Pasar al contenido principal

MEJORA DEL PROCESADOR, SUBSISTEMA DE MEMORIA, ACELERADORES Y REDES

PID2024-158682OB-C31

•
Nombre agencia financiadora Agencia Estatal de Investigación
Acrónimo agencia financiadora AEI
Programa Programa Estatal para la Investigación y el Desarrollo Experimental
Subprograma Subprograma Estatal de Generación de Conocimiento Científico-Técnico y Desarrollo Experimental
Convocatoria Proyectos de I+D+i (Generación de Conocimiento y Retos Investigación)
Año convocatoria 2024
Plan Estatal 2024-2027 Plan Estatal de Investigación Científica y Técnica y de Innovación
Centro beneficiario UNIVERSITAT POLITÈCNICA DE VALÈNCIA

Resultados relacionados

Resultados totales (Incluyendo duplicados): 6
Encontrada(s) 1 página(s)

RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Synthesizing an optimal spin-qubit shuttling-bus architecture for the surface code

  • Escofet, Pau
  • Alarcon, Eduard
  • Abadal, S
  • Semenov, Andrii
  • Murphy, Niall
  • Blokhina, Elena
  • GARCIA ALMUDEVER, CARMEN|||0000-0002-3800-2357
[EN] As quantum computers scale toward millions of physical qubits, it becomes essential to robustly encode individual logical qubits to ensure fault tolerance under realistic noise. A high-quality foundational encoding allows future compilation techniques and heuristics to build on optimal or near-optimal layouts, improving scalability and error resilience. In this work, we synthesize a one-dimensional shuttling-bus architecture for the rotated surface code, leveraging coherent spin-qubit shuttling, following a methodology we name quantum reverse mapping. We formulate a mixed-integer optimization model that yields optimal solutions with relatively low execution time for small code distances, and propose a scalable heuristic that matches optimal results while maintaining linear computational complexity. We evaluate the synthesized architecture using architectural metrics, such as shuttling distance and cycle time, and full quantum simulations under realistic noise models, showing that the proposed design can sustain logical error rates as low as 2 & times; 10-10 per round at code distance 21, showcasing its feasibility for scalable quantum error correction in spin-based quantum processors., Authors acknowledge funding from the EC through Grant No. HORIZON-EIC-2022-PATHFINDEROPEN-01-101099697 (QUADRATURE) . E.A. acknowledges funding from the Spanish Ministry of Science, Innovation, and Universities and European ERDF (QCOMM-CAT-Planes Complementarios) , Generalitat de Catalunya, and ICREA Academia Grant 2024. S.A. acknowledges funding from the EC through Grant No. HORIZON-ERC-2021-101042080 (WINC) . E.B. acknowledges funding from Science Foundation Ireland under Grant No. 21/RP-2TF/10019. C.G.A. acknowledges funding from QuantERA EQUIP Grant No.PCI2022-133004, funded by Agencia Estatal de Investigacion, Ministerio de Ciencia e Innovacion, Gobierno de Espana, Project No. MCIN/AEI/10.13039/501100011033, from the European Union NextGenerationEU/PRTR, and from the Spanish Ministry of Science, Innovation, and Universities and European ERDF under Grants No. PID2021-123627OB-C51 and No. PID2024-158682OB-C31. P.E. acknowledges funding from the predoctoral FPI-UPC grant supported by UPC and Banco Santander.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

TailBench plus plus : flexible multi-client, multi-server benchmarking for latency-critical workloads

  • Li, Zhibin
  • Pons-Escat, Lucía|||0000-0002-4582-7744
  • Petit Martí, Salvador Vicente|||0000-0003-2426-4134
  • Pons Terol, Julio|||0000-0002-5654-6753
  • Sahuquillo Borrás, Julio|||0000-0001-8630-4846
[EN] Cloud systems have rapidly expanded worldwide in the last decade, shifting computational tasks to cloud servers where clients submit their requests. Among cloud workloads, latency-critical applications -characterized by high-percentile response times - have gained special interest. These applications are present in modern services, representing an important fraction of cloud workloads. This work analyzes common cloud benchmarking suites and identifies TailBench as the most suitable to assess cloud performance with latency-critical workloads. Unfortunately, this suite presents key limitations, especially in multi-server scenarios or environments with variable client arrival patterns and fluctuating loads. To address these limitations, we propose TailBench++, an enhanced benchmark suite that extends TailBench to enable cloud evaluation studies to be performed in dynamic multi-client, multi-server environments. It allows reproducing experiments with varying client arrival times, dynamic query per second (QPS) fluctuations, and multiple servers handling requests. Case studies show that TailBench++ enables more realistic evaluations by capturing a wider range of real-world scenarios., This work has been supported by the Spanish Ministerio de Ciencia e Innovacion and European ERDF under grants PID2021-123627OB-C51 and PID2024-158682OB-C31, as well as by the Generalitat Valenciana under grant CIPROM/2024/007.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

Power-aware scheduling strategies for improving performance and energy consumption in heterogeneous multicore platforms

  • Mazón, Ricardo
  • Hassan Mohamed, Houcine|||0000-0002-2038-0587
[EN] Energy consumption is an increasingly critical concern in multicore embedded real-time systems, especially when it is referred to systems that operate under battery constraints as mobile devices. Moreover, these systems must enhance performance such as the execution time while meeting temporal constraints to reduce deadline miss rates. Although Heterogeneous Multi-Processing (HMP) architectures offer opportunities for power optimization using dynamic voltage and frequency scaling (DVFS), standard operating system (OS) governors do not consider benchmark characteristics, memory or CPU usage, and optimal co-runner configurations. This limitation leads to scheduling decisions that trigger early battery depletion or extreme deadline violations under high system utilization. To solve these challenges, this paper proposes a two-level scheduling and DVFS framework. The proposed solution performs profiling benchmark-pair combinations to construct a static co-runner compatibility matrix. Online, a supervisor level (Partitioner) dynamically evaluates global system constraints, tracking remaining battery status and runtime deadline miss rates to adaptively select an heuristic strategy (A_max, B_max, B_eff). In the second level, a Task Scheduler executes tasks based on Earliest Deadline First (EDF) policy, integrated with a power estimation model. This structure translates a nonlinear, NP-hard multi-objective optimization problem into a deterministic, constant-time O(1) runtime overhead per task. Experimental evaluations are conducted on an ODROID XU4 platform containing a Samsung Exynos 5422 heterogeneous multicore processor, based on an ARM big.LITTLE architecture, using a set of embedded real-time benchmarks. The results show that: (i) B_max strategy delivers a suitable trade-off by simultaneously reducing mean energy consumption by about 7% and execution time by around 8%, compared to the OS., at the same time reducing drastically deadline failures by 72.61% under a 50% workload; (ii) the aggressive B_eff configuration, decreases power consumption by 50% in situations with ample slack time; and (iii) the A_max recovery strategy achieves about 29% acceleration, hence preventing system degradation under extreme workload saturation. Finally, integrating these context-aware mechanisms prevents the hardware from performance collapse, maximizing battery longevity while ensuring suitable real-time predictability., This work has been supported by the Spanish Ministerio de Ciencia e Innovación and European ERDF under grants PID2021-123627OB-C51 and PID2024-158682OB-C31, and Generalitat Valenciana under grant CIPROM/2024/007.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

SYNPA: understanding the impact of the interference modeling on thread-to-core allocation policies for SMT ARM processors

  • Navarro-Edo, Marta|||0000-0002-4334-460X
  • Feliu-Pérez, Josué|||0000-0003-3017-4266
  • Petit Martí, Salvador Vicente|||0000-0003-2426-4134
  • Gómez Requena, María Engracia|||0000-0003-1466-4118
  • Sahuquillo Borrás, Julio|||0000-0001-8630-4846
[EN] Modern high-performance servers increasingly rely on Simultaneous Multithreading (SMT) processors to enhance throughput with minimal area overhead. However, SMT architectures introduce inter-application interference, often resulting in degraded performance for individual applications. To address this issue, interference-aware thread-to-core (T2C) allocation policies are essential. This paper explores the design and implementation of such policies using real performance counters on ARM processors. We introduce the Instructions and Stalls Cycles (ISC) stack¿a simple yet effective model for characterizing application behavior and identifying synergistic thread pairings. Building on our previous work, SYNPA, we improve the accuracy of the model by accounting for horizontal waste (that is, unused dispatch slots) and proposing methods to address limitations in ARM¿s Performance Monitoring Unit (PMU), which prevent complete attribution of processor cycles. These enhancements result in a family of SYNPA schedulers, each based on a different ISC stack variant. Detailed discussions are provided on the pros and cons that researchers typically face when building a performance stack on commercial processors. These analyses are intended to assist researchers in their work., This work has been supported by the Spanish Ministerio de Ciencia e Innovación and European ERDF under grant PID2024-158682OB-C31 and Generalitat Valenciana under grant CIPROM/2024/007. Marta Navarro is supported by Subvenciones para la contratación de personal investigador predoctoral by CIACIF/2021/413. Josué Feliu is supported by the RYC2021-030862-I contract, funded by the MCIN/AEI /10.13039/501100011033 and the European Union NextGenerationEU/PRTR.
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Tesis doctoral (DoctoralThesis)

Optimizing Prefetch with Machine Learning and Analyzing its Impact on the Performance of Deep Memories Hierarchies

  • Lurbe Sempere, Manel
[ES] La creciente presión sobre los recursos compartidos de la jerarquía de memoria en los sistemas de cómputo modernos se ha convertido en un factor crítico que limita el rendimiento. La proliferación de aplicaciones concurrentes, impulsada por cargas de trabajo de aprendizaje automático y entornos virtualizados en la nube, ha transformado la concepción tradicional de las jerarquías de memoria y acentuado la necesidad de ampliar la capacidad de la memoria principal. Sin embargo, la memoria DRAM convencional enfrenta restricciones físicas y tecnológicas que comprometen su escalabilidad, lo que ha impulsado la incorporación de memorias no volátiles (NVRAM) como alternativa emergente. A pesar de sus ventajas en densidad, la mayor latencia de acceso de la NVRAM plantea nuevos desafíos que requieren rediseñar y optimizar los mecanismos de gestión de memoria. En este contexto, los mecanismos de prefetching resultan fundamentales para mitigar las penalizaciones de latencia y aprovechar plenamente las nuevas tecnologías de memoria. No obstante, la configuración de prefetchers en arquitecturas multinivel plantea tensiones significativas, dado que su carácter especulativo incrementa el consumo de ancho de banda y puede deteriorar el rendimiento en sistemas con alta concurrencia. Así, la selección adecuada de configuraciones de prefetching requiere un equilibrio entre las características de acceso de las aplicaciones y la eficiencia en el uso del ancho de banda compartido. De manera complementaria, la integración de memorias NVRAM en la jerarquía introduce profundos cambios estructurales. La menor velocidad de estas memorias en comparación con la DRAM exige el diseño de nuevas organizaciones de caché y controladores de memoria capaces de mitigar sus limitaciones de latencia y preservar el rendimiento de los procesadores de altas prestaciones. Esta tesis aborda de forma conjunta dos problemáticas clave en la evolución de las jerarquías de memoria: (i) la selección óptima de configuraciones de prefetching en cargas de trabajo multiprogramadas mediante técnicas basadas en redes neuronales, y (ii) el diseño de jerarquías profundas con NVRAM orientadas a aumentar la capacidad de memoria principal sin comprometer el rendimiento., [EN] The increasing pressure on shared resources within the memory hierarchy of modern computing systems has become a critical performance-limiting factor. The proliferation of concurrent applications, driven by machine learning workloads and cloud-based virtualized environments, has reshaped the traditional conception of memory hierarchies and highlighted the need to expand main memory capacity. However, conventional DRAM faces physical and technological constraints that hinder its scalability, fostering the integration of non-volatile memories (NVRAM) as an emerging alternative. Despite their advantages in density, NVRAM technologies introduce higher access latency, which demands new strategies for memory management and optimization. In this context, prefetching mechanisms play a fundamental role in mitigating latency penalties and fully exploiting the potential of emerging memory technologies. Nevertheless, the configuration of prefetchers in multilevel architectures poses significant challenges, as their speculative nature increases bandwidth consumption and can degrade performance in highly concurrent systems. Therefore, the optimal selection of prefetching configurations requires balancing application-specific access patterns with efficient shared bandwidth utilization. Complementarily, the integration of NVRAM into the hierarchy introduces profound structural changes. The slower speed of these memories compared to DRAM necessitates the design of new cache organizations and memory controllers to mitigate latency limitations and sustain high-performance processor efficiency. This dissertation %jointly addresses two key challenges in the evolution of memory hierarchies: (i) the optimal selection of prefetching configurations in multiprogrammed workloads through neural network-based techniques, and (ii) the design of deep NVRAM hierarchies aimed at increasing main memory capacity without compromising system performance., [CA] La creixent pressió sobre els recursos compartits dins de la jerarquia de memòria dels sistemes de computació moderns s'ha convertit en un factor crític que limita el rendiment. La proliferació d'aplicacions concurrents, impulsada per les càrregues de treball d'aprenentatge automàtic i els entorns virtualitzats al núvol, ha transformat la concepció tradicional de les jerarquies de memòria i ha posat de manifest la necessitat d'ampliar la capacitat de la memòria principal. No obstant això, la DRAM convencional afronta restriccions físiques i tecnològiques que en dificulten l'escalabilitat, fet que ha propiciat la incorporació de memòries no volàtils (NVRAM) com a alternativa emergent. Tot i els seus avantatges en densitat, les NVRAM introdueixen una latència d'accés superior, la qual cosa exigeix noves estratègies de gestió i optimització de la memòria. En aquest context, els mecanismes de prefetching són fonamentals per a mitigar les penalitzacions de latència i aprofitar plenament el potencial de les noves tecnologies de memòria. Tanmateix, la configuració dels prefetchers en arquitectures multinivell planteja reptes importants, atés que el seu caràcter especulatiu incrementa el consum d'ample de banda i pot degradar el rendiment en sistemes amb alta concurrència. Per tant, la selecció òptima de configuracions de prefetching requereix equilibrar els patrons d'accés específics de les aplicacions amb una utilització eficient dels recursos compartits. De manera complementària, la integració de NVRAM en la jerarquia introdueix canvis estructurals profunds. La menor velocitat d'aquestes memòries en comparació amb la DRAM obliga a dissenyar noves organitzacions de memòria cau i nous controladors de memòria capaços de mitigar les seues limitacions de latència i mantindre l'eficiència dels processadors d'altes prestacions. Aquesta tesi aborda conjuntament dos reptes clau en l'evolució de les jerarquies de memòria: (i) la selecció òptima de configuracions de prefetching en càrregues de treball multiprogramades mitjançant tècniques basades en xarxes neuronals, i (ii) el disseny de jerarquies profundes amb NVRAM orientades a augmentar la capacitat de la memòria principal sense comprometre el rendiment del sistema., This work has been supported by the Spanish Ministerio de Ciencia e Innovación and European ERDF under grants PID2021-123627OB-C51, PID2024-158682OB-C31, by MCIN/AEI/10.13039/501100011033 and European Union NextGenerationEU/PRTR under grant TED2021-130233B-C32, and Generalitat Valenciana under grant CIPROM/2024/007. Manel Lurbe is supported by Ayudas Contratos predoctorales UPV-subprograma 1 (PAID-01-20).
RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Publicación Artículo científico (JournalArticle)

WAPA: A Microarchitecture- and Workload-Agnostic Universal SMT Scheduler

  • Navarro-Edo, Marta|||0000-0002-4334-460X
  • Pons-Escat, Lucía|||0000-0002-4582-7744
  • Petit Martí, Salvador Vicente|||0000-0003-2426-4134
  • Gómez Requena, María Engracia|||0000-0003-1466-4118
  • Sahuquillo Borrás, Julio|||0000-0001-8630-4846
  • Pallardó-Julià, Vicent
[EN] This paper presents WAPA, a thread-to-core allocation policy for SMT processors grounded in OT theory, and demonstrates its effectiveness in achieving two key design goals: microarchitecture agnosticism and workload agnosticism. WAPA is implemented on both Intel Xeon and ARM ThunderX2 platforms without requiring modifications, leveraging universally available performance counters. Experimental results show that WAPA consistently outperforms both the default Linux scheduler and state-of-the-art model-based approaches, achieving up to 110% improvement in turnaround time on Intel and over 30% average gains on ARM. Furthermore, we show WAPA is extensible to support alternative cost metrics beyond CPI, such as energy consumption or fairness. These findings position WAPA as a practical and portable solution for integration into modern operating system schedulers., The work of Marta Navarro was supported by Generalitat Valencianathrough "Subvenciones para la contratacion de personal investigador pre-doctoral" under Grant CIACIF/2021/413. This work was supported in partby MICIU/AEI/10.13039/501100011033/ FEDER, UE under Grant PID2024-158682OB-C31 and Grant PID2021-123627OB-C51 and in part by GeneralitatValenciana under Grant CIPROM/2024/7.