Vito Giovanni Castellana
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Vito Giovanni Castellana has a total of 53 co-authors. Here are the top co-authors by number of shared publications:
- Antonino Tumeo
Pacific Northwest National Laboratory
20 shared publications - Fabrizio Ferrandi
Politecnico di Milano
18 shared publications - Marco Minutoli
Pacific Northwest National Laboratory, Richland, WA, USA
12 shared publications David Haglin
9 shared publicationsJohn Feo
8 shared publicationsAlessandro Morari
7 shared publications
Latest Publications
- Open Access
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 138
Towards Scheduling of Pipelined Dataflow Graphs in MLIR
- 21 February 2026
- Published by Association for Computing Machinery (ACM)
- p. 138
Open AccessTowards Scheduling of Pipelined Dataflow Graphs in MLIR
- 3 February 2026
- Published by Office of Scientific and Technical Information (OSTI)
AMAIS - Advanced Memory to Support AI for Science: Report February 2026
The AMAIS (Advanced Memory to Support AI for Science) program addresses the growing imbalance between computation and memory performance in modern scientific and AI workloads—commonly known as the memory wall. The project introduces an integrated hardware–software co-design approach that leverages fabric-attached memory, near-memory compute, and advanced memory-centric software tools to enable scalable, high-performance execution of data-intensive applications. Key innovations include the Memory Archipelago architecture, the RRAPID software suite for accelerator synthesis and analysis, CXL-based platforms such as Two Sisters and Crete, and OCEAN, an open-source emulation environment that enables hyperscale experimentation with emerging memory fabrics. Through algorithm– architecture co-design, novel HLS methodologies, performance analysis frameworks, distributed runtimes, and new compiler technologies, AMAIS demonstrates how large, heterogeneous memory pools can support next-generation AI, HPC, and converged simulation workflows. This work establishes foundational methods and platforms for exploiting disaggregated memory systems at scale and provides early evidence of the transformative potential of CXL-enabled architectures for scientific computing.
- 3 February 2026
- Published by Office of Scientific and Technical Information (OSTI)
AMAIS - Advanced Memory to Support AI for Science: Report February 2026
The AMAIS (Advanced Memory to Support AI for Science) program addresses the growing imbalance between computation and memory performance in modern scientific and AI workloads—commonly known as the memory wall. The project introduces an integrated hardware–software co-design approach that leverages fabric-attached memory, near-memory compute, and advanced memory-centric software tools to enable scalable, high-performance execution of data-intensive applications. Key innovations include the Memory Archipelago architecture, the RRAPID software suite for accelerator synthesis and analysis, CXL-based platforms such as Two Sisters and Crete, and OCEAN, an open-source emulation environment that enables hyperscale experimentation with emerging memory fabrics. Through algorithm– architecture co-design, novel HLS methodologies, performance analysis frameworks, distributed runtimes, and new compiler technologies, AMAIS demonstrates how large, heterogeneous memory pools can support next-generation AI, HPC, and converged simulation workflows. This work establishes foundational methods and platforms for exploiting disaggregated memory systems at scale and provides early evidence of the transformative potential of CXL-enabled architectures for scientific computing.
- Open Access
- 20 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 1016-1022
A Synthesis Methodology for Intelligent Memory Interfaces in Accelerator Systems
Domain-specific systems improve the performance of specific applications compared to general-purpose processing systems by deploying custom hardware accelerators. These hardware accelerators are generated using high-level synthesis (HLS) tools. The HLS tools enable a comprehensive design space exploration, optimizing the accelerators' compute performance. However, they often ignore the challenges of implementing the accelerators in a system-on-chip, particularly how they access memory. Our work introduces a buffering system design that improves accelerators' memory accesses by intelligently employing burst transactions to prefetch useful data from external memory to on-chip local buffers. Our design is dynamic, parametric, and transparent to the accelerators generated by HLS tools. We derive the buffering system parameters using appropriate compiler-based analysis passes and memory channel latency constraints. The proposed buffering system design results in, on average, 8.8× performance improvements while lowering memory channel utilization by 53.2% for a set of PolyBench kernels.
- 20 January 2025
- Published by Association for Computing Machinery (ACM)
- p. 1016-1022
Open AccessA Synthesis Methodology for Intelligent Memory Interfaces in Accelerator Systems
Domain-specific systems improve the performance of specific applications compared to general-purpose processing systems by deploying custom hardware accelerators. These hardware accelerators are generated using high-level synthesis (HLS) tools. The HLS tools enable a comprehensive design space exploration, optimizing the accelerators' compute performance. However, they often ignore the challenges of implementing the accelerators in a system-on-chip, particularly how they access memory. Our work introduces a buffering system design that improves accelerators' memory accesses by intelligently employing burst transactions to prefetch useful data from external memory to on-chip local buffers. Our design is dynamic, parametric, and transparent to the accelerators generated by HLS tools. We derive the buffering system parameters using appropriate compiler-based analysis passes and memory channel latency constraints. The proposed buffering system design results in, on average, 8.8× performance improvements while lowering memory channel utilization by 53.2% for a set of PolyBench kernels.
- Open Access
- 25 December 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Reconfigurable Technology and Systems
- Vol. 18 (1) , 1-30
SPARTA: High-Level Synthesis of Parallel Multi-Threaded Accelerators
This article presents a methodology for the Synthesis of PARallel multi-Threaded Accelerators (SPARTA) from OpenMP annotated C/C++ specifications. SPARTA extends an open-source HLS tool, enabling the generation of accelerators that provide latency tolerance for irregular memory accesses through multithreading, support fine-grained memory-level parallelism through a hot-potato deflection-based network-on-chip (NoC), support synchronization constructs, and can instantiate memory-side caches. Our approach is based on a custom runtime OpenMP library, providing flexibility and extensibility. Experimental results show high scalability when synthesizing irregular graph kernels. The accelerators generated with our approach are, on average, 2.29 \(\times\) faster than state-of-the-art HLS methodologies.- 25 December 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Reconfigurable Technology and Systems
- Vol. 18 (1) , 1-30
Open AccessSPARTA: High-Level Synthesis of Parallel Multi-Threaded Accelerators
This article presents a methodology for the Synthesis of PARallel multi-Threaded Accelerators (SPARTA) from OpenMP annotated C/C++ specifications. SPARTA extends an open-source HLS tool, enabling the generation of accelerators that provide latency tolerance for irregular memory accesses through multithreading, support fine-grained memory-level parallelism through a hot-potato deflection-based network-on-chip (NoC), support synchronization constructs, and can instantiate memory-side caches. Our approach is based on a custom runtime OpenMP library, providing flexibility and extensibility. Experimental results show high scalability when synthesizing irregular graph kernels. The accelerators generated with our approach are, on average, 2.29 \(\times\) faster than state-of-the-art HLS methodologies. - Open Access
- 27 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Extending High-Level Synthesis with AI/ML Methods
- 27 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-6
Open AccessExtending High-Level Synthesis with AI/ML Methods
- Open Access
- 30 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 210-218
To Cache or not to Cache? Exploring the Design Space of Tunable, HLS-generated Accelerators
- 30 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 210-218
Open AccessTo Cache or not to Cache? Exploring the Design Space of Tunable, HLS-generated Accelerators
- Open Access
- 30 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 186-191
Programming the Future: the Essential Role of System Topology Awareness in Heterogeneous Disaggregated Environments
- 30 September 2024
- Published by Association for Computing Machinery (ACM)
- p. 186-191
Open AccessProgramming the Future: the Essential Role of System Topology Awareness in Heterogeneous Disaggregated Environments
- Open Access
- 17 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-5
Exploring Architectural-Aware Affinity Policies in Modern HPC Runtimes
Modern commodity and High-Performance Computing (HPC) systems are evolving with complex CPU architectures. These architectures now feature higher core and NUMA domain counts and implement features such as hyperthreading. When considering significant differences in hardware configurations, library availability, and hardware-tailored system/software stacks, which could substantially vary from one system to another, performance portability is hard to achieve. Throughout the years, this trend resulted in an increasingly high burden on application developers to fine-tune their workloads for each architecture. This work explores how hardware-dependent aspects such as locality/process/thread affinity affect performance in modern CPU architectures. We focus our study on the Global Memory and Threading (GMT) distributed runtime system as a representative of Partitioned Global Address Space (PGAS) software stacks commonly adopted for productivity. In particular, to appreciate performance implications, we evaluate GMT’s thread affinity policies, and, introduce two new ones which exploit architectural awareness. Finally, we explore alternative NUMA configurations via different process bindings and perform a scalability study on three HPC clusters with varying CPU architectures and NUMA layouts. Our analysis indicates that more complex architectures are more affected by affinity and binding policies and highlights the importance of setting proper runtime configurations to achieve superior performance.
- 17 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-5
Open AccessExploring Architectural-Aware Affinity Policies in Modern HPC Runtimes
Modern commodity and High-Performance Computing (HPC) systems are evolving with complex CPU architectures. These architectures now feature higher core and NUMA domain counts and implement features such as hyperthreading. When considering significant differences in hardware configurations, library availability, and hardware-tailored system/software stacks, which could substantially vary from one system to another, performance portability is hard to achieve. Throughout the years, this trend resulted in an increasingly high burden on application developers to fine-tune their workloads for each architecture. This work explores how hardware-dependent aspects such as locality/process/thread affinity affect performance in modern CPU architectures. We focus our study on the Global Memory and Threading (GMT) distributed runtime system as a representative of Partitioned Global Address Space (PGAS) software stacks commonly adopted for productivity. In particular, to appreciate performance implications, we evaluate GMT’s thread affinity policies, and, introduce two new ones which exploit architectural awareness. Finally, we explore alternative NUMA configurations via different process bindings and perform a scalability study on three HPC clusters with varying CPU architectures and NUMA layouts. Our analysis indicates that more complex architectures are more affected by affinity and binding policies and highlights the importance of setting proper runtime configurations to achieve superior performance.
- 1 May 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 442-451
SHAD: The Scalable High-Performance Algorithms and Data-Structures Library
The unprecedented amount of data that needs to be processed in emerging data analytics applications poses novel challenges to industry and academia. Scalability and high performance become more than a desirable feature because, due to the scale and the nature of the problems, they draw the line between what is achievable and what is unfeasible. In this paper, we propose SHAD, the Scalable High-performance Algorithms and Data-structures library. SHAD adopts a modular design that confines low level details and promotes reuse. SHAD's core is built on an Abstract Runtime Interface which enhances portability and identifies the minimal set of features of the underlying system required by the framework. The core library includes common data-structures such as: Array, Vector, Map and Set. These are designed to accommodate significant amount of data which can be accessed in massively parallel environments, and used as building blocks for SHAD extensions, i.e. higher level software libraries. We have validated and evaluated our design with a performance and scalability study of the core components of the library. We have validated the design flexibility by proposing a Graph Library as an example of SHAD extension, which implements two different graph data-structures; we evaluate their performance with a set of graph applications. Experimental results show that the approach is promising in terms of both performance and scalability. On a distributed system with 320 cores, SHAD Arrays are able to sustain a throughput of 65 billion operations per second, while SHAD Maps sustain 1 billion of operations per second. Algorithms implemented using the Graph Library exhibit performance and scalability comparable to a custom solution, but with smaller development effort.
- 1 May 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 442-451
SHAD: The Scalable High-Performance Algorithms and Data-Structures Library
The unprecedented amount of data that needs to be processed in emerging data analytics applications poses novel challenges to industry and academia. Scalability and high performance become more than a desirable feature because, due to the scale and the nature of the problems, they draw the line between what is achievable and what is unfeasible. In this paper, we propose SHAD, the Scalable High-performance Algorithms and Data-structures library. SHAD adopts a modular design that confines low level details and promotes reuse. SHAD's core is built on an Abstract Runtime Interface which enhances portability and identifies the minimal set of features of the underlying system required by the framework. The core library includes common data-structures such as: Array, Vector, Map and Set. These are designed to accommodate significant amount of data which can be accessed in massively parallel environments, and used as building blocks for SHAD extensions, i.e. higher level software libraries. We have validated and evaluated our design with a performance and scalability study of the core components of the library. We have validated the design flexibility by proposing a Graph Library as an example of SHAD extension, which implements two different graph data-structures; we evaluate their performance with a set of graph applications. Experimental results show that the approach is promising in terms of both performance and scalability. On a distributed system with 320 cores, SHAD Arrays are able to sustain a throughput of 65 billion operations per second, while SHAD Maps sustain 1 billion of operations per second. Algorithms implemented using the Graph Library exhibit performance and scalability comparable to a custom solution, but with smaller development effort.
- 19 May 2017
- Published by Taylor & Francis
- p. 279-296
Considerations on the Use of Custom Accelerators for Big Data Analytics
Accelerators, including graphic processing units (GPUs) for general-purpose computation, many-core designs with wide vector units (e.g., Intel Phi), have become a common component of many high-performance clusters. The appearance of more stable and reliable tools that can automatically convert code written in high-level specifications with annotations (such as C or C++) to hardware description languages (high-level synthesis—HLS) is also setting the stage for a broader use of reconfigurable devices (e.g., field programmable gate arrays—FPGAs) in high-performance system for the implementation of custom accelerators, helped by the fact that new processors include advanced cache-coherent interconnects for these components. In this chapter, we briefly survey the status of the use of accelerators in high-performance systems targeted at big data analytics applications. Although the recent progress in the use of accelerators for this class of applications has been significant, we argue that, differently from scientific simulations, there are still gaps to close. This is particularly true for the “irregular” behaviors exhibited by emerging no-SQL and graph databases. We focus our attention on the limits of HLS tools for data analytics and graph methods, and discuss a new architectural template that better fits the requirement of this class of applications. We validate the new architectural templates by modifying the Graph Engine for Multithreaded System (GEMS) framework to support accelerators generated with such a methodology, and by testing it 280with queries coming from the Lehigh University Benchmark (LUBM). The architectural template enables better supporting the task- and memory-level parallelism present in graph methods by supporting a new control model and an enhanced memory interface. We show that our solution allows generating parallel accelerators, providing speed ups with respect to conventional HLS flows. We finally draw conclusions and present a perspective on the use of reconfigurable devices and design automation tools for data analytics.
- 19 May 2017
- Published by Taylor & Francis
- p. 279-296
Considerations on the Use of Custom Accelerators for Big Data Analytics
Accelerators, including graphic processing units (GPUs) for general-purpose computation, many-core designs with wide vector units (e.g., Intel Phi), have become a common component of many high-performance clusters. The appearance of more stable and reliable tools that can automatically convert code written in high-level specifications with annotations (such as C or C++) to hardware description languages (high-level synthesis—HLS) is also setting the stage for a broader use of reconfigurable devices (e.g., field programmable gate arrays—FPGAs) in high-performance system for the implementation of custom accelerators, helped by the fact that new processors include advanced cache-coherent interconnects for these components. In this chapter, we briefly survey the status of the use of accelerators in high-performance systems targeted at big data analytics applications. Although the recent progress in the use of accelerators for this class of applications has been significant, we argue that, differently from scientific simulations, there are still gaps to close. This is particularly true for the “irregular” behaviors exhibited by emerging no-SQL and graph databases. We focus our attention on the limits of HLS tools for data analytics and graph methods, and discuss a new architectural template that better fits the requirement of this class of applications. We validate the new architectural templates by modifying the Graph Engine for Multithreaded System (GEMS) framework to support accelerators generated with such a methodology, and by testing it 280with queries coming from the Lehigh University Benchmark (LUBM). The architectural template enables better supporting the task- and memory-level parallelism present in graph methods by supporting a new control model and an enhanced memory interface. We show that our solution allows generating parallel accelerators, providing speed ups with respect to conventional HLS flows. We finally draw conclusions and present a perspective on the use of reconfigurable devices and design automation tools for data analytics.