Modern scienti!c research faces increasingly complex computational demands that span multiple computing paradigms. A single research project may require high-performance computing (HPC) for parallel simulations, high-throughput computing (HTC) for processing large volumes of independent tasks, and specialized environments for machine learning and arti!cial intelligence (ML/AI) applications. Traditional computing infrastructures-designed primarily for a single computing paradigm-often struggle to accommodate this diversity of needs, creating bottlenecks in scienti!c work"ows and hindering interdisciplinary collaboration.
We designed and developed the Recon!gurable Advanced Platform for Transdisciplinary Open Research (RAPTOR) to address some of these challenges through a novel architecture that dynamically adapts to diverse computational requirements. Unlike static and monolithic computing environments, RAPTOR implements a recon!gurable design that allows compute nodes to be provisioned on demand between an HPC cluster and the Chameleon Cloud platform [8]. This "exibility enables researchers to access the most suitable computing environment for their speci!c applicationswhether that involves bare-metal provisioning for specialized software stacks, integration with the Open Science Grid (OSG) [12] for high-throughput computing, or HPC cluster resources for tightly coupled parallel applications.
This paper demonstrates RAPTOR's e#ectiveness in supporting diverse scienti!c work"ows across multiple domains, with a speci!c focus on how its recon!gurability features address key computational challenges. We present performance benchmarks from four scienti!c applications-proteomics analysis, sea level prediction, weather research, and hurricane risk assessment-highlighting RAPTOR's speci!c advantages compared to traditional computing environments.
The remainder of this paper is organized as follows. Section 2 presents RAPTOR's design considerations in addressing the various computational demands for modern research. In Section 3, we discuss RAPTOR's architecture which provides recon!gurability and adaptability to a diverse range of scienti!c work"ows based on data characteristics and research requirements. In Section 4, we examine several representative applications to demonstrate RAP-TOR's capabilities in advancing scienti!c discovery. By outlining these aspects, this study highlights how RAPTOR enhances computational e$ciency, fosters interdisciplinary collaboration, and supports a broad spectrum of research domains. We conclude the paper and outline future directions in Section 5.
Traditional computing infrastructures are typically optimized for a single computing paradigm, making them ine$cient for the increasingly diverse computational demands of modern scienti!c research. We identi!ed key challenges that informed RAPTOR's recon!gurable design:
• Resource allocation in"exibility: Fixed hardware allocations in traditional HPC environments lead to resource under-utilization for work"ows with varying computational needs over time. • Software environment constraints: Standard HPC clusters impose strict software environments that limit the deployment of specialized applications, such as emerging AI frameworks or domain-speci!c tools. • Interoperability barriers: Institutional boundaries restrict access to specialized computing resources, hindering collaborative research that could bene!t from shared infrastructure. • Work"ow transition overhead: Moving between computing paradigms (e.g., from HPC simulation to ML analysis) typically requires manual data transfer and environment recon!guration, creating work"ow bottlenecks. • Heterogeneous resources: Modern scienti!c applications increasingly require the seamless integration of diverse computing resources, including CPUs, GPUs, high-memory nodes, and storage systems, which would be challenging to achieve in any monolithic system environments.
To address these challenges, RAPTOR implements a recon!gurable architecture with !ve key polymorphic capabilities:
(1) High-performance computing: RAPTOR connects highend compute nodes to FIU's existing HPC cluster, providing a uni!ed environment for parallel applications with dynamic scaling capabilities that overcome !xed allocation limitations of traditional clusters. (2) High-throughput computing: Through OSG integration [5], RAPTOR enables the opportunistic execution of independent jobs across distributed resources, extending computational reach beyond institutional boundaries. (3) Data-intensive computing: RAPTOR nodes feature multiprocessor, multi-core servers with GPU acceleration, large memory (up to 1.5 TB/node), and scalable storage for big data processing with dynamic resource allocation. (4) Real-time on-demand provisioning: Using Chameleon's bare-metal provisioning, researchers gain full control over software stacks, enabling customized environments for applications with strict deadlines or specialized con!gurations. While container solutions like Kubernetes or Singularity provide portability on HPC and cloud platforms, RAPTOR's emphasis on bare-metal provisioning o#ers a complementary path that ensures unrestricted hardware access and allows full OS/kernel customization. (5) Federated resource integration: RAPTOR operates as a federated resource with Chameleon and OSG, enabling researchers to access and utilize resources based on their speci!c requirements while maintaining secure isolation.
These polymorphic capabilities are enabled through several key recon!gurability mechanisms: (1) network-level recon!guration using VLANs and software-de!ned networking, (2) system-level recon!guration with automated OS provisioning, (3) resource management integration across multiple scheduling systems, and (4) uni!ed storage access across computing environments. Together, these mechanisms allow RAPTOR to dynamically adapt its architecture to meet the diverse computational demands of scienti!c applications while maintaining e$cient resource utilization.
RAPTOR's design intentionally balances two key sets of tradeo#s. First, it weighs the increased management complexity and switching overhead of a recon!gurable system against the "exibility gains o#ered by specialized con!gurations. Second, it considers the security challenges introduced by federated access alongside the collaborative research opportunities that federation makes possible [2,3]. Our evaluation of diverse scienti!c work"ows demonstrates that these trade-o#s are well-justi!ed by the signi!cant performance improvements and research productivity gains achieved through RAPTOR's recon!gurable architecture.
RAPTOR's architecture is designed around the principle of recon-!gurability, enabling dynamic resource allocation across di#erent computing environments. As shown in Figure 1, the system consists of three interconnected zones: the CHI Zone (Chameleon Cloud), the HPC Zone, and the Storage Zone [1].
The CHI Zone features RAPTOR Nodes with AMD EPYC processors, DDR4 memory, and NVIDIA A100 GPUs, provisioned via Chameleon's CHI-in-a-Box [9]. It includes Management Nodes for lease management and a Storage Server providing 1 TB of scratch storage per user. RAPTOR's bare-metal access eliminates virtualization overhead for performance-sensitive applications.
The HPC Zone is connected to FIU's High-Performance Cluster [4], which features over 3,000 Intel cores, high-memory nodes, and 22 GPUs. The cluster comprises Master Nodes, compute nodes, a Login Node, and a Visualization Node, utilizing both Ethernet and In!niBand interconnects. It federates with OSG, extending capabilities beyond institutional boundaries, and uses DDN Storage for high-throughput data access.
A key innovation is the Provisioning Server which dynamically recon!gures nodes between zones using IPMI. This process is managed through a CLI tool that automates network settings, boot environments, and resource allocation. By default, nodes reside in the CHI Zone until speci!cally requested for HPC use.
The Storage Zone provides "exible, high-performance storage through a combination of technologies. The Storage Zone provides "exible storage through S3-compatible object storage, CEPH distributed storage [15], and iSCSI block storage, enabling researchers to scale resources according to speci!c work"ow requirements.
While the current system uses speci!c core zones, its architecture could theoretically support more. However, the scalability of managing numerous zones, particularly their control plane, requires future investigation.
The network infrastructure consists of two key switches that enable RAPTOR's recon!gurability. A Dell Switch is positioned between the CHI Zone and HPC Zone, facilitating the dynamic reassignment of computing resources between these environments. An Enterprise Switch connects the Storage Zone to both the HPC Zone and CHI Zone, ensuring that storage resources are accessible regardless of which computing environment is being used. These switches implement network-level recon!guration, enabling RAPTOR's dynamic resource allocation capabilities.
Network-level recon!guration: The system uses Virtual Local Area Networks (VLANs) and software-de!ned networking [10] to dynamically reassign nodes between the CHI and HPC environments. This recon!guration is managed by the Dell Switch and Enterprise Switch, which provide connectivity across zones.
System-level recon!guration: The Provisioning Server handles operating system deployment and con!guration, allowing nodes to boot into di#erent environments based on workload requirements. This includes specialized OS images for bare-metal deployments through Chameleon and standard HPC environments for cluster-based workloads.
Resource management integration: RAPTOR integrates with both the Slurm workload manager [16] (for HPC workloads) and Chameleon's leasing system (for bare-metal provisioning). This allows researchers to request resources through familiar interfaces while leveraging RAPTOR's recon!gurability.
Automated work"ow support: The system includes tools for transitioning data and work"ows between environments, enabling researchers to move seamlessly between computing paradigms as their needs evolve. Importantly, RAPTOR's node recon!guration between zones (e.g., from the HPC Zone to the CHI Zone or vice versa) occurs between application executions. Running applications, including any tightly-coupled modules, are allowed to complete their execution on their allocated resources before these resources become eligible for reprovisioning into a di#erent environment for subsequent workloads. This operational model ensures that the recon!guration process itself does not impact the performance of active computations. Table 1 summarizes RAPTOR's recon!gurability features and their bene!ts for scienti!c work"ows.
To demonstrate RAPTOR's e#ectiveness in supporting diverse scienti!c work"ows, we evaluated its performance across four representative applications spanning di#erent computational paradigms. We highlight the speci!c recon!gurability features for each application that addressed key computational challenges.
Computational proteomics relies on comparing experimental mass spectrometry (MS) spectra against large databases of theoretical spectra to identify peptides. As database sizes grow to the terabyte scale for proteogenomic and metaproteomics analyses [6], traditional computing approaches become impractical, with processing times extending to weeks or months. RAPTOR's recon!gurability features enabled signi!cant performance improvements for the GiCOPS framework [7], which implements a distributed database search algorithm for peptide identi!cation. By leveraging bare-metal provisioning, researchers con!gured specialized environments with optimized CUDA and MPI installations-the ability to dynamically allocate GPU-accelerated nodes on demand provided "exibility unavailable on traditional HPC resources.
Without making any modi!cation to the code base, we recon!gured the RAPTOR environment (using bare-metal nodes), compiled MPI/CUDA GiCOPS code base, and then executed the experiments. Our experiments demonstrate that GiCOPS outperforms the existing GPU-based database search algorithms by more than 10 times in both closed-and open-search modes and 2 times as compared to the CPU-only HiCOPS. The code for GICOPS (MPI/CUDA) was compiled on RAPTOR without any modi!cations. The speedup attained when running GiCOPS on RAPTOR is comparable to the speedup obtained on traditional distributed memory architecture for same number of nodes.
Experiments used PRIDE datasets PXD055735, PXD055119, and PXD015384 [13], scaling MPI processes while maintaining constant parameters. Figure 2 shows execution time and speedup results. Our results show that we were able to successfully compile, troubleshoot, and execute MPI and CUDA based code bases with minimal modi!cation on the RAPTOR machines.
By scaling work"ows in the omics disciplines, RAPTOR empowers researchers to push the boundaries of computational biology, leading to faster biomarker discovery, improved disease modeling, and more e#ective drug development strategies. More speci!cally, our results show that RAPTOR's GPU-accelerated nodes provided a 6x-12x speedup (due to GPU node's availability) compared to CPU-only implementations (for 4 MPI processes with one GPU per process). This acceleration reduced processing time from days to hours, enabling more comprehensive analyses that would be impractical on traditional resources.
For this particular application, key bene!ts included customized CUDA/MPI con!gurations, on-demand GPU allocation, and seamless development-to-production transitions.
Coastal "ooding prediction requires processing large multivariate time series datasets to identify patterns and make forecasts. Traditional statistical approaches struggle with the complex, nonlinear relationships in these datasets, leading researchers to explore machine-learning techniques.
We used RAPTOR to evaluate PatchTST [11], a state-of-the-art multivariate time series transformer model, for predicting sea level variability at 21 National Oceanic and Atmospheric Administration (NOAA) tide gauges. The recon!gurable nature of RAPTOR allowed researchers to create custom environments with specialized deeplearning frameworks.
We compared training times for the PatchTST model on RAP-TOR with those on local workstations for a multivariate dataset comprising 21 variables across 52,696 time steps. The results show a consistent speedup on RAPTOR compared to local workstations. This acceleration enabled researchers to explore more hyperparameter con!gurations and longer prediction horizons, signi!cantly enhancing model accuracy for coastal "ooding prediction.
This research signi!cantly bene!ted from RAPTOR's recon!gurability, including: 1) a customized environment with specialized deep-learning frameworks; 2) having on-demand access to highmemory nodes for extensive dataset processing; and 3) the ability to scale computations based on model complexity. The Weather Research and Forecasting (WRF) model [14] is a numerical weather prediction system designed for atmospheric research and operational forecasting. Running high-resolution WRF simulations requires substantial computational resources and specialized software con!gurations. RAPTOR's recon!gurability enabled researchers to deploy the WRF model for South Florida with customized environments optimized for meteorological simulations.
The bare-metal provisioning capability enabled the installation of specialized libraries and dependencies that are challenging to con!gure in traditional HPC environments. We evaluated WRF's performance on RAPTOR for high-resolution (1 km) simulations covering a 500 x 500 km domain centered on South Florida. Table 2 compares di#erent simulation con!gurations.
The RAPTOR environment enabled high-resolution WRF simulations that were previously impractical due to software con!guration challenges. These simulations o#er crucial insights into local weather patterns that impact coastal communities, including severe weather events such as heavy rainfall and thunderstorms.
Key bene!ts for this application included specialized environment con!guration for meteorological simulations, access to highperformance I/O for multi-source data integration, and ability to adjust resource allocation based on simulation complexity.
The Florida Public Hurricane Loss Model (FPHLM) [3] requires simulating over 140,000 synthetic storms to estimate property losses and insurance risks. This computationally intensive process involves independent simulations, which are ideal for high-throughput computing approaches. RAPTOR's integration with the Open Science Grid (OSG) enabled the e$cient execution of these simulations across distributed resources. By federating RAPTOR with OSG, researchers gained access to computational capacity far beyond local resources, significantly reducing the time required for comprehensive risk assessments.
For the 140,000 storm simulations, attempts to use only the local FIU High-Performance Cluster (RAPTOR's HPC Zone) faced significant throughput limitations. However, RAPTOR's orchestration of this workload across the federated Open Science Grid (OSG) dramatically reduced computation time from months to weeks. This highlights RAPTOR's cross-environment capability-leveraging OSG beyond the local HPC Zone-delivering quantitative performance gains compared to standard execution on institutional resources alone.
This research is bene!ted from RAPTOR's recon!gurability. More speci!cally, the federation with OSG enabled access to distributed computing resources beyond what has been o#ered at the university. Also, the high-throughput computing environment can be optimized for many independent simulations. E$cient data management allows for large simulation outputs.
RAPTOR demonstrates that a recon!gurable computing platform can e#ectively address the diverse computational needs of modern scienti!c research. RAPTOR overcomes key limitations of traditional computing infrastructures by providing dynamic allocation between HPC and cloud environments, bare-metal provisioning for specialized applications, and federation with national resources. Our performance evaluations across four scienti!c domains show signi!cant advantages compared to traditional approaches:
• Accelerated computation: For GPU-accelerated proteomics work"ows, RAPTOR provided speedups of up to 12x. • Work"ow "exibility: The recon!gurable architecture enabled seamless transitions between computing paradigms, supporting complex work"ows that span HPC, HTC, and ML/AI applications.
• Resource e#ciency: Dynamic resource allocation optimizes utilization based on current research demands, avoiding the underutilization common in static computing environments. • Enhanced collaboration: The federation with national resources extended computational reach beyond institutional boundaries, fostering multi-institutional research collaborations that transcend geographical boundaries.
These bene!ts directly address the challenges identi!ed in the FlexScience workshop's focus on "exible computing infrastructures for scienti!c applications. RAPTOR's approach to recon!gurability o#ers valuable insights for designing the next generation of research computing platforms that can adapt to evolving computational demands.
Future work will integrate RAPTOR with our petabyte-scale storage system designed to support data-intensive applications. This integration will further enhance RAPTOR's capabilities for realtime simulations, high-throughput data processing, and machine learning-driven research.