arXiv subjects

Oana Balmau

Publications and source records attributed to Oana Balmau.

9 recordsLinked to original sources

arXiv2026

DeltaServe: Host-Agnostic Co-Serving of Inference and Fine-Tuning for LLMs

LLM serving systems are provisioned for peak load to meet strict latency targets, leaving substantial GPU compute idle whenever traffic falls below peak. We present DeltaServe, a host-agnostic co-serving design that converts this idle inference capacity into LoRA fine-tuning throughput while preserving inference service-level objectives (SLOs). DeltaServe integrates with existing inference engines through a compact hook interface that requires only multi-LoRA batching support. It exploits the shared execution structure of inference prefill and LoRA fine-tuning forward passes, and uses an SLO-aware scheduler to admit and execute fine-tuning only when sufficient inference headroom is available. The scheduler is driven by a CUDA-graph-aware latency model calibrated offline and refined online. We integrate DeltaServe with vLLM, SGLang, and S-LoRA. On a production trace from Company X, DeltaServe on vLLM delivers 2.9x higher fine-tuning throughput than LLMStation at 100% inference SLO compliance, versus 85% for LLMStation. It also achieves 39% higher fine-tuning throughput than a baseline running vLLM+torchtune, using no additional hardware and maintaining full SLO compliance.

cs.DC↗

arXiv2026

FOLD: Fuzzy Online Deduplication for Very Large Evolving Datasets via Approximate Nearest Neighbor Search

Fuzzy deduplication is key to constructing large language model training corpora. However, classic Locality-Sensitive Hashing (LSH) pipelines scale poorly as corpora grow and are ill-suited to continuous ingestion. The main issue is that each new document batch must be checked against the admitted corpus before insertion. As the corpus grows, the LSH buckets grow: each query can hit several large buckets and must scan the returned candidates. To solve this problem, we present RAD (Retrieval-Augmented Deduplication), an online fuzzy deduplication system that delivers both high recall and throughput for evolving datasets. RAD maintains an incrementally updated HNSW index over admitted documents, retrieving a small, high-quality candidate neighborhood for each incoming document instead of repeatedly re-scanning the accumulated corpus. RAD is the first online fuzzy deduplication system to use HNSW, leading to stable throughput as datasets grow. However, it is not easy to maintain high recall when using HNSW-style indexes. The core issue is the distance metric between graph nodes. Jaccard similarity, the metric used for fuzzy deduplication, yields low recall when applied out-of-the-box with an HNSW index. It leads to distance score crowding, making graph traversal unreliable within a bounded number of steps. RAD addresses this with a bitmap representation that provides a more discriminative, Jaccard-aligned signal during HNSW search. Across four LLM-scale datasets (LM1B, C4, RealNews, and Common Crawl), RAD preserves the scaling trajectory needed for online fuzzy deduplication: at 30M documents, it maintains 0.94-0.97 recall relative to state-of-the-art LSH solutions, and delivers up to an 8x throughput increase.

cs.DC↗

arXiv2026

A Pragmatic Approach to Learned Indexing in RocksDB: Targeted Optimizations with Minimal System Modification

Learned indexes have emerged as a promising alternative to traditional index structures, offering higher throughput and lower memory usage by approximating the cumulative key distribution function with lightweight models. Despite these benefits, adoption in production systems remains limited, partly because learned indexes that support concurrency and persistence as effectively as, e.g., the B+-Tree, do not yet exist, while many research prototypes introduce substantial complexity. In this paper, we investigate whether off-the-shelf learned indexes can be integrated into a production database with minimal storage-engine redesign. Using RocksDB as a case study, we exploit its separation between in-memory Memtables and immutable on-disk files to deploy specialized indexes at each level. We show that directly applying existing learned indexes is insufficient under write-heavy workloads because frequent Memtable replacement prevents models from fully adapting. To address this, we introduce a reuse mechanism that preserves structural knowledge across Memtable instances. At the storage level, we replace RocksDB's disk index with a learned index without modifying the storage layer or read path. We further adapt a read-only learned index to be block-aware, enabling worst-case single-I/O lookups. We implement these techniques in MountDB, an extension of RocksDB. Experiments on large-scale workloads with diverse data distributions and access patterns show up to 1.5X higher write throughput and 2.1X higher read throughput than state-of-the-art systems, demonstrating that established learned indexes can be integrated into production systems with minimal overhead and substantial performance benefits.

cs.DB↗

arXiv2025

MinatoLoader: Accelerating Machine Learning Training Through Efficient Data Preprocessing

Data loaders are used by Machine Learning (ML) frameworks like PyTorch and TensorFlow to apply transformations to data before feeding it into the accelerator. This operation is called data preprocessing. Data preprocessing plays an important role in the ML training workflow because if it is inefficiently pipelined with the training, it can yield high GPU idleness, resulting in important training delays. Unfortunately, existing data loaders turn out to waste GPU resources, with $76\%$ GPU idleness when using the PyTorch data loader, for example. One key source of inefficiency is the variability in preprocessing time across samples within the same dataset. Existing data loaders are oblivious to this variability, and they construct batches without any consideration of slow or fast samples. In this case, the entire batch is delayed by a single slow sample, stalling the training pipeline and resulting in head-of-line blocking. To address these inefficiencies, we present MinatoLoader, a general-purpose data loader for PyTorch that accelerates training and improves GPU utilization. MinatoLoader is designed for a single-server setup, containing multiple GPUs. It continuously prepares data in the background and actively constructs batches by prioritizing fast-to-preprocess samples, while slower samples are processed in parallel. We evaluate MinatoLoader on servers with V100 and A100 GPUs. On a machine with four A100 GPUs, MinatoLoader improves the training time of a wide range of workloads by up to $7.5\times$ ($3.6\times$ on average) over PyTorch DataLoader and Pecan, and up to $3\times$ ($2.2\times$ on average) over DALI. It also increases average GPU utilization from 46.4\% with PyTorch to 90.45\%, while preserving model accuracy and enabling faster convergence.

cs.DC↗

arXiv2025

MaLV-OS: Rethinking the Operating System Architecture for Machine Learning in Virtualized Clouds

A large body of research has employed Machine Learning (ML) models to develop learned operating systems (OSes) and kernels. The latter dynamically adapts to the job load and dynamically adjusts resources (CPU, IO, memory, network bandwidth) allocation to respond to the actual user demand. What this work has in common is that it utilizes ML to improve kernel decisions. To this day, and to the best of our knowledge, no work has taken the opposite direction, i.e., using OS to improve ML. While some work proposes applying system-level optimizations to ML algorithms, they do not tailor the OS to adapt to the ML context. To address this limitation, we take an orthogonal approach in this paper by leveraging the OS to enhance the performance of ML models and algorithms. We explore the path towards an ML-specialized OS, MaLV-OS. MaLV-OS rethinks the OS architecture to make it specifically tailored to ML workloads, especially in virtualized clouds, which are now widely used to run ML applications. MaLV-OS envisioned architecture includes (1) a micro-kernel, Micro-LAKE, which allows kernel space applications to use the GPU, and (2) an MLaaS (ML as a Service) subsystem that gathers ML models to help Micro-LAKE with memory management and CPU scheduling. MaLV-OS architecture also offloads system-sensitive parts of the models to the OS, to lighten the model complexity and programming, and speed up its execution. Finally, MaLV-OS integrates an open-source GPU virtualization software, merged directly into the hypervisor. For more flexibility, MaLV-OS vision is to enable the virtual machine to dynamically select MLaaS policies that can improve the performance of the model the user is running. Because MLaaS is designed as loadable kernel modules, the MaLV-OS architecture enables the dynamic addition of new capabilities to the MLaaS subsystem.

cs.OS↗

arXiv2025

HarMoEny: Efficient Multi-GPU Inference of MoE Models

Mixture-of-Experts (MoE) models offer computational efficiency during inference by activating only a subset of specialized experts for a given input. This enables efficient model scaling on multi-GPU systems that use expert parallelism without compromising performance. However, load imbalance among experts and GPUs introduces waiting times, which can significantly increase inference latency. To address this challenge, we propose HarMoEny, a novel solution to address MoE load imbalance through two simple techniques: (i) dynamic token redistribution to underutilized GPUs and (ii) asynchronous prefetching of experts from the system to GPU memory. These techniques achieve a near-perfect load balance among experts and GPUs and mitigate delays caused by overloaded GPUs. We implement HarMoEny and compare its latency and throughput with four MoE baselines using real-world and synthetic datasets. Under heavy load imbalance, HarMoEny increases throughput by 37%-70% and reduces time-to-first-token by 34%-41%, compared to the next-best baseline. Moreover, our ablation study demonstrates that HarMoEny's scheduling policy reduces the GPU idling time by up to 84% compared to the baseline policies.

cs.DC↗

arXiv2025

Keigo: Co-designing Log-Structured Merge Key-Value Stores with a Non-Volatile, Concurrency-aware Storage Hierarchy (Extended Version)

We present Keigo, a concurrency- and workload-aware storage middleware that enhances the performance of log-structured merge key-value stores (LSM KVS) when they are deployed on a hierarchy of storage devices. The key observation behind Keigo is that there is no one-size-fits-all placement of data across the storage hierarchy that optimizes for all workloads. Hence, to leverage the benefits of combining different storage devices, Keigo places files across different devices based on their parallelism, I/O bandwidth, and capacity. We introduce three techniques - concurrency-aware data placement, persistent read-only caching, and context-based I/O differentiation. Keigo is portable across different LSMs, is adaptable to dynamic workloads, and does not require extensive profiling. Our system enables established production KVS such as RocksDB, LevelDB, and Speedb to benefit from heterogeneous storage setups. We evaluate Keigo using synthetic and realistic workloads, showing that it improves the throughput of production-grade LSMs up to 4x for write- and 18x for read-heavy workloads when compared to general-purpose storage systems and specialized LSM KVS.

cs.DC↗

arXiv2025

Practical Federated Learning without a Server

Federated Learning (FL) enables end-user devices to collaboratively train ML models without sharing raw data, thereby preserving data privacy. In FL, a central parameter server coordinates the learning process by iteratively aggregating the trained models received from clients. Yet, deploying a central server is not always feasible due to hardware unavailability, infrastructure constraints, or operational costs. We present Plexus, a fully decentralized FL system for large networks that operates without the drawbacks originating from having a central server. Plexus distributes the responsibilities of model aggregation and sampling among participating nodes while avoiding network-wide coordination. We evaluate Plexus using realistic traces for compute speed, pairwise latency and network capacity. Our experiments on three common learning tasks and with up to 1000 nodes empirically show that Plexus reduces time-to-accuracy by 1.4-1.6x, communication volume by 15.8-292x and training resources needed for convergence by 30.5-77.9x compared to conventional decentralized learning algorithms.

cs.DC↗

arXiv2018

Limiting the Spread of Fake News on Social Media Platforms by Evaluating Users' Trustworthiness

Today's social media platforms enable to spread both authentic and fake news very quickly. Some approaches have been proposed to automatically detect such "fake" news based on their content, but it is difficult to agree on universal criteria of authenticity (which can be bypassed by adversaries once known). Besides, it is obviously impossible to have each news item checked by a human. In this paper, we a mechanism to limit the spread of fake news which is not based on content. It can be implemented as a plugin on a social media platform. The principle is as follows: a team of fact-checkers reviews a small number of news items (the most popular ones), which enables to have an estimation of each user's inclination to share fake news items. Then, using a Bayesian approach, we estimate the trustworthiness of future news items, and treat accordingly those of them that pass a certain "untrustworthiness" threshold. We then evaluate the effectiveness and overhead of this technique on a large Twitter graph. We show that having a few thousands users exposed to one given news item enables to reach a very precise estimation of its reliability. We thus identify more than 99% of fake news items with no false positives. The performance impact is very small: the induced overhead on the 90th percentile latency is less than 3%, and less than 8% on the throughput of user operations.

cs.SI↗