Jump to Content

Large-scale optimization

Our mission is to develop large-scale optimization techniques and use them to improve the efficiency and robustness of infrastructure at Google.

About the team

We apply techniques from large-scale combinatorial optimization, online algorithms, and control theory to make Google’s computing infrastructure do more with less. We combine online and offline optimizations to achieve goals such as reducing search query latency, increasing model inference throughput and prediction quality, minimizing resource contention, maximizing the efficacy of caches, and eliminating unnecessary work in distributed systems. Our research is used in critical infrastructure that supports Search, Ads, Gemini, YouTube, and Cloud products.

Team focus summaries

Load balancing

We developed and deployed load balancing tools based on the power-of-d-choices paradigm for Google’s distributed serving systems. Our tools drive most of the YouTube serving stack and have also found success in Search Ads, logs processing, and serving ML models, where they have greatly improved tail latencies and error rates while saving resources by allowing our systems to run at higher utilizations.

Large-scale linear programming

We developed a linear programming (LP) solver that can solve instances with 100 billion non-zeros, which is about 1000x larger than the previous state-of-the-art. To achieve this scale, we rely on first-order methods where the bottleneck operations are matrix-vector multiplications, so we can avoid factoring matrices. This allows us to circumvent the memory scaling bottleneck faced by traditional LP solvers. The in-memory, multi-core version of our solver is open-sourced as PDLP, within Google's OR-Tools library.

ML model structure optimization

Optimizing ML model structures (e.g., feature selection, embedding table optimization, matrix sparsification, and neural architecture search) is a core problem with a significant impact on prediction quality, resource efficiency, and ultimately revenue. These are often NP-hard combinatorial optimization problems, so we developed efficient differentiable search techniques (e.g., Sequential Attention) to tackle these problems at scale, while offering provable guarantees.

Search infrastructure optimization

Several of our projects have partnered with product teams to help improve the efficiency of Google's search infrastructure. In the backend for web search, we introduced a distributed feedback control loop to govern the way queries are fanned out to worker machines, in order to balance load better by taking advantage of real-time load information on the workers. We also improved the efficacy of caching by increasing the homogeneity of the stream of queries seen by any single worker machine. To accomplish this, we clustered query terms, used these clusters to define voting weights, and assigned queries to workers on the basis of votes cast by the terms in the query. In a pub-sub system that powers some of our display ads, we saved computation by clustering some components of subscriptions in a way that lets us take advantage of bitwise-parallel CPU operations.

Distributed optimization based on core-sets

Composable core-sets provide an effective method for solving optimization problems on massive datasets. The main idea is to partition data among some number of machines, and use each machine to compute some small summary/sketch of the data. After gathering all summaries on one machine, we solve the original optimization problem on the combined sketch.

Large-scale set cover

Maximum coverage and minimum set-cover problems are among the fundamental problems in combinatorial optimization and have a variety of applications in machine learning and data mining. We study these problems in the streaming and MapReduce models of computation, and develop practical algorithms with tight theoretical bounds for runtime, space and approximation guarantee. Our main idea is a novel sketching technique that compresses the input into a small space, independent of the size of the ground set, without hurting the quality by much.

Read more

Consistent hashing

We design memoryless balanced allocation algorithms to assign a dynamic set of tasks to a dynamic set of servers such that the load (the number of assigned tasks) on each server is bounded, and the allocation does not change by much for every update operation.

Read more

Featured publications

Load is not what you should balance: Introducing PReQuaL
Bartek Wydrowski
Bobby Kleinberg
Steve Rumble
(2024)
Preview abstract We present Prequal (\emph{Probing to Reduce Queuing and Latency}), a load balancer for distributed multi-tenant systems. Prequal aims to minimize real-time request latency in the presence of heterogeneous server capacities and non-uniform, time-varying antagonist load. It actively probes server load to leverage the \emph{power of $d$ choices} paradigm, extending it with asynchronous and reusable probes. Cutting against received wisdom, Prequal does not balance CPU load, but instead selects servers according to estimated latency and active requests-in-flight (RIF). We explore its major design features on a testbed system and evaluate it on YouTube, where it has been deployed for more than two years. Prequal has dramatically decreased tail latency, error rates, and resource use, enabling YouTube and other production systems at Google to run at much higher utilization. View details
Sequential Attention for Feature Selection
Taisuke Yasuda
Proceedings of the 11th International Conference on Learning Representations (2023)
Preview abstract Feature selection is the problem of selecting a subset of features for a machine learning model that maximizes model quality subject to a budget constraint. For neural networks, prior methods, including those based on L1 regularization, attention, and other techniques, typically select the entire feature subset in one evaluation round, ignoring the residual value of features during selection, i.e., the marginal contribution of a feature given that other features have already been selected. We propose a feature selection algorithm called Sequential Attention that achieves state-of-the-art empirical results for neural networks. This algorithm is based on an efficient one-pass implementation of greedy forward selection and uses attention weights at each step as a proxy for feature importance. We give theoretical insights into our algorithm for linear regression by showing that an adaptation to this setting is equivalent to the classical Orthogonal Matching Pursuit (OMP) algorithm, and thus inherits all of its provable guarantees. Our theoretical and empirical analyses offer new explanations towards the effectiveness of attention and its connections to overparameterization, which may be of independent interest. View details
Practical Large-Scale Linear Programming using Primal-Dual Hybrid Gradient
Mateo Díaz
Oliver Hinder
Haihao Lu
Miles Lubin
Brendan O'Donoghue
Warren Schudy
NeurIPS 2021
Preview abstract We present PDLP, a practical first-order method for linear programming (LP) that can solve to the high levels of accuracy that are expected in traditional LP applications. In addition, it can scale to very large problems because its core operation is matrix-vector multiplications. PDLP is derived by applying the primal-dual hybrid gradient (PDHG) method, popularized by Chambolle and Pock (2011), to a saddle-point formulation of LP. PDLP enhances PDHG for LP by combining several new techniques with older tricks from the literature; the enhancements include diagonal preconditioning, presolving, adaptive step sizes, and adaptive restarting. PDLP improves the state of the art for first-order methods applied to LP. We compare PDLP with SCS, an ADMM-based solver, on a set of 383 LP instances derived from MIPLIB 2017. With a target of 10^(-8) relative accuracy and 1 hour time limit, PDLP achieves a 6.3x reduction in the geometric mean of solve times and a 4.6x reduction in the number of instances unsolved (from 227 to 49). Furthermore, we highlight standard benchmark instances and a large-scale application (PageRank) where our open-source prototype of PDLP, written in Julia, outperforms a commercial LP solver. View details
Edge-Weighted Online Bipartite Matching
Runzhou Tao
Zhiyi Huang
Journal of the ACM, vol. 69 (2022), 45:1-45:35
Preview abstract Online bipartite matching is one of the most fundamental problems in the online algorithms literature. Karp, Vazirani, and Vazirani (STOC 1990) introduced an elegant algorithm for the unweighted problem that achieves an optimal competitive ratio of 1 - 1/e. Aggarwal et al. (SODA 2011) later generalized their algorithm and analysis to the vertex-weighted case. Little is known, however, about the most general edge-weighted problem aside from the trivial 1/2-competitive greedy algorithm. In this paper, we present the first online algorithm that breaks the long standing 1/2 barrier and achieves a competitive ratio of at least 0.5086. In light of the hardness result of Kapralov, Post, and Vondrák (SODA 2013) that restricts beating a 1/2 competitive ratio for the more general problem of monotone submodular welfare maximization, our result can be seen as strong evidence that edge-weighted bipartite matching is strictly easier than submodular welfare maximization in the online setting. The main ingredient in our online matching algorithm is a novel subroutine called online correlated selection (OCS), which takes a sequence of pairs of vertices as input and selects one vertex from each pair. Instead of using a fresh random bit to choose a vertex from each pair, the OCS negatively correlates decisions across different pairs and provides a quantitative measure on the level of correlation. We believe our OCS technique is of independent interest and will find further applications in other online optimization problems. View details
Cache-aware load balancing of data center applications
Aaron Schild
Ray Yang
Richard Zhuang
Proceedings of the VLDB Endowment, vol. 12 (2019), pp. 709-723
Preview abstract Our deployment of cache-aware load balancing in the Google web search backend reduced cache misses by ~0.5x, contributing to a double-digit percentage increase in the throughput of our serving clusters by relieving a bottleneck. This innovation has benefited all production workloads since 2015, serving billions of queries daily. A load balancer forwards each query to one of several identical serving replicas. The replica pulls each term's postings list into RAM from flash, either locally or over the network. Flash bandwidth is a critical bottleneck, motivating an application-directed RAM cache on each replica. Sending the same term reliably to the same replica would increase the chance it hits cache, and avoid polluting the other replicas' caches. However, most queries contain multiple terms and we have to send the whole query to one replica, so it is not possible to achieve a perfect partitioning of terms to replicas. We solve this via a voting scheme, whereby the load balancer conducts a weighted vote by the terms in each query, and sends the query to the winning replica. We develop a multi-stage scalable algorithm to learn these weights. We first construct a large-scale term-query graph from logs and apply a distributed balanced graph partitioning algorithm to cluster each term to a preferred replica. This yields a good but simplistic initial voting table, which we then iteratively refine via cache simulation to capture feedback effects. View details
Preview abstract Submodular optimization generalizes many classic problems in combinatorial optimization and has recently found a wide range of applications in machine learning (e.g., feature engineering and active learning). For many large-scale optimization problems, we are often concerned with the adaptivity complexity of an algorithm, which quantifies the number of sequential rounds where polynomially-many independent function evaluations can be executed in parallel. While low adaptivity is ideal, it is not sufficient for a distributed algorithm to be efficient, since in many practical applications of submodular optimization the number of function evaluations becomes prohibitively expensive. Motivated by these applications, we study the adaptivity and query complexity of adaptive submodular optimization. Our main result is a distributed algorithm for maximizing a monotone submodular function with cardinality constraint k that achieves a (1 − 1/e − ε)-approximation in expectation. This algorithm runs in O(log(n)) adaptive rounds and makes O(n) calls to the function evaluation oracle in expectation. The approximation guarantee and query complexity are optimal, and the adaptivity is nearly optimal. Moreover, the number of queries is substantially less than in previous works. We also extend our results to the submodular cover problem to demonstrate the generality of our algorithm and techniques. View details
Consistent Hashing with Bounded Loads
Mikkel Thorup
Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms (2018), pp. 587-604
Preview abstract Designing algorithms for balanced allocation of clients to servers in dynamic settings is a challenging problem for a variety of reasons. Both servers and clients may be added and/or removed from the system periodically, and the main objectives of allocation algorithms are: the uniformity of the allocation, and the number of moves after adding or removing a server or a client. The most popular solution for our dynamic settings is Consistent Hashing. However, the load balancing of consistent hashing is no better than a random assignment of clients to servers, so with n of each, we expect many servers to be overloaded with Θ(logn/loglogn) clients. In this paper, with n clients and n servers, we get a guaranteed max-load of 2 while only moving an expected constant number of clients for each update. We take an arbitrary user specified balancing parameter c=1+ϵ>1. With m balls and n bins in the system, we want no load above ⌈cm/n⌉. Meanwhile we want to bound the expected number of balls that have to be moved when a ball or server is added or removed. Compared with general lower bounds without capacity constraints, we show that in our algorithm when a ball or bin is inserted or deleted, the expected number of balls that have to be moved is increased only by a multiplicative factor O(1ϵ2) for ϵ≤1 (Theorem 4) and by a factor 1+O(logcc) for ϵ≥1 (Theorem 3). Technically, the latter bound is the most challenging to prove. It implies that we for superconstant c only pay a negligible cost in extra moves. We also get the same bounds for the simpler problem where we instead of a user specified balancing parameter have a fixed bin capacity C for all bins. View details
Almost Optimal Streaming Algorithms for Coverage Problems
Hossein Esfandiari
29th ACM Symposium on Parallelism in Algorithms and Architectures (2017)
Preview abstract Maximum coverage and minimum set cover problems --collectively called coverage problems-- have been studied extensively in streaming models. However, previous research not only achieve sub-optimal approximation factors and space complexities, but also study a restricted set arrival model which makes an explicit or implicit assumption on oracle access to the sets, ignoring the complexity of reading and storing the whole set at once. In this paper, we address the above shortcomings, and present algorithms with improved approximation factor and improved space complexity, and prove that our results are almost tight. Moreover, unlike most of previous work, our results hold on a more general edge arrival model. More specifically, we present (almost) optimal approximation algorithms for maximum coverage and minimum set cover problems in the streaming model with an (almost) optimal space complexity of O~(n), i.e., the space is {\em independent of the size of the sets or the size of the ground set of elements}. These results not only improve over the best known algorithms for the set arrival model, but also are the first such algorithms for the more powerful {\em edge arrival} model. In order to achieve the above results, we introduce a new general sketching technique for coverage functions: This sketching scheme can be applied to convert an α-approximation algorithm for a coverage problem to a $(1-\eps)\alpha$-approximation algorithm for the same problem in streaming, or RAM models. We show the significance of our sketching technique by ruling out the possibility of solving coverage problems via accessing (as a black box) a $(1 \pm \eps)$-approximate oracle (e.g., a sketch function) that estimates the coverage function on any subfamily of the sets. View details
Preview abstract An effective technique for solving optimization problems over massive data sets is to partition the data into smaller pieces, solve the problem on each piece and compute a representative solution from it, and finally obtain a solution inside the union of the representative solutions for all pieces. This technique can be captured via the concept of composable core-sets, and has been recently applied to solve diversity maximization problems as well as several clustering problems [7,15,8]. However, for coverage and submodular maximization problems, impossibility bounds are known for this technique [15]. In this paper, we focus on efficient construction of a randomized variant of composable core-sets where the above idea is applied on a random clustering of the data. We employ this technique for the coverage, monotone and non-monotone submodular maximization problems. Our results significantly improve upon the hardness results for non-randomized core-sets, and imply improved results for submodular maximization in a distributed and streaming settings. The effectiveness of this technique has been confirmed empirically for several machine learning applications [22], and our proof provides a theoretical foundation to this idea. In summary, we show that a simple greedy algorithm results in a 1/3-approximate randomized composable core-set for submodular maximization under a cardinality constraint. Our result also extends to non-monotone submodular functions, and leads to the first 2-round MapReduce-based constant-factor approximation algorithm with O(n) total communication complexity for either monotone or non-monotone functions. Finally, using an improved analysis technique and a new algorithm PseudoGreedy, we present an improved 0.545-approximation algorithm for monotone submodular maximization, which is in turn the first MapReduce-based algorithm beating factor 1/2 in a constant number of rounds. View details
HyperAttention: Large-scale Attention in Linear Time
Amin Karbasi
Amir Zandieh
Insu Han
David Woodruff
HyperAttention: Long-context Attention in Near-Linear Time (2024) (to appear)
Preview abstract In this paper, we introduce a novel approximate attention mechanism dubbed ``HyperAttention``. Despite the rapidly increasing size and complexity of contexts used with Large Language Models (LLM), there is still a dire lack of computationally efficient attention mechanisms scaling better than the naive quadratic time. HyperAttention addresses this gap: it delivers provably linear time complexity with respect to the size of the context, while only incurring a negligible loss in downstream quality. Distinctively, it integrates the principles of Locality Sensitive Hashing (LSH), for efficient detection of heavy elements, along with uniform column sampling, allowing for a good approximation both of the heavy and light components of the attention matrix. HyperAttention provably approximates the attention layer in \textit{linear time}, making it the first practical linear time approximate attention mechanism. Crucially, HyperAttention has a highly-modular design, allowing seamless integration of other rapid low-level implementations, most notably FlashAttention. Empirical evaluations indicate that HyperAttention surpasses the existing methods, achieving orders of magnitude speed-up when compared to prevalent state-of-the-art solutions such as Flash Attention. This breakthrough presents significant implications for enabling the scalability of LLMs to significantly larger contexts. View details

Highlighted work