arXiv subjects

Matthias Langer

Publications and source records attributed to Matthias Langer.

At least 19 recordsLinked to original sources

arXiv2026

Spectral coarse spaces based on indefinite operators: the $H_k$-GenEO method

GenEO (`Generalised Eigenvalue problems on the Overlap') is a method for constructing coarse spaces used in the preconditioning of iterative solvers for discrete PDEs. This method combines a (small) number of modes of local PDE eigenproblems to obtain a global coarse space. A coarse solve is then combined with local solves of the global PDE to obtain the preconditioner. A substantial theory for GenEO has been developed for the case when the local elgenproblems are positive semi-definite. This has been applied mostly to positive definite global PDEs, but also recently extended to the case of convection--diffusion--reaction problems, which may be neither self-adjoint, nor positive definite. However, when the global problem is highly indefinite, coarse spaces built from positive semi-definite local eigenproblems fail to be robust in practice. In this paper we consider highly indefinite global PDE problems, characterised by a large parameter $k$ (allowing also highly variable coefficients), and we develop a new spectral coarse space built from solving eigenvalue problems based on \textit{local copies of the global problem}. We put no constraint on the diameters of the local domains, thus allowing the local eigenvalue problems to be indefinite. The new method (which we call $H_k$-GenEO) is seen to be much more robust as $k$ increases than methods based on positive semi-definite eigenproblems. We provide sufficient conditions for robustness of the preconditioned GMRES iterative method, in terms of the tolerance of the local eigenproblems and the size of the subdomains for the local PDE solves. In practice the method is observed to be robust with respect to $k$ under even weaker conditions on the local eigenproblem tolerance. The experiments also suggest the method can be resilient to high variation in PDE coefficients.

math.NA↗

arXiv2026

The direct spectral problem for indefinite canonical systems

For indefinite (Pontryagin space) canonical systems that contain an inner singularity we prove the existence of generalised boundary values at the singularity, which are used to formulate interface conditions. With the help of such interface conditions we construct the monodromy matrix of the canonical system and write it as a product of matrices, which separates the contributions of the Hamiltonian function and the finitely many discrete parameters that are associated with the singularity.

math.SP↗

arXiv2026

HierarchicalKV: A GPU Hash Table with Cache Semantics for Continuous Online Embedding Storage

Traditional GPU hash tables preserve every inserted key -- a dictionary assumption that wastes scarce High Bandwidth Memory (HBM) when embedding tables routinely exceed single-GPU capacity. We challenge this assumption with cache semantics, where policy-driven eviction is a first-class operation. We introduce HierarchicalKV (HKV), the first general-purpose GPU hash table library whose normal full-capacity operating contract is cache-semantic: each full-bucket upsert (update-or-insert) is resolved in place by eviction or admission rejection rather than by rehashing or capacity-induced failure. HKV co-designs four core mechanisms -- cache-line-aligned buckets, in-line score-driven upsert, score-based dynamic dual-bucket selection, and triple-group concurrency -- and uses tiered key-value separation as a scaling enabler beyond HBM. On an NVIDIA H100 NVL GPU, HKV achieves up to 3.9 billion key-value pairs per second (B-KV/s) find throughput, stable across load factors 0.50-1.00 (<5% variation), and delivers 1.4x higher find throughput than WarpCore (the strongest dictionary-semantic GPU baseline at lambda=0.50) and up to 2.6-9.4x over indirection-based GPU baselines. Since its open-source release in October 2022, HKV has been integrated into multiple open-source recommendation frameworks.

cs.DB↗

arXiv2025

Continuous fragmentation equations in weighted $L^1$ spaces

We investigate an integro-differential equation that models the evolution of fragmenting clusters. We assume cluster size to be a continuous variable and allow for situations in which mass is not necessarily conserved during each fragmentation event. We formulate the initial-value problem as an abstract Cauchy problem (ACP) in an appropriate weighted $L^1$ space, and apply perturbation results to prove that a unique, physically relevant classical solution of the ACP is given by a strongly continuous semigroup for a wide class of initial conditions. Moreover, we show that it is often possible to identify a weighted $L^1$ space in which this semigroup is analytic, leading to the existence of a unique, physically relevant classical solution for all initial conditions belonging to that space. For some specific fragmentation coefficients, we provide examples of weighted $L^1$ spaces where our results can be applied.

math.FA↗

arXiv2025

Achieving wavenumber robustness in domain decomposition for heterogeneous Helmholtz equation: an overview of spectral coarse spaces

Solving time-harmonic wave propagation problems in the frequency domain within heterogeneous media poses significant mathematical and computational challenges, particularly in the high-frequency regime. Among the available numerical approaches, domain decomposition methods are widely regarded as effective due to their suitability for parallel computing and their capacity to maintain robustness with respect to physical parameters, such as the wavenumber. These methods can achieve near-constant time-to-solution as the wavenumber increases, though often at the expense of a computationally intensive coarse correction step. This work focuses on identifying the best algorithms and numerical strategies for benchmark problems modelled by the Helmholtz equation. Specifically, we examine and compare several coarse spaces which are part of different families, e.g. GenEO (Generalised Eigenvalue Overlap) type coarse spaces and harmonic coarse spaces, that underpin two-level domain decomposition methods. By leveraging spectral information and multiscale approaches, we aim to provide a comprehensive overview of the strengths and weaknesses of these methods. Numerical experiments demonstrate that the effectiveness of these coarse spaces depends on the specific problem and numerical configuration, highlighting the trade-offs between computational cost, robustness, and practical applicability.

math.NA↗

arXiv2025

Discrete coagulation--fragmentation systems in weighted $\ell^1$ spaces

We study an infinite system of ordinary differential equations that models the evolution of coagulating and fragmenting clusters, which we assume to be composed of identical units. Under very mild assumptions on the coefficients we prove existence, uniqueness and positivity of solutions of a corresponding semi-linear Cauchy problem in a weighted $\ell^1$ space. This requires the application of novel results, which we prove for abstract semi-linear Cauchy problems in Banach lattices where the non-linear term is defined only on a dense subspace.

math.FA↗

arXiv2024

Eigenvalue distribution of canonical systems: trace class and sparse spectrum

In this paper we consider two-dimensional canonical systems with discrete spectrum and study their eigenvalue densities. We develop a formula that determines the Stieltjes transform of the eigenvalue counting function up to universal multiplicative constants. An explicit criterion is given for the resolvents of the model operator to belong to a Schatten-von Neumann class with index 0<p<2, thus giving an answer to the long-standing question which canonical systems have trace class resolvents. For canonical systems with two limit circle endpoints we develop an algorithm for determining the growth of the monodromy matrix up to a small error. Moreover, we present examples to illustrate our results, show their sharpness and prove an inverse result giving explicit formulae.

math.SP↗

arXiv2024

Schwarz preconditioner with $H_k$-GenEO coarse space for the indefinite Helmholtz problem

GenEO (`Generalised Eigenvalue problems on the Overlap') is a method from the family of spectral coarse spaces that can efficiently rely on local eigensolves in order to build a robust parallel domain decomposition preconditioner for elliptic PDEs. When used as a preconditioner in a conjugate gradient, this method is extremely efficient in the positive-definite case, yielding an iteration count completely independent of the number of subdomains and heterogeneity. In a previous work this theory was extended to the cased of convection--diffusion--reaction problems, which may be non-self-adjoint and indefinite, and whose discretisations are solved with preconditioned GMRES. The GenEO coarse space was then defined here using a generalised eigenvalue problem based on a self-adjoint and positive definite subproblem. The resulting method, called $Δ$-GenEO becomes robust with respect to the variation of the coefficient of the diffusion term in the operator and depends only very mildly on variations of the other coefficients. However, the iteration number estimates get worse as the non-self-adjointness and indefiniteness of the operator increases, which is often the case for the high frequency Helmholtz problems. In this work, we will improve on this aspect by introducing a new version, called $H_k$-GenEO, which uses a generalised eigenvalue problem based directly on the indefinite operator which will lead to a robust method with respect to the increase in the wave-number. We provide theoretical estimates showing the dependence of the size of the coarse space on the wave-number.

math.NA↗

arXiv2024

Can Symmetric Positive Definite (SPD) coarse spaces perform well for indefinite Helmholtz problems?

Wave propagation problems governed by the Helmholtz equation remain among the most challenging in scientific computing, due to their indefinite nature. Domain decomposition methods with spectral coarse spaces have emerged as some of the most effective preconditioners, yet their theoretical guarantees often lag behind practical performance. In this work, we introduce and analyse the $\Delta_k$-GenEO coarse space within the two-level additive Schwarz preconditioners for heterogeneous Helmholtz problems. This is an adaptation of the $\Delta$-GenEO coarse space. Our results sharpen the $k$-explicit conditions for GMRES convergence, reducing the restrictions on the subdomain size and eigenvalue threshold. This narrows the long-standing gap between pessimistic theory and empirical evidence, and reveals why GenEO spaces based on SPD (symmetric positive definite) eigenvalue problems remain surprisingly effective despite their apparent limitations. Numerical experiments confirm the theory, demonstrating scalability, robustness to heterogeneity for low to moderate frequencies (while experiencing limitations in the high frequency cases), and significantly milder coarse-space growth than conservative estimates predict.

math.NA↗

arXiv2023

Karamata's theorem for regularised Cauchy transforms

We prove Abelian and Tauberian theorems for regularised Cauchy transforms of positive Borel measures on the real line whose distribution functions grow at most polynomially at infinity. In particular, we relate the asymptotics of the distribution functions to the asymptotics of the regularised Cauchy transform.

math.CV↗

arXiv2023

The Radial Hedgehog Solution in the Landau-de Gennes Theory: Effects of the Bulk Potentials

We study equilibrium configurations in spherical droplets of nematic liquid crystal with strong radial anchoring, within the Landau-de Gennes theory with a sixth-order bulk potential. The sixth-order potential predicts a bulk biaxial phase for sufficiently low temperatures, which the conventional fourth-order potential cannot predict. We prove the existence of a radial hedgehog solution, which is a uniaxial solution with a single isotropic point defect at the droplet centre, for all temperatures and droplet sizes, and prove that there is a unique radial hedgehog solution for moderately low temperatures, but not deep in the nematic phase. We numerically compute critical points of the Landau-de Gennes free energy with the sixth order bulk potential, with rotational and mirror symmetry, and find at least two competing stable critical points: the biaxial torus and split core solutions, which have biaxial regions around the centre, for low temperatures. The size of the biaxial regions increases with decreasing temperature. We also compare the properties of the radial hedgehog solution with the fourth-order and sixth-order potentials respectively, in terms of the Morse indices as a function of the temperature and droplet radius; the role of the radial hedgehog solution as a transition state in switching processes; and compare the bifurcation plots with temperature, with the fourth- and sixth-order potentials. Overall, the sixth-order potential has a stabilising effect on biaxial critical points and a de-stabilising effect on uniaxial critical points and we discover an altogether novel bulk biaxial critical point of the Landau-de Gennes energy with the sixth-order potential, for which the bulk biaxiality is driven by the sixth-order potential.

cond-mat.soft↗

arXiv2022

Discrete fragmentation equations with time-dependent coefficients

We examine an infinite, linear system of ordinary differential equations that models the evolution of fragmenting clusters, where each cluster is assumed to be composed of identical units. In contrast to previous investigations into such discrete-size fragmentation models, we allow the fragmentation coefficients to vary with time. By formulating the initial-value problem for the system as a non-autonomous abstract Cauchy problem, posed in an appropriately weighted $\ell^1$ space, and then applying results from the theory of evolution families, we prove the existence and uniqueness of physically relevant, classical solutions for suitably constrained coefficients.

math.FA↗

arXiv2022

A GPU-specialized Inference Parameter Server for Large-Scale Deep Recommendation Models

Recommendation systems are of crucial importance for a variety of modern apps and web services, such as news feeds, social networks, e-commerce, search, etc. To achieve peak prediction accuracy, modern recommendation models combine deep learning with terabyte-scale embedding tables to obtain a fine-grained representation of the underlying data. Traditional inference serving architectures require deploying the whole model to standalone servers, which is infeasible at such massive scale. In this paper, we provide insights into the intriguing and challenging inference domain of online recommendation systems. We propose the HugeCTR Hierarchical Parameter Server (HPS), an industry-leading distributed recommendation inference framework, that combines a high-performance GPU embedding cache with an hierarchical storage architecture, to realize low-latency retrieval of embeddings for online model inference tasks. Among other things, HPS features (1) a redundant hierarchical storage system, (2) a novel high-bandwidth cache to accelerate parallel embedding lookup on NVIDIA GPUs, (3) online training support and (4) light-weight APIs for easy integration into existing large-scale recommendation workflows. To demonstrate its capabilities, we conduct extensive studies using both synthetically engineered and public datasets. We show that our HPS can dramatically reduce end-to-end inference latency, achieving 5~62x speedup (depending on the batch size) over CPU baseline implementations for popular recommendation models. Through multi-GPU concurrent deployment, the HPS can also greatly increase the inference QPS.

cs.IR↗

arXiv2022

Merlin HugeCTR: GPU-accelerated Recommender System Training and Inference

In this talk, we introduce Merlin HugeCTR. Merlin HugeCTR is an open source, GPU-accelerated integration framework for click-through rate estimation. It optimizes both training and inference, whilst enabling model training at scale with model-parallel embeddings and data-parallel neural networks. In particular, Merlin HugeCTR combines a high-performance GPU embedding cache with an hierarchical storage architecture, to realize low-latency retrieval of embeddings for online model inference tasks. In the MLPerf v1.0 DLRM model training benchmark, Merlin HugeCTR achieves a speedup of up to 24.6x on a single DGX A100 (8x A100) over PyTorch on 4x4-socket CPU nodes (4x4x28 cores). Merlin HugeCTR can also take advantage of multi-node environments to accelerate training even further. Since late 2021, Merlin HugeCTR additionally features a hierarchical parameter server (HPS) and supports deployment via the NVIDIA Triton server framework, to leverage the computational capabilities of GPUs for high-speed recommendation model inference. Using this HPS, Merlin HugeCTR users can achieve a 5~62x speedup (batch size dependent) for popular recommendation models over CPU baseline implementations, and dramatically reduce their end-to-end inference latency.

cs.DC↗

arXiv2022

Canonical systems whose Weyl coefficients have regularly varying asymptotics

For a two-dimensional canonical system $y'(t)=zJH(t)y(t)$ on an interval $(0,L)$ with $0<L\le\infty$ whose Hamiltonian $H$ is a.e.\ positive semidefinite, denote by $q_H$ its Weyl coefficient. De~Branges' inverse spectral theorem states that the assignment $H\mapsto q_H$ is a bijection between trace-normalised Hamiltonians and Nevanlinna functions. We prove that $q_H$ has an asymptotics towards $i\infty$ whose leading term is some (complex) multiple of a regularly varying function if and only if the primitive $M$ of $H$ is regularly or rapidly varying at $0$ and its off-diagonal entries do not oscillate too much. The leading term in the asymptotics of $q_H$ towards $i\infty$ is related to the behaviour of $M$ towards $0$ by explicit formulae. The speed of growth in absolute value depends only on the diagonal entries of $M$, while the argument of the leading coefficient corresponds to the relative size of the off-diagonal entries. Translated to the spectral measure $\mu_H$ and the Hamiltonian $H$, this means that the diagonal of $H$ determines the growth of the symmetrised distribution function of $\mu_H$, and the relative size and sign distribution of its off-diagonal is a measure for the asymmetry of $\mu_H$. The results are applied to Sturm--Liouville equations, Krein strings and generalised indefinite strings to prove similar characterisations for the asymptotics of the corresponding Weyl coefficients.

math.SP↗

arXiv2021

Canonical systems whose Weyl coefficients have dominating real part

For a two-dimensional canonical system $y'(t)=zJH(t)y(t)$ on the half-line $(0,\infty)$ whose Hamiltonian $H$ is a.e. positive semi-definite, denote by $q_H$ its Weyl coefficient. De Branges' inverse spectral theorem states that the assignment $H\mapsto q_H$ is a bijection between Hamiltonians (suitably normalised) and Nevanlinna functions. The main result of the paper is a criterion when the singular integral of the spectral measure, i.e. Re $q_H(iy)$, dominates its Poisson integral Im $q_H(iy)$ for $y\to+\infty$. Two equivalent conditions characterising this situation are provided. The first one is analytic in nature, very simple, and explicit in terms of the primitive $M$ of $H$. It merely depends on the relative size of the off-diagonal entries of $M$ compared with the diagonal entries. The second condition is of geometric nature and technically more complicated, but explicit in terms of $H$ itself. It involves the relative size of the off-diagonal entries of $H$, a measurement for oscillations of the diagonal of $H$, and a condition on the speed and smoothness of the rotation of $H$.

math.SP↗

arXiv2021

Estimates for the Weyl coefficient of a two-dimensional canonical system

For a two-dimensional canonical system $y'(t)=zJH(t)y(t)$ on some interval $(a,b)$ whose Hamiltonian $H$ is a.e. positive semi-definite and which is regular at $a$ and in the limit point case at $b$, denote by $q_H$ its Weyl coefficient. De Branges' inverse spectral theorem states that the assignment $H\mapsto q_H$ is a bijection between Hamiltonians (suitably normalised) and Nevanlinna functions. We give upper and lower bounds for $|q_H(z)|$ and $\operatorname{Im} q_H(z)$ when $z$ tends to $i\infty$ non-tangentially. These bounds depend on the Hamiltonian $H$ near the left endpoint $a$ and determine $|q_H(z)|$ up to universal multiplicative constants. We obtain that the growth of $|q_H(z)|$ is independent of the off-diagonal entries of $H$ and depends monotonically on the diagonal entries in a natural way. The imaginary part is, in general, not fully determined by our bounds (in forthcoming work we shall prove that for "most" Hamiltonians also $\operatorname{Im} q_H(z)$ is fully determined). We translate the asymptotic behaviour of $q_H$ to the behaviour of the spectral measure $\mu_H$ of $H$ by means of Abelian--Tauberian results and obtain conditions for membership of growth classes defined by weighted integrability condition (Kac classes) or by boundedness of tails at $\pm\infty$ w.r.t. a weight function. Moreover, we apply our results to Krein strings and Sturm--Liouville equations.

math-ph↗

arXiv2020

Distributed Training of Deep Learning Models: A Taxonomic Perspective

Distributed deep learning systems (DDLS) train deep neural network models by utilizing the distributed resources of a cluster. Developers of DDLS are required to make many decisions to process their particular workloads in their chosen environment efficiently. The advent of GPU-based deep learning, the ever-increasing size of datasets and deep neural network models, in combination with the bandwidth constraints that exist in cluster environments require developers of DDLS to be innovative in order to train high quality models quickly. Comparing DDLS side-by-side is difficult due to their extensive feature lists and architectural deviations. We aim to shine some light on the fundamental principles that are at work when training deep neural networks in a cluster of independent machines by analyzing the general properties associated with training deep learning models and how such workloads can be distributed in a cluster to achieve collaborative model training. Thereby we provide an overview of the different techniques that are used by contemporary DDLS and discuss their influence and implications on the training process. To conceptualize and compare DDLS, we group different techniques into categories, thus establishing a taxonomy of distributed deep learning systems.

cs.DC↗