues to grow, hyperscale providers have used power oversubscription to increase efficiency and reduce costs. Power oversubscription requires power capping systems to smooth out the spikes that risk overloading power equipment by throttling workloads. Modern compute clusters run latency-sensitive serving and throughput-oriented batch workloads on the same servers, provisioning resources to ensure low latency for the former while using the latter to achieve high server utilization. When power capping occurs, it is desirable to maintain low latency for serving tasks and throttle the throughput of batch tasks. To achieve this, we seek a system that can gracefully throttle batch workloads and has task-level quality-of-service (QoS) differentiation.
In this paper we present Thunderbolt, a hardware-agnostic power capping system that ensures safe power oversubscription while minimizing impact on both long-running throughput-oriented tasks and latency-sensitive tasks. It uses a two-threshold, randomized unthrottling/multiplicative decrease control policy to ensure power safety with minimized performance degradation. It leverages the Linux kernel's CPU bandwidth control feature to achieve task-level QoS-aware throttling. It is robust even in the face of power telemetry unavailability. Evaluation results at the node and cluster levels demonstrate the system's responsiveness, effectiveness for reducing power, capability of QoS differentiation, and minimal impact on latency and task health. We have deployed this system at scale, in multiple production clusters. As a result, we enabled power oversubscription gains of 9%--25%, where none was previously possible.
View details
Preview abstract
A significant effort has been made to train neural networks that replicate algorithmic reasoning, but they often fail to learn the abstract concepts underlying these algorithms. This is evidenced by their inability to generalize to data distributions that are outside of their restricted training sets, namely larger inputs and unseen data. We study these generalization issues at the level of numerical subroutines that comprise common algorithms like sorting, shortest paths, and minimum spanning trees. First, we observe that transformer-based sequence-to-sequence models can learn subroutines like sorting a list of numbers, but their performance rapidly degrades as the length of lists grows beyond those found in the training set. We demonstrate that this is due to attention weights that lose fidelity with longer sequences, particularly when the input numbers are numerically similar. To address the issue, we propose a learned conditional masking mechanism, which enables the model to strongly generalize far outside of its training range with near-perfect accuracy on a variety of algorithms. Second, to generalize to unseen data, we show that encoding numbers with a binary representation leads to embeddings with rich structure once trained on downstream tasks like addition or multiplication. This allows the embedding to handle missing data by faithfully interpolating numbers not seen during training.
View details
Highlighted work
-
Reimagining video infrastructure to empower YouTubeVideo infrastructure for YouTube -
Unlocking the Full Potential of Datacenter ML Accelerators with Platform-Aware Neural ArchitectureHardware-specific ML models -
Machine Learning for Computer ArchitectureAutomating computer architecture using machine learning -
Offline Optimization for Architecting Hardware AcceleratorsLearned models for microarchitectural simulation
Some of our locations
Some of our people
-
Milad Hashemi
- Distributed Systems and Parallel Computing
- Hardware and Architecture
- Machine Intelligence
-
Marisabel Guevara Hechtman
- Hardware and Architecture
-
Sundar Dev
- Distributed Systems and Parallel Computing
- Hardware and Architecture
- Software Systems
-
Liqun Cheng
- Distributed Systems and Parallel Computing
- Hardware and Architecture
-
Rama Govindaraju
- Data Mining and Modeling
- Distributed Systems and Parallel Computing
- Algorithms and Theory
-
Parthasarathy Ranganathan
- Distributed Systems and Parallel Computing
- Hardware and Architecture
- Software Systems
-
Jichuan Chang
- Distributed Systems and Parallel Computing
- Hardware and Architecture
-
Sheng Li
- Distributed Systems and Parallel Computing
- Hardware and Architecture
- Machine Intelligence
-
Kun Lin
- Data Mining and Modeling
- Human-Computer Interaction and Visualization
- Machine Intelligence
