Oana Balmau
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Oana Balmau has a total of 29 co-authors. Here are the top co-authors by number of shared publications:
- Stella Bitchebe
McGill University, Montreal, Quebec, Canada
3 shared publications - Alexandre da Silva Veith
Nokia Bell Labs, Antwerp, Belgium
2 shared publications - Anne-Marie Kermarrec
EPFL, Lausanne, Switzerland
2 shared publications - Eyal de Lara
University of Toronto, Toronto, Canada
2 shared publications - Jiaxuan Chen
McGill University, Montreal, Canada
2 shared publications - Martijn de Vos
EPFL, Lausanne, Switzerland
2 shared publications
Latest Publications
- Open Access
- 20 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 275-280
Wagg: Cost-aware Aggregation of Windowing Operators in Stream Processing
- 20 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 275-280
Open AccessWagg: Cost-aware Aggregation of Windowing Operators in Stream Processing
- Open Access
- 27 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 466-472
The Cost of Expertise: Understanding MoE Decode Performance
- 27 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 466-472
Open AccessThe Cost of Expertise: Understanding MoE Decode Performance
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2037-2053
MinatoLoader: Accelerating Machine Learning Training Through Efficient Data Preprocessing
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 2037-2053
Open AccessMinatoLoader: Accelerating Machine Learning Training Through Efficient Data Preprocessing
- Open Access
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 20-25
PostLearn: Towards A Learned Index For PostgreSQL
Recent advances in machine learning based data structures, such as learned indexes, show that models can be trained to approximate the cumulative distribution of keys, thereby predicting key positions more efficiently than traditional indexes. While research prototypes have shown significant promise, so far there has been no comprehensive integration of learned indexes into a relational database system. In this paper, we explore the feasibility of learned indexes for relational systems by introducing PostLearn, an integration of the ALEX+ learned index as a native PostgreSQL index access method. We detail the design and the implementation challenges of embedding a model-based index within PostgreSQL's existing infrastructure and explore its potential performance benefits. While PostLearn can achieve up to 1.5x speedups for point lookups and small range scans compared to the built-in B+-Tree under selected workloads, end-to-end performance benefits are considerably lower than when tested in isolation.
- 26 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 20-25
Open AccessPostLearn: Towards A Learned Index For PostgreSQL
Recent advances in machine learning based data structures, such as learned indexes, show that models can be trained to approximate the cumulative distribution of keys, thereby predicting key positions more efficiently than traditional indexes. While research prototypes have shown significant promise, so far there has been no comprehensive integration of learned indexes into a relational database system. In this paper, we explore the feasibility of learned indexes for relational systems by introducing PostLearn, an integration of the ALEX+ learned index as a native PostgreSQL index access method. We detail the design and the implementation challenges of embedding a model-based index within PostgreSQL's existing infrastructure and explore its potential performance benefits. While PostLearn can achieve up to 1.5x speedups for point lookups and small range scans compared to the built-in B+-Tree under selected workloads, end-to-end performance benefits are considerably lower than when tested in isolation.
- Open Access
- 30 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 192-199
Accelerating MoE Model Inference with Expert Sharding
- 30 March 2025
- Published by Association for Computing Machinery (ACM)
- p. 192-199
Open AccessAccelerating MoE Model Inference with Expert Sharding
- Open Access
- 2 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 417-430
vPIM: Processing-in-Memory Virtualization
Data movement is the leading cause of performance degradation and energy consumption in modern data centers. Processing inmemory (PIM) is an architecture that addresses data movement by bringing computation inside the memory chips. This paper is the first to study the virtualization of PIM devices by designing and implementing vPIM, an open-source UPMEM-based virtualization system for the cloud. Our vPIM design considers four requirements: Compatibility such that no hardware and no hypervisor changes are needed; Multiplexing and isolation for a higher utilization ratio; Utilizability and transparency such that applications written for PIM can be efficiently run out-of-the-box, leading to rapid adoption; Minimalization of virtualization performance overhead. We prototype vPIM in Firecracker, expanding the virtio standard. Our experimental evaluation uses 16 applications provided by PrIM, a recent PIM benchmark suite. The virtualization overhead is between 1.01× and 2.07× for untouched PrIM applications. To keep overhead low, vPIM introduces several optimizations: zero-copy from guest OS to Firecracker, efficient virtio queues management, efficient Guest Physical Address to Host Virtual Address translation, parallel processing on multiple ranks, automatic data batching and prefetching, and the reimplementation of some specific functionalities in C instead of Rust. We hope this work will lay the foundation for future research on PIM for cloud computing.
- 2 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 417-430
Open AccessvPIM: Processing-in-Memory Virtualization
Data movement is the leading cause of performance degradation and energy consumption in modern data centers. Processing inmemory (PIM) is an architecture that addresses data movement by bringing computation inside the memory chips. This paper is the first to study the virtualization of PIM devices by designing and implementing vPIM, an open-source UPMEM-based virtualization system for the cloud. Our vPIM design considers four requirements: Compatibility such that no hardware and no hypervisor changes are needed; Multiplexing and isolation for a higher utilization ratio; Utilizability and transparency such that applications written for PIM can be efficiently run out-of-the-box, leading to rapid adoption; Minimalization of virtualization performance overhead. We prototype vPIM in Firecracker, expanding the virtio standard. Our experimental evaluation uses 16 applications provided by PrIM, a recent PIM benchmark suite. The virtualization overhead is between 1.01× and 2.07× for untouched PrIM applications. To keep overhead low, vPIM introduces several optimizations: zero-copy from guest OS to Firecracker, efficient virtio queues management, efficient Guest Physical Address to Host Virtual Address translation, parallel processing on multiple ranks, automatic data batching and prefetching, and the reimplementation of some specific functionalities in C instead of Rust. We hope this work will lay the foundation for future research on PIM for cloud computing.
- Open Access
- 22 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 65-72
SpeedyLoader
Data preprocessing consisting of tasks like sample resizing, cropping, and filtering, is a crucial step in machine learning (ML) workflows. Even though the preprocessing step is largely ignored by work that focuses on optimizing training algorithms, in practice for many workloads preprocessing and training are pipelined. Popular ML frameworks like PyTorch use data loaders to feed data into model training. If the pipeline between preprocessing and training is not done carefully, it can cause significant waiting times on the GPU side. To address this limitation, we introduce SpeedyLoader, a system that overlaps preprocessing and training by leveraging asynchronous data preprocessing and avoiding head-of-line blocking. SpeedyLoader incorporates dedicated data loading threads, which organize preprocessed samples into queues based on their predicted processing times. Concurrently, GPUs fetch samples from these queues, ensuring training is not impeded by preprocessing completion. Compared to the default PyTorch DataLoader, SpeedyLoader reduces training time by up to 30% and increases GPU usage by 4.3×, all while maintaining a consistent evaluation accuracy of 91%.
- 22 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 65-72
Open AccessSpeedyLoader
Data preprocessing consisting of tasks like sample resizing, cropping, and filtering, is a crucial step in machine learning (ML) workflows. Even though the preprocessing step is largely ignored by work that focuses on optimizing training algorithms, in practice for many workloads preprocessing and training are pipelined. Popular ML frameworks like PyTorch use data loaders to feed data into model training. If the pipeline between preprocessing and training is not done carefully, it can cause significant waiting times on the GPU side. To address this limitation, we introduce SpeedyLoader, a system that overlaps preprocessing and training by leveraging asynchronous data preprocessing and avoiding head-of-line blocking. SpeedyLoader incorporates dedicated data loading threads, which organize preprocessed samples into queues based on their predicted processing times. Concurrently, GPUs fetch samples from these queues, ensuring training is not impeded by preprocessing completion. Compared to the default PyTorch DataLoader, SpeedyLoader reduces training time by up to 30% and increases GPU usage by 4.3×, all while maintaining a consistent evaluation accuracy of 91%.
- Open Access
- 22 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 55-60
PathFS
As IoT devices multiply and produce vast volumes of data, there is a heightened demand for instantaneous data processing. However, traditional cloud computing cannot adequately address these demands due to its latency and bandwidth limitations. Edge computing has emerged as a viable alternative with a hierarchical deployment of datacenters. However, this introduces additional layers of infrastructure and management that increase application development complexity. Using a shared file system is an attractive method for enhancing communication between components in an edge computing application. In this paper we introduce PathFS, a shared file system designed for the hierarchical edge-cloud infrastructure. PathFS adopts a treelike structure, with cloud datacenters at the root, edge datacenters as leaves, and a variable number of network datacenters in between. We evaluate PathFS through benchmarks on an emulated hierarchical edge deployment and compare it with NFS and ownCloud. The results show that PathFS offers lower latency than these systems by an order of magnitude, and scales to a larger number of concurrent clients without performance impacts, providing an end-to-end latency reduction of at least 80%.
- 22 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 55-60
Open AccessPathFS
As IoT devices multiply and produce vast volumes of data, there is a heightened demand for instantaneous data processing. However, traditional cloud computing cannot adequately address these demands due to its latency and bandwidth limitations. Edge computing has emerged as a viable alternative with a hierarchical deployment of datacenters. However, this introduces additional layers of infrastructure and management that increase application development complexity. Using a shared file system is an attractive method for enhancing communication between components in an edge computing application. In this paper we introduce PathFS, a shared file system designed for the hierarchical edge-cloud infrastructure. PathFS adopts a treelike structure, with cloud datacenters at the root, edge datacenters as leaves, and a variable number of network datacenters in between. We evaluate PathFS through benchmarks on an emulated hierarchical edge deployment and compare it with NFS and ownCloud. The results show that PathFS offers lower latency than these systems by an order of magnitude, and scales to a larger number of concurrent clients without performance impacts, providing an end-to-end latency reduction of at least 80%.
- Open Access
- 22 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 37-42
Stream Processing with Adaptive Edge-Enhanced Confidential Computing
Stream processing is becoming increasingly significant in various scenarios, including security-sensitive sectors. It benefits from keeping data in memory, which exposes large volumes of data in use, thereby emphasising the need for protection. The recent development of confidential computing makes such protection technologically feasible. However, these new hardware-based protection methods incur performance overhead. Our evaluation shows that replacing legacy VMs with confidential VMs to run streaming applications incurs up to 8.5% overhead on the throughput of the queries we tested in the NEXMark benchmark suite. Pursuing specialised protection for broader attacks, such as attacks at the edge with more physical exposure, can push this overhead further. In this paper, we propose a resource scheduling strategy for stream processing applications tailored to the privacy needs of specific application functions. We implement this system model using Apache Flink, a widely-used stream processing framework, making it aware of the underlying cluster's protection capability and scheduling the application functions across resources with different protections tailored to the privacy requirements of an application and the available deployment environment.
- 22 April 2024
- Published by Association for Computing Machinery (ACM)
- p. 37-42
Open AccessStream Processing with Adaptive Edge-Enhanced Confidential Computing
Stream processing is becoming increasingly significant in various scenarios, including security-sensitive sectors. It benefits from keeping data in memory, which exposes large volumes of data in use, thereby emphasising the need for protection. The recent development of confidential computing makes such protection technologically feasible. However, these new hardware-based protection methods incur performance overhead. Our evaluation shows that replacing legacy VMs with confidential VMs to run streaming applications incurs up to 8.5% overhead on the throughput of the queries we tested in the NEXMark benchmark suite. Pursuing specialised protection for broader attacks, such as attacks at the edge with more physical exposure, can push this overhead further. In this paper, we propose a resource scheduling strategy for stream processing applications tailored to the privacy needs of specific application functions. We implement this system model using Apache Flink, a widely-used stream processing framework, making it aware of the underlying cluster's protection capability and scheduling the application functions across resources with different protections tailored to the privacy requirements of an application and the available deployment environment.
- Open Access
- 8 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 27-34
Is Bare-metal I/O Performance with User-defined Storage Drives Inside VMs Possible?: Benchmarking libvfio-user vs. Common Storage Virtualization Configurations
It is a critical concern for cloud computing providers to identify and adopt optimal virtual machine (VM) I/O storage paths. In Linux, these involve NVMe hardware, PCIe passthrough with vfio, and userspace drivers (e.g. SPDK). One project tying together these host-side improvements with the guests is vfio-user, a library and protocol for inter-process communication of I/O operations over sockets. QEMU/KVM VMs using vfio-user and SPDK are theoretically expected to attain low latencies due to mostly userspace processing, but this VM datapath configuration is still experimental and performance data is lacking in the literature. This paper presents the setup and results of benchmarks comparing the I/O performance of a VM using vfio-user and SPDK as its storage backend against different datapath configurations. Additionally, we present a bpftrace-based tool we call vfio_user_snoop to analyze vfio-user packets in transit and a layer-by-layer breakdown of I/O latencies using this setup. We observe negligible differences for baremetal, vfio-user, and passthrough in fio benchmarks, at 26.66 ± 0.18μs mean latency in random reads and 27.77 ± 0.77μs in random writes. For RocksDB, on random reads, vfio-user attains 19.75% fewer mean ops/sec vs. bare-metal and 1.66% higher mean ops/sec vs. passthrough; on 50/50 random reads and writes, we see 5.59% and 2.39% fewer mean ops/sec vs. bare-metal and passthrough, respectively.
- 8 May 2023
- Published by Association for Computing Machinery (ACM)
- p. 27-34
Open AccessIs Bare-metal I/O Performance with User-defined Storage Drives Inside VMs Possible?: Benchmarking libvfio-user vs. Common Storage Virtualization Configurations
It is a critical concern for cloud computing providers to identify and adopt optimal virtual machine (VM) I/O storage paths. In Linux, these involve NVMe hardware, PCIe passthrough with vfio, and userspace drivers (e.g. SPDK). One project tying together these host-side improvements with the guests is vfio-user, a library and protocol for inter-process communication of I/O operations over sockets. QEMU/KVM VMs using vfio-user and SPDK are theoretically expected to attain low latencies due to mostly userspace processing, but this VM datapath configuration is still experimental and performance data is lacking in the literature. This paper presents the setup and results of benchmarks comparing the I/O performance of a VM using vfio-user and SPDK as its storage backend against different datapath configurations. Additionally, we present a bpftrace-based tool we call vfio_user_snoop to analyze vfio-user packets in transit and a layer-by-layer breakdown of I/O latencies using this setup. We observe negligible differences for baremetal, vfio-user, and passthrough in fio benchmarks, at 26.66 ± 0.18μs mean latency in random reads and 27.77 ± 0.77μs in random writes. For RocksDB, on random reads, vfio-user attains 19.75% fewer mean ops/sec vs. bare-metal and 1.66% higher mean ops/sec vs. passthrough; on 50/50 random reads and writes, we see 5.59% and 2.39% fewer mean ops/sec vs. bare-metal and passthrough, respectively.