Smarter Matchmaking for Ride-Hailing Platforms Ride-hailing platforms such as Uber, Lyft, and DiDi must assign drivers to riders every minute without knowing who will request a ride next. Most systems optimize each batch myopically, and this can leave some future riders waiting much longer than necessary. In “Matching Drivers to Riders: A Two-Stage Robust Approach,” accepted for publication in Operations Research, the authors propose a two-stage robust matching model that incorporates uncertainty about future demand. The first stage matches current riders, reserving enough nearby drivers for plausible future scenarios; the second stage then tests these decisions against the most adverse demand patterns to guarantee good performance even in the worst case. The authors prove that finding the best such policy is computationally hard, but they develop constant factor approximation algorithms for several practically relevant cases and test them on large-scale taxi data from Shenzhen, China. Their methods substantially reduce maximum rider waiting times with little or no sacrifice in average travel distances.

The emergence of Multi-Instance GPU (MIG) technology enables us to run smaller machine learning models on partitions of a GPU rather than the entire device, thus improving utilization and reducing energy consumption, albeit with potential performance trade-offs. Meanwhile, the growing energy demands of GPU-equipped data centers motivate the development of online partitioning and scheduling schemes that not only ensure fast job processing but also achieve high energy efficiency. However, achieving energy-tardiness efficiency with manageable algorithmic complexity in large-scale scheduling remains a great challenge, due to the dual objectives of deciding on the GPU partitions and scheduling jobs onto the slices of the heterogeneous partitions. To address this challenge, we propose SMART-MIG, a parallel computing system that combines Mean-Field Multi-Agent Reinforcement Learning (MF-MARL) for large-scale MIG repartitioning with tailored heuristic algorithms for job scheduling. We demonstrate that the complexity of the repartitioning component remains constant even as the number of jobs and GPUs increases. We also establish theoretical lower bounds on energy consumption and tardiness to rigorously benchmark system performance. Finally, extensive experiments show that SMART-MIG improves the energy-tardiness efficiency by 18% compared to its corresponding static-partitioning counterpart, while being only 27% above the theoretical lower bound on energy consumption.

We design efficient approximation algorithms for maximizing the expectation of the supremum of families of Gaussian random variables. In particular, let OPT:=max_σ_1,⋯,σ_n𝔼[∑_j=1^mmax_i∈ S_j X_i], where X_i are Gaussian, S_j⊂[n] and ∑_iσ_i^2=1, then our theoretical results include: - We characterize the optimal variance allocation – it concentrates on a small subset of variables as |S_j| increases, - A polynomial time approximation scheme (PTAS) for computing OPT when m=1, and - An O(log n) approximation algorithm for computing OPT for general m>1. Such expectation maximization problems occur in diverse applications, ranging from utility maximization in auctions markets to learning mixture models in quantitative genetics.

The goal in the stochastic vertex cover problem is to obtain an approximately minimum vertex cover for a graph G^⋆ that is realized by sampling each edge independently with some probability p∈ (0, 1] in a base graph G = (V, E). The algorithm is given the base graph G and the probability p as inputs, but its only access to the realized graph G^⋆ is through queries on individual edges in G that reveal the existence (or not) of the queried edge in G^⋆. In this paper, we resolve the central open question for this problem: to find a (1+ε)-approximate vertex cover using only O_ε(n/p) edge queries. Prior to our work, there were two incomparable state-of-the-art results for this problem: a (3/2+ε)-approximation using O_ε(n/p) queries (Derakhshan, Durvasula, and Haghtalab, 2023) and a (1+ε)-approximation using O_ε((n/p)·RS(n)) queries (Derakhshan, Saneian, and Xun, 2025), where RS(n) is known to be at least 2^Ω(log n/loglog n) and could be as large as n/2^Θ(log^* n). Our improved upper bound of O_ε(n/p) matches the known lower bound of Ω(n/p) for any constant-factor approximation algorithm for this problem (Behnezhad, Blum, and Derakhshan, 2022). A key tool in our result is a new concentration bound for the size of minimum vertex cover on random graphs, which might be of independent interest.

Matching demand (riders) to supply (drivers) efficiently is a fundamental problem for ride-sharing platforms who need to match the riders (almost) as soon as the request arrives with only partial knowledge about future ride requests. A myopic approach that computes an optimal matching for current requests ignoring future uncertainty can be highly sub-optimal. In this paper, we consider a two-stage robust optimization framework for this matching problem where future demand uncertainty is modeled using a set of demand scenarios (specified explicitly or implicitly). The goal is to match the current request to drivers (in the first stage) so that the cost of first stage matching and the worst case cost over all scenarios for the second stage matching is minimized. We show that the two-stage robust matching is NP-hard under various cost functions and present constant approximation algorithms for different settings of our two-stage problem. Furthermore, we test our algorithms on real-life taxi data from the city of Shenzhen and show that they substantially improve upon myopic solutions and reduce the maximum wait time of the second-stage riders by an average of $30\%$ in our experimental results.

High-Bandwidth Memory (HBM) is a decade-old memory technology that is increasingly commonly being used in highly-parallel machines such as GPUs and multicores. Comparatively, HBM has higher bandwidth, smaller capacity, and similar latency to other DRAM technologies. Many systems use both HBM and other DRAM technologies, where HBM is naturally closer to the processor in the conceptual memory hierarchy. Thus, a natural resulting question is how one should best manage a collection of processes running on a HBM/DRAM memory hierarchy. Prior work introduced a theoretical model for addressing this question, and gave a competitive policy for the objective of minimizing makespan. Our main technical contribution is to give a competitive policy for the more commonly appropriate total/average response/completion time objective. However, we believe the broader, and more important contribution, is to make explicit the case (hinted at in the prior literature) that managing an HBM/DRAM hierarchy should be thought of as a parallel scheduling problem. To that end, we introduce a new online scheduling model that we call the semi-normal model. We then show how to use a competitive algorithm for scheduling in the semi-normal model as a black box to obtain a competitive algorithm for managing a HBM/DRAM memory hierarchy. Thus, as a result of this black-box conversion, competitiveness results in the semi-normal model translate (essentially) automatically into competitiveness results in the HBM/DRAM management model. Our main technical result is then an application of such a translation. That is, we show that a natural variant of the Round Robin (processor sharing) algorithm, naturally adapted for the seminormal model, is competitive for the objective of average/total completion time. Thus, we obtain an algorithm for managing a HBM/DRAM hierarchy that is competitive for the objective of average/total completion time, using this black-box reduction.

Increasing demand from AI/ML workloads is exacerbating the rising energy consumption of data centers. Recent advances in hardware such as NVIDIA's Multi Instance GPUs (MIGs) offer improvements in flexibility and computational power and the opportunity for data centers to manage incoming jobs in energy-efficient ways, while maintaining acceptable performance. The challenge in achieving this multi-objective in a MIG environment through job scheduling is multi-faceted. Firstly, for a given MIG configuration, one seeks an easy-toimplement scheduling algorithm which selects a job from the queue as well as decides on which slice in the configuration the job runs. Secondly, for the identified scheduling algorithm, a particular MIG configuration may not always be suitable (as the workload fluctuates) and may need to be repartitioned. We tackle both problems using simulations and reinforcement learning (RL). We present a dynamic repartitioning scheduling framework for a single MIG as a solution to a multi-objective heterogeneous machine scheduling problem with preemption. In particular, we compare four scheduling algorithms and identify a promising one. Then, we employ reinforcement learning to perform dynamic repartitioning over a day. Furthermore, using a diurnal workload pattern based on real-world data center traces, we demonstrate the superiority of our dynamic repartitioning algorithm over twice-daily repartitioning (26%), static partitioning (31%) and no partitioning at all (68%) according to a multi-objective function of energy consumption and tardiness. Our results indicate specific preferred configurations at different times of the day under different queue conditions, suggesting a policy for predictive and automatic reconfiguration.

We give an algorithm for the fully-dynamic carpooling problem with recourse: Edges arrive and depart online from a graph G with n nodes according to an adaptive adversary. Our goal is to maintain an orientation H of G that keeps the discrepancy, defined as max_v ∈ V |deg_H^+(v) - deg_H^-(v)|, small at all times. We present a simple algorithm and analysis for this problem with recourse based on cycles that simplifies and improves on a result of Gupta et al. [SODA '22].

We revisit the well-known online traveling salesman problem (OLTSP) and its extension, the online dial-a-ride problem (OLDARP). A server starting at a designated origin in a metric space, is required to serve online requests, and return to the origin such that the completion time is minimized. The SmartStart algorithm, introduced by Ascheuer et al., incorporates a waiting approach into an online schedule-based algorithm and attains the optimal upper bound of 2 for the OLTSP and the OLDARP if each schedule is optimal. Using the Christofides' heuristic to approximate each schedule leads to the currently best upper bound of (7 + sqrt(13)) / 4 approximately 2.6514 in polynomial time. In this study, we investigate how an online algorithm with predictions, a recent popular framework (i.e. the so-called learning-augmented algorithms), can be used to improve the best competitive ratio in polynomial time. In particular, we develop a waiting strategy with online predictions, each of which is only a binary decision-making for every schedule in a whole route, rather than forecasting an entire set of requests in the beginning (i.e. offline predictions). That is, it does not require knowing the number of requests in advance. The proposed online schedule-based algorithm can achieve 1.1514 * lambda + 1.5-consistency and 1.5 + 1.5 / (2.3028 * lambda - 1)-robustness in polynomial time, where lambda lies in the interval (1/theta, 1] and theta is set to (1 + sqrt(13)) / 2 approximately 2.3028. The best consistency tends to approach to 2 when lambda is close to 1/theta. Meanwhile, we show any online schedule-based algorithms cannot derive a competitive ratio of less than 2 even with perfect online predictions.

We ported the Sony Pictures Imageworks version of the Arnold Renderer to the GPU using NVIDIA's OptiX ray tracing toolkit. This required modifying algorithms to run efficiently on the GPU, the use of new software methodologies to better share source code between the host and device renderers, and a reevaluation of what contributes to poor performance on the device. We share here the key decisions we made to overcome these challenges and the valuable lessons we learned during our journey in implementing the Sony Pictures Evolved Arnold Renderer (Spear) on the GPU.

Computing is an indispensable tool in addressing climate change, but it also contributes to a significant and steadily increasing carbon footprint, partly due to the exponential growth in energy-demanding workloads, such as artificial intelligence (AI). While hardware specialization has become the primary driver of operational energy efficiency improvements, it introduces new challenges including increased embodied emission, and a rise in complexity of operations of heterogeneous and dynamic datacenters. We posit that while specialization is necessary for sustainable computing, to fully harness its power, the academic and technical community must address the specific challenges arising from embracing it. We enumerate and analyze key challenges that specialization introduces across software, system design, and operations, and their potential impact on carbon cost, and propose a way forward for each identified area. Furthermore, we argue that intricate relationships exist across the life-cycle of compute systems, which must be understood, modeled, and analyzed to identify the most beneficial Pareto frontiers for carbon life-cycle efficiency. We analyze these trade-offs and offer an approach to address them using a unified metric and framework.

This paper addresses point-to-point packet routing in undirected networks, which is the most important communication primitive in most networks. The main result proves the existence of routing tables that deterministically guarantee a polylog-competitive completion-time: In any undirected network, it is possible to give each node simple stateless deterministic local forwarding rules, such that, any adversarially chosen set of packets are delivered as fast as possible, up to polylog factors. All previous routing strategies crucially required randomization for both route selection and packet scheduling. The core technical contribution of this paper is a newlocal packet scheduling result of independent interest. This scheduling strategy integrates well with recent sparse semi-oblivious path selection strategies. Such strategies deterministically select not one but several candidate paths for each packet and require a global coordinator to know all packets to adaptively select a single good path from those candidates for each packet. Of course, global knowledge of all packets is exactly what local routing tables cannot have. Another challenge is that, even if a single path is selected for each packet, no strategy for scheduling packets along low-congestion paths that is both local and deterministic is known. Our novel scheduling strategy utilizes the fact that every semi-oblivious routing strategy uses only a small (polynomial) subset of candidate routes. It overcomes the issue of global coordination by furthermore being provably robust to adversarial noise. This avoids the issue of having to choose a single path per packet by treating congestion caused by ineffective candidate paths as noise. Beyond more efficient routing tables, our results can be seen as making progress on fundamental questions regarding the importance and power of randomization in network communications and distributed computing. For example, our results imply the first deterministic universally-optimal algorithms in the distributed supported-CONGEST model for many important global distributed tasks, including computing minimum spanning trees, approximate shortest paths, and part-wise aggregates.

We consider circuit routing with an objective of minimizing energy, in a network of routers that are speed scalable and that may be shutdown when idle. It is known that this energy minimization problem can be reduced to a capacitated flow network design problem, where vertices have a common capacity but arbitrary costs, and the goal is to choose a minimum cost collection of vertices whose induced subgraph will support the specified flow requirements. For the multicast (single-sink) capacitated design problem we give a polynomial-time algorithm that is O(log3n)- approximate with O(log4n) congestion. This translates back to a O(log4α+3n)-approximation for the multicast energy-minimization routing problem, where α is the polynomial exponent in the dynamic power used by a router. For the unicast (multicommodity) capacitated design problem we give a polynomial-time algorithm that is O(log5n)-approximate with O(log12n) congestion, which translates back to a O(log12α+5n)-approximation for the unicast energy-minimization routing problem.

In the process of redistricting, one important metric is the number of competitive districts, that is, districts where both parties have a reasonable chance of winning a majority of votes. Competitive districts are important for achieving proportionality, responsiveness, and other desirable qualities; some states even directly list competitiveness in their legally-codified districting requirements. In this work, we discuss the problem of drawing plans with at least a fixed number of competitive districts. In addition to the standard, “vote-band” measure of competitivenesss (i.e., how close was the last election?), we propose a measure that explicitly considers “swing voters” - the segment of the population that may choose to vote either way, or not vote at all, in a given election. We present two main, contrasting results. First, from a computational complexity perspective, we show that the task of drawing plans with competitive districts is NP-hard, even on very natural instances where the districting task itself is easy (e.g., small rectangular grids of population-balanced cells). Second, however, we show that a simple hill-climbing procedure can in practice find districtings on real states in which all the districts are competitive. We present the results of the latter on the precinct-level graphs of the U.S. states of North Carolina and Arizona, and discuss trade-offs between competitiveness and other desirable qualities.

Gale and Shapley’s stability criterion enjoys a rich mathematical structure, which propelled its application in various settings. Although immensely popular, the approach by Gale and Shapley cannot encompass all the different features that arise in applications, motivating the search for alternative solution concepts. We investigate alternatives that rely on the concept of internal stability, a notion introduced for abstract games by von Neumann and Morgenstern and motivated by the need of finding a set of mutually compatible solutions. The set of stable matchings is internally stable. However, the class of internally stable sets is much richer, for an internally stable set of matchings may also include unstable matchings and/or exclude stable ones. In this paper, we focus on two families of internally stable sets of matchings: von Neumann-Morgenstern stable and internally closed. We study algorithmic questions around those concepts in both the marriage and the roommate models. One of our results implies that, in the marriage model, internally closed sets are an alternative to stable matchings that is as tractable as stable matchings themselves, a fairly rare occurrence in the area. Both our positive and negative results rely on new structural insights and extensions of classical algebraic structures associated with sets of matchings, which we believe to be of independent interest.

The modern network aims to prioritize critical traffic over non-critical traffic and effectively manage traffic flow. This necessitates proper buffer management to prevent the loss of crucial traffic while minimizing the impact on non-critical traffic. Therefore, the algorithm's objective is to control which packets to transmit and which to discard at each step. In this study, we initiate the learning-augmented online packet scheduling with deadlines and provide a novel algorithmic framework to cope with the prediction. We show that when the prediction error is small, our algorithm improves the competitive ratio while still maintaining a bounded competitive ratio, regardless of the prediction error.

We give a randomized algorithm for online metric b-matching that is O(log2⁡k) competitive, where k is the number of server locations, by giving a black box reduction from b-matching on a hierarchically separated tree to a uniform metric space.

Algorithms with predictions is a recent framework that has been used to overcome pessimistic worst-case bounds in incomplete information settings. In the context of scheduling, very recent work has leveraged machine-learned predictions to design algorithms that achieve improved approximation ratios in settings where the processing times of the jobs are initially unknown. In this paper, we study the speed-robust scheduling problem where the speeds of the machines, instead of the processing times of the jobs, are unknown and augment this problem with predictions. Our main result is an algorithm that achieves a $\min\{\eta^2(1+\alpha), (2 + 2/\alpha)\}$ approximation, for any $\alpha \in (0,1)$, where $\eta \geq 1$ is the prediction error. When the predictions are accurate, this approximation outperforms the best known approximation for speed-robust scheduling without predictions of $2-1/m$, where $m$ is the number of machines, while simultaneously maintaining a worst-case approximation of $2 + 2/\alpha$ even when the predictions are arbitrarily wrong. In addition, we obtain improved approximations for three special cases: equal job sizes, infinitesimal job sizes, and binary machine speeds. We also complement our algorithmic results with lower bounds. Finally, we empirically evaluate our algorithm against existing algorithms for speed-robust scheduling.