arXiv subjects

Sergey Pupyrev

Publications and source records attributed to Sergey Pupyrev.

At least 19 recordsLinked to original sources

arXiv2026

Tensor Seeks Layout: Formalizing Layout Selection for ML Compilers

Modern machine learning compilers select tensor memory layouts to minimize execution cost under hardware constraints. Layout selection is global: an operator may be fastest under one layout while its consumers prefer another, and aligning these preferences requires explicit layout conversions that can hurt model performance. Despite its practical importance, layout selection lacks a formal basis, so current compilers rely on ad-hoc heuristics. This paper presents the first formal study of layout selection in machine learning compilers. We formulate the problem as combinatorial optimization over dataflow graphs, minimizing the sum of operator execution costs and the per-tensor cost of these conversions. Our theoretical analysis shows that optimal layout selection is computationally hard, even for programs containing only matrix multiplications over two-dimensional tensors. We design an optimal polynomial-time algorithm for dataflow graphs of bounded treewidth. For general instances, we give a weighted MaxSAT encoding that an off-the-shelf solver can optimize. The formulation unifies several existing layout optimization strategies, including XLA's layout assignment, partition dimension selection in systolic array compilers, and layout planning in mobile GPU optimizers. We implement the formalization in a production compiler for an AI accelerator and measure the execution time of the compiled models under greedy heuristics, the compiler's rule-based strategy, and an optimal solver. Simple heuristics degrade execution time by up to $5\times$ on some workloads. Where the compiler's cost model is accurate, the solver matches or beats the rule-based strategy. On workloads with complex data movement it falls behind, and since the solver minimizes the stated objective exactly, that gap isolates cost-model error from search quality, showing where compiler effort actually pays off.

cs.PL↗

arXiv2026

A Gap in the 42-Queue Layout Algorithm for Planar Graphs

A queue layout of a graph consists of a linear order of the vertices and a partition of the edges into queues so that no two edges in a single queue are nested. The minimum number of queues needed in a queue layout of a graph is called its queue number. The planar product structure theorem states that every planar graph is a subgraph of the strong product of a graph of simple treewidth at most $3$, a clique $K_3$, and a path. Such a strong product admits a queue layout with $49$ queues (Wood, 2005), which implies that the queue number of planar graphs is at most $49$. Recently, Bekos, Gronemann, and Raftopoulou (Algorithmica, 2023) investigated how the general approach based on the product structure can be optimized for planar graphs. They claim that by appropriately reordering the three vertices in each bag arising from a tripod, it is possible to reduce the queue number of planar graphs to~$42$. In this note we highlight a gap in their queue layout algorithm: one of the choices required by the algorithm is not guaranteed to exist. Hence the claimed upper bound of $42$ queues is not established by the published proof.

cs.DM↗

arXiv2026

How to Draw a Planar Graph: An Experimental Evaluation

Planar graphs are central to graph drawing, with extensive results on planar layouts and related structures. Every planar graph admits a planar straight-line drawing, and algorithms can guarantee additional geometric or combinatorial properties. However, it is unclear which algorithms work best in practice. Even for small graphs with near-perfect manual drawings, standard algorithms might produce poor spacing, distorted faces, or small angles. We present an experimental evaluation of planar graph drawing algorithms on a large benchmark collection of small and medium-sized planar graphs ($10$--$400$ vertices). The study compares established algorithms from the graph drawing literature, practical force-directed and pressure-based heuristics, and new optimization-based methods that directly improve visual properties such as edge-length uniformity, face-area balance, and angular resolution. The results show that no evaluated algorithm is best across all aesthetic criteria, and optimizing one visual property often worsens another. Directly optimizing visual criteria improves targeted scores, and score-guided combination of several methods gives the best aggregate results, but no simple algorithm emerges as a clear universal default. Designing a simple, robust algorithm that performs well across graph families and aesthetic criteria therefore remains an open practical problem.

cs.CG↗

arXiv2024

Forbidden Patterns in Mixed Linear Layouts

An ordered graph is a graph with a total order over its vertices. A linear layout of an ordered graph is a partition of the edges into sets of either non-crossing edges, called stacks, or non-nesting edges, called queues. The stack (queue) number of an ordered graph is the minimum number of required stacks (queues). Mixed linear layouts combine these layouts by allowing each set of edges to form either a stack or a queue. The minimum number of stacks plus queues is called the mixed page number. It is well known that ordered graphs with small stack number are characterized, up to a function, by the absence of large twists (that is, pairwise crossing edges). Similarly, ordered graphs with small queue number are characterized by the absence of large rainbows (that is, pairwise nesting edges). However, no such characterization via forbidden patterns is known for mixed linear layouts. We address this gap by introducing patterns similar to twists and rainbows, which we call thick patterns; such patterns allow a characterization, again up to a function, of mixed linear layouts of bounded-degree graphs. That is, we show that a family of ordered graphs with bounded maximum degree has bounded mixed page number if and only if the size of the largest thick pattern is bounded. In addition, we investigate an exact characterization of ordered graphs whose mixed page number equals a fixed integer $ k $ via a finite set of forbidden patterns. We show that for every $ k \ge 2 $, there is no such characterization, which supports the nature of our first result.

cs.DM↗

arXiv2024

Transforming Stacks into Queues: Mixed and Separated Layouts of Graphs

Some of the most important open problems for linear layouts of graphs ask for the relation between a graph's queue number and its stack number or mixed number. In such, we seek a vertex order and edge partition of $G$ into parts with pairwise non-crossing edges (a stack) or with pairwise non-nesting edges (a queue). Allowing only stacks, only queues, or both, the minimum number of required parts is the graph's stack number $sn(G)$, queue number $qn(G)$, and mixed number $mn(G)$, respectively. Already in 1992, Heath and Rosenberg asked whether $qn(G)$ is bounded in terms of $sn(G)$, that is, whether stacks "can be transformed into" queues. This is equivalent to bipartite $3$-stack graphs having bounded queue number (Dujmovi\'c and Wood, 2005). Recently, Alam et al. asked whether $qn(G)$ is bounded in terms of $mn(G)$, which we show to also be equivalent to the previous questions. We approach the problem by considering separated linear layouts of bipartite graphs. In this natural setting all vertices of one part must precede all vertices of the other part. Separated stack and queue numbers coincide, and for fixed vertex orders, graphs with bounded separated stack/queue number can be characterized and efficiently recognized, whereas the separated mixed layouts are more challenging. In this work, we thoroughly investigate the relationship between separated and non-separated, mixed and pure linear layouts.

math.CO↗

arXiv2024

The Price of Upwardness

Not every directed acyclic graph (DAG) whose underlying undirected graph is planar admits an upward planar drawing. We are interested in pushing the notion of upward drawings beyond planarity by considering upward $k$-planar drawings of DAGs in which the edges are monotonically increasing in a common direction and every edge is crossed at most $k$ times for some integer $k \ge 1$. We show that the number of crossings per edge in a monotone drawing is in general unbounded for the class of bipartite outerplanar, cubic, or bounded pathwidth DAGs. However, it is at most two for outerpaths and it is at most quadratic in the bandwidth in general. From the computational point of view, we prove that testing upward-$k$-planarity is NP-complete already for $k=1$ and even for restricted instances for which upward planarity testing is polynomial. On the positive side, we can decide in linear time whether a single-source DAG admits an upward 1-planar drawing in which all vertices are incident to the outer face.

cs.CG↗

arXiv2024

Stale Profile Matching

Profile-guided optimizations rely on profile data for directing compilers to generate optimized code. To achieve the maximum performance boost, profile data needs to be collected on the same version of the binary that is being optimized. In practice however, there is typically a gap between the profile collection and the release, which makes a portion of the profile invalid for optimizations. This phenomenon is known as profile staleness, and it is a serious practical problem for data-center workloads both for compilers and binary optimizers. In this paper we thoroughly study the staleness problem and propose the first practical solution for utilizing profiles collected on binaries built from several revisions behind the release. Our algorithm is developed and implemented in a mainstream open-source post-link optimizer, BOLT. An extensive evaluation on a variety of standalone benchmarks and production services indicates that the new method recovers up to $0.8$ of the maximum BOLT benefit, even when most of the input profile data is stale and would have been discarded by the optimizer otherwise.

cs.PL↗

arXiv2023

On Families of Planar DAGs with Constant Stack Number

A $k$-stack layout (or $k$-page book embedding) of a graph consists of a total order of the vertices, and a partition of the edges into $k$ sets of non-crossing edges with respect to the vertex order. The stack number of a graph is the minimum $k$ such that it admits a $k$-stack layout. In this paper we study a long-standing problem regarding the stack number of planar directed acyclic graphs (DAGs), for which the vertex order has to respect the orientation of the edges. We investigate upper and lower bounds on the stack number of several families of planar graphs: We improve the constant upper bounds on the stack number of single-source and monotone outerplanar DAGs and of outerpath DAGs, and improve the constant upper bound for upward planar 3-trees. Further, we provide computer-aided lower bounds for upward (outer-) planar DAGs.

math.CO↗

arXiv2023

Linear Layouts of Bipartite Planar Graphs

A linear layout of a graph $ G $ consists of a linear order $\prec$ of the vertices and a partition of the edges. A part is called a queue (stack) if no two edges nest (cross), that is, two edges $ (v,w) $ and $ (x,y) $ with $ v \prec x \prec y \prec w $ ($ v \prec x \prec w \prec y $) may not be in the same queue (stack). The best known lower and upper bounds for the number of queues needed for planar graphs are 4 [Alam et al., Algorithmica 2020] and 42 [Bekos et al., Algorithmica 2022], respectively. While queue layouts of special classes of planar graphs have received increased attention following the breakthrough result of [Dujmović et al., J. ACM 2020], the meaningful class of bipartite planar graphs has remained elusive so far, explicitly asked for by Bekos et al. In this paper we investigate bipartite planar graphs and give an improved upper bound of 28 by refining existing techniques. In contrast, we show that two queues or one queue together with one stack do not suffice; the latter answers an open question by Pupyrev [GD 2018]. We further investigate subclasses of bipartite planar graphs and give improved upper bounds; in particular we construct 5-queue layouts for 2-degenerate quadrangulations.

math.CO↗

arXiv2022

Optimizing Function Layout for Mobile Applications

Function layout, also referred to as function reordering or function placement, is one of the most effective profile-guided compiler optimizations. By reordering functions in a binary, compilers are able to greatly improve the performance of large-scale applications or reduce the compressed size of mobile applications. Although the technique has been studied in the context of large-scale binaries, no recent study has investigated the impact of function layout on mobile applications. In this paper we develop the first principled solution for optimizing function layouts in the mobile space. To this end, we identify two important optimization goals, the compressed code size and the cold start-up time of a mobile application. Then we propose a formal model for the layout problem, whose objective closely matches the goals. Our novel algorithm to optimize the layout is inspired by the classic balanced graph partitioning problem. We carefully engineer and implement the algorithm in an open source compiler, LLVM. An extensive evaluation of the new method on large commercial mobile applications indicates up to 2% compressed size reduction and up to 3% start-up time improvement on top of the state-of-the-art approach.

cs.PL↗

arXiv2022

Minimum Coverage Instrumentation

Modern compilers leverage block coverage profile data to carry out downstream profile-guided optimizations to improve the runtime performance and the size of a binary. Given a control-flow graph $G=(V, E)$ of a function in the binary, where nodes in $V$ correspond to basic blocks (sequences of instructions that are always executed sequentially) and edges in $E$ represent jumps in the control flow, the goal is to know for each block $u \in V$ whether $u$ was executed during a session. To this end, extra instrumentation code that records when a block is executed needs to be added to the binary. This extra code creates a time and space overhead, which one would like to minimize as much as possible. Motivated by this application, we study the Minimum Coverage Instrumentation problem, where the goal is to find a minimum size subset of blocks to instrument such that the coverage of the remaining blocks in the graph can be inferred from the coverage status of the instrumented subset. Our main result is an algorithm to find an optimal instrumentation strategy and to carry out the inference in $O(|E|)$ time. We also study variants of this basic problem in which we are interested in learning the coverage of edges instead of the nodes, or when we are only allowed to instrument edges instead of the nodes.

cs.DS↗

arXiv2022

Queue Layouts of Two-Dimensional Posets

The queue number of a poset is the queue number of its cover graph when the vertex order is a linear extension of the poset. Heath and Pemmaraju conjectured that every poset of width $w$ has queue number at most $w$. The conjecture has been confirmed for posets of width $w=2$ and for planar posets with $0$ and $1$. In contrast, the conjecture has been refused by a family of general (non-planar) posets of width $w>2$. In this paper, we study queue layouts of two-dimensional posets. First, we construct a two-dimensional poset of width $w > 2$ with queue number $2(w - 1)$, thereby disproving the conjecture for two-dimensional posets. Second, we show an upper bound of $w(w+1)/2$ on the queue number of such posets, thus improving the previously best-known bound of $(w-1)^2+1$ for every $w > 3$.

math.CO↗

arXiv2022

Robust and fair work allocation

In today's digital world, interaction with online platforms is ubiquitous, and thus content moderation is important for protecting users from content that do not comply with pre-established community guidelines. Having a robust content moderation system throughout every stage of planning is particularly important. We study the short-term planning problem of allocating human content reviewers to different harmful content categories. We use tools from fair division and study the application of competitive equilibrium and leximin allocation rules. Furthermore, we incorporate, to the traditional Fisher market setup, novel aspects that are of practical importance. The first aspect is the forecasted workload of different content categories. We show how a formulation that is inspired by the celebrated Eisenberg-Gale program allows us to find an allocation that not only satisfies the forecasted workload, but also fairly allocates the remaining reviewing hours among all content categories. The resulting allocation is also robust as the additional allocation provides a guardrail in cases where the actual workload deviates from the predicted workload. The second practical consideration is time dependent allocation that is motivated by the fact that partners need scheduling guidance for the reviewers across days to achieve efficiency. To address the time component, we introduce new extensions of the various fair allocation approaches for the single-time period setting, and we show that many properties extend in essence, albeit with some modifications. Related to the time component, we additionally investigate how to satisfy markets' desire for smooth allocation (e.g., partners for content reviewers prefer an allocation that does not vary much from time to time, to minimize staffing switch). We demonstrate the performance of our proposed approaches through real-world data obtained from Meta.

cs.GT↗

arXiv2021

Matching Algorithms for Blood Donation

Global demand for donated blood far exceeds supply, and unmet need is greatest in low- and middle-income countries; experts suggest that large-scale coordination is necessary to alleviate demand. Using the Facebook Blood Donation tool, we conduct the first large-scale algorithmic matching of blood donors with donation opportunities. While measuring actual donation rates remains a challenge, we measure donor action (e.g., making a donation appointment) as a proxy for actual donation. We develop automated policies for matching donors with donation opportunities, based on an online matching model. We provide theoretical guarantees for these policies, both regarding the number of expected donations and the equitable treatment of blood recipients. In simulations, a simple matching strategy increases the number of donations by 5-10%; a pilot experiment with real donors shows a 5% relative increase in donor action rate (from 3.7% to 3.9%). When scaled to the global Blood Donation tool user base, this corresponds to an increase of around one hundred thousand users taking action toward donation. Further, observing donor action on a social network can shed light onto donor behavior and response to incentives. Our initial findings align with several observations made in the medical and social science literature regarding donor behavior.

cs.AI↗

arXiv2021

On the Extended TSP Problem

We initiate the theoretical study of Ext-TSP, a problem that originates in the area of profile-guided binary optimization. Given a graph $G=(V, E)$ with positive edge weights $w: E \rightarrow R^+$, and a non-increasing discount function $f(\cdot)$ such that $f(1) = 1$ and $f(i) = 0$ for $i > k$, for some parameter $k$ that is part of the problem definition. The problem is to sequence the vertices $V$ so as to maximize $\sum_{(u, v) \in E} f(|d_u - d_v|)\cdot w(u,v)$, where $d_v \in \{1, \ldots, |V| \}$ is the position of vertex~$v$ in the sequence. We show that \prob{Ext-TSP} is APX-hard to approximate in general and we give a $(k+1)$-approximation algorithm for general graphs and a PTAS for some sparse graph classes such as planar or treewidth-bounded graphs. Interestingly, the problem remains challenging even on very simple graph classes; indeed, there is no exact $n^{o(k)}$ time algorithm for trees unless the ETH fails. We complement this negative result with an exact $n^{O(k)}$ time algorithm for trees.

cs.DS↗

arXiv2021

The Mixed Page Number of Graphs

A linear layout of a graph typically consists of a total vertex order, and a partition of the edges into sets of either non-crossing edges, called stacks, or non-nested edges, called queues. The stack (queue) number of a graph is the minimum number of required stacks (queues) in a linear layout. Mixed linear layouts combine these layouts by allowing each set of edges to form either a stack or a queue. In this work we initiate the study of the mixed page number of a graph which corresponds to the minimum number of such sets. First, we study the edge density of graphs with bounded mixed page number. Then, we focus on complete and complete bipartite graphs, for which we derive lower and upper bounds on their mixed page number. Our findings indicate that combining stacks and queues is more powerful in various ways compared to the two traditional layouts.

cs.DS↗

arXiv2020

Lazy Queue Layouts of Posets

We investigate the queue number of posets in terms of their width, that is, the maximum number of pairwise incomparable elements. A long-standing conjecture of Heath and Pemmaraju asserts that every poset of width w has queue number at most w. The conjecture has been confirmed for posets of width w=2 via so-called lazy linear extension. We extend and thoroughly analyze lazy linear extensions for posets of width w > 2. Our analysis implies an upper bound of $(w-1)^2 +1$ on the queue number of width-w posets, which is tight for the strategy and yields an improvement over the previously best-known bound. Further, we provide an example of a poset that requires at least w+1 queues in every linear extension, thereby disproving the conjecture for posets of width w > 2.

cs.DS↗

arXiv2020

The Turing Test for Graph Drawing Algorithms

Do algorithms for drawing graphs pass the Turing Test? That is, are their outputs indistinguishable from graphs drawn by humans? We address this question through a human-centred experiment, focusing on `small' graphs, of a size for which it would be reasonable for someone to choose to draw the graph manually. Overall, we find that hand-drawn layouts can be distinguished from those generated by graph drawing algorithms, although this is not always the case for graphs drawn by force-directed or multi-dimensional scaling algorithms, making these good candidates for Turing Test success. We show that, in general, hand-drawn graphs are judged to be of higher quality than automatically generated ones, although this result varies with graph size and algorithm.

cs.SI↗