arXiv subjects

Miquel Pericàs

Publications and source records attributed to Miquel Pericàs.

11 recordsLinked to original sources

arXiv2026

FlashAttention for Scalable Vector Architectures

Inference with transformer models on CPUs is increasingly important, especially for Small Language Models (SLMs), where vector architectures are emerging as a promising execution substrate. The attention module is a major bottleneck due to high memory bandwidth requirements; FlashAttention mitigates this by fusing operations to improve data locality and reduce intermediate memory traffic. In this paper, we present FlashAttention-V, a blocked FlashAttention for scalable vector architectures that adapts efficiently from short to very long vectors by exploiting parallelism across attention heads, inter-head packing to enable efficient utilization of vector lengths beyond the head dimension, and improving vector register utilization and memory access locality. We integrate FlashAttention-V into ggml within llama.cpp and evaluate it on TinyLlama, Llama 3.2, Qwen2.5, and Pythia-410M using gem5 and a Banana Pi BPI-F3. On the Banana Pi BPI-F3, we confirm that loop reordering and loop unrolling across attention heads are effective optimization principles, scaling performance gains with larger models and most pronounced with short contexts and during decoding. Simulation-based analysis shows that FlashAttention-V achieves 22x-42x speedup over scalar FlashAttention at 512-bit VL in prefill, with an additional 2x-2.5x gain scaling to 64 lanes and 4096-bit VL. During decode, FlashAttention-V achieves 8x-11x speedup using 512-bit vector lengths over scalar FlashAttention, with performance showing diminishing sensitivity to vector width and lane count due to single-token, memory-bound execution. We further identify structural bottlenecks in Q8_0 quantized linear layers that limit arithmetic amortization under long-vector execution, consistent across RVV and Arm SVE, indicating that current quantization formats pose a fundamental challenge to long-vector scalability.

cs.LG↗

arXiv2023

Analysis and Characterization of Performance Variability for OpenMP Runtime

In the high performance computing (HPC) domain, performance variability is a major scalability issue for parallel computing applications with heavy synchronization and communication. In this paper, we present an experimental performance analysis of OpenMP benchmarks regarding the variation of execution time, and determine the potential factors causing performance variability. Our work offers some understanding of performance distributions and directions for future work on how to mitigate variability for OpenMP-based applications. Two representative OpenMP benchmarks from the EPCC OpenMP micro-benchmark suite and BabelStream are run across two x86 multicore platforms featuring up to 256 threads. From the obtained results, we characterize and explain the execution time variability as a function of thread-pinning, simultaneous multithreading (SMT) and core frequency variation.

cs.DC↗

arXiv2023

Challenges and Opportunities in the Co-design of Convolutions and RISC-V Vector Processors

The RISC-V "V" extension introduces vector processing to the RISC-V architecture. Unlike most SIMD extensions, it supports long vectors which can result in significant improvement of multiple applications. In this paper, we present our ongoing research to implement and optimize a vectorized Winograd algorithm used in convolutional layers on RISC-V Vector(RISC-VV) processors. Our study identifies effective techniques for optimizing the kernels of Winograd on RISC-VV using intrinsic instructions, and showcases how certain instructions offer better performance. Our co-design findings suggest that the Winograd algorithm benefits from vector lengths up to 2048 bits and cache sizes up to 64MB. We use our experience with Winograd to highlight potential enhancements for the standard that would simplify code generation and aid low-level programming. Finally, we share our experience from experimenting with forks of gem5 for RISC-VV and stress the importance of a mature software ecosystem, to facilitate design space exploration and architectural optimization.

cs.DC↗

arXiv2023

At the Locus of Performance: Quantifying the Effects of Copious 3D-Stacked Cache on HPC Workloads

Over the last three decades, innovations in the memory subsystem were primarily targeted at overcoming the data movement bottleneck. In this paper, we focus on a specific market trend in memory technology: 3D-stacked memory and caches. We investigate the impact of extending the on-chip memory capabilities in future HPC-focused processors, particularly by 3D-stacked SRAM. First, we propose a method oblivious to the memory subsystem to gauge the upper-bound in performance improvements when data movement costs are eliminated. Then, using the gem5 simulator, we model two variants of a hypothetical LARge Cache processor (LARC), fabricated in 1.5 nm and enriched with high-capacity 3D-stacked cache. With a volume of experiments involving a broad set of proxy-applications and benchmarks, we aim to reveal how HPC CPU performance will evolve, and conclude an average boost of 9.56x for cache-sensitive HPC applications, on a per-chip basis. Additionally, we exhaustively document our methodological exploration to motivate HPC centers to drive their own technological agenda through enhanced co-design.

cs.DC↗

arXiv2023

JOSS: Joint Exploration of CPU-Memory DVFS and Task Scheduling for Energy Efficiency

Energy-efficient execution of task-based parallel applications is crucial as tasking is a widely supported feature in many parallel programming libraries and runtimes. Currently, state-of-the-art proposals primarily rely on leveraging core asymmetry and CPU DVFS. Additionally, these proposals mostly use heuristics and lack the ability to explore the trade-offs between energy usage and performance. However, our findings demonstrate that focusing solely on CPU energy consumption for energy-efficient scheduling while neglecting memory energy consumption leaves room for further energy savings. We propose JOSS, a runtime scheduling framework that leverages both CPU DVFS and memory DVFS in conjunction with core asymmetry and task characteristics to enable energy-efficient execution of task-based applications. JOSS also enables the exploration of energy and performance trade-offs by supporting user-defined performance constraints. JOSS uses a set of models to predict task execution time, CPU and memory power consumption, and then selects the configuration for the tunable knobs to achieve the desired energy performance trade-off. Our evaluation shows that JOSS achieves 21.2% energy reduction, on average, compared to the state-of-the-art. Moreover, we demonstrate that even in the absence of a memory DVFS knob, taking energy consumption of both CPU and memory into account achieves better energy savings compared to only accounting for CPU energy. Furthermore, JOSS is able to adapt scheduling to reduce energy consumption while satisfying the desired performance constraints.

cs.DC↗

arXiv2023

ODIN: Overcoming Dynamic Interference in iNference pipelines

As an increasing number of businesses becomes powered by machine-learning, inference becomes a core operation, with a growing trend to be offered as a service. In this context, the inference task must meet certain service-level objectives (SLOs), such as high throughput and low latency. However, these targets can be compromised by interference caused by long- or short-lived co-located tasks. Prior works focus on the generic problem of co-scheduling to mitigate the effect of interference on the performance-critical task. In this work, we focus on inference pipelines and propose ODIN, a technique to mitigate the effect of interference on the performance of the inference task, based on the online scheduling of the pipeline stages. Our technique detects interference online and automatically re-balances the pipeline stages to mitigate the performance degradation of the inference task. We demonstrate that ODIN successfully mitigates the effect of interference, sustaining the latency and throughput of CNN inference, and outperforms the least-loaded scheduling (LLS), a common technique for interference mitigation. Additionally, it is effective in maintaining service-level objectives for inference, and it is scalable to large network models executing on multiple processing elements.

cs.DC↗

arXiv2022

Shisha: Online scheduling of CNN pipelines on heterogeneous architectures

Chiplets have become a common methodology in modern chip design. Chiplets improve yield and enable heterogeneity at the level of cores, memory subsystem and the interconnect. Convolutional Neural Networks (CNNs) have high computational, bandwidth and memory capacity requirements owing to the increasingly large amount of weights. Thus to exploit chiplet-based architectures, CNNs must be optimized in terms of scheduling and workload distribution among computing resources. We propose Shisha, an online approach to generate and schedule parallel CNN pipelines on chiplet architectures. Shisha targets heterogeneity in compute performance and memory bandwidth and tunes the pipeline schedule through a fast online exploration technique. We compare Shisha with Simulated Annealing, Hill Climbing and Pipe-Search. On average, the convergence time is improved by ~35x in Shisha compared to other exploration algorithms. Despite the quick exploration, Shisha's solution is often better than that of other heuristic exploration algorithms.

cs.PF↗

arXiv2022

Energy-Efficiency Evaluation of OpenMP Loop Transformations and Runtime Constructs

OpenMP is the de facto API for parallel programming in HPC applications. These programs are often computed in data centers, where energy consumption is a major issue. Whereas previous work has focused almost entirely on performance, we here analyse aspects of OpenMP from an energy consumption perspective. This analysis is accomplished by executing novel microbenchmarks and common benchmark suites on data center nodes and measuring the energy consumption. Three main aspects are analysed: directive-generated loop tiling and unrolling, parallel for loops and explicit tasking, and the policy of handling blocked threads. For loop tiling and unrolling, we find that tiling can yield significant energy savings for some, mostly unoptimised programs, while directive-generated unrolling provides very minor improvement in the best case and degenerates performance majorly in the worst case. For the second aspect, we find that parallel for loops yield better results than explicit tasking loops in cases where both can be used. This becomes more prominent with more fine-grained workloads. For the third, we find that significant energy savings can be made by not descheduling waiting threads, but instead having them spin, at the cost of a higher power consumption. We also analyse how the choice of compiler affects the above questions by compiling programs with each of ICC, Clang and GCC, and find that while neither is strictly better than the others, they can produce very different results for the same compiled programs. As a final step, we combine the findings of all results and suggest novel compiler directives as well as general recommendations on how to reduce energy consumption in OpenMP programs.

cs.DC↗

arXiv2022

ERASE: Energy Efficient Task Mapping and Resource Management for Work Stealing Runtimes

Parallel applications often rely on work stealing schedulers in combination with fine-grained tasking to achieve high performance and scalability. However, reducing the total energy consumption in the context of work stealing runtimes is still challenging, particularly when using asymmetric architectures with different types of CPU cores. A common approach for energy savings involves dynamic voltage and frequency scaling (DVFS) wherein throttling is carried out based on factors like task parallelism, stealing relations and task criticality. This paper makes the following observations: (i) leveraging DVFS on a per-task basis is impractical when using fine-grained tasking and in environments with cluster/chip-level DVFS; (ii) task moldability, wherein a single task can execute on multiple threads/cores via work-sharing, can help to reduce energy consumption; and (iii) mismatch between tasks and assigned resources (i.e.~core type and number of cores) can detrimentally impact energy consumption. In this paper, we propose ERASE (EneRgy Aware SchedulEr), an intra-application task scheduler on top of work stealing runtimes that aims to reduce the total energy consumption of parallel applications. It achieves energy savings by guiding scheduling decisions based on per-task energy consumption predictions of different resource configurations. In addition, ERASE is capable of adapting to both given static frequency settings and externally controlled DVFS. Overall, ERASE achieves up to 31% energy savings and improves performance by 44% on average, compared to the state-of-the-art DVFS-based schedulers.

cs.DC↗

arXiv2021

CBP: Coordinated management of cache partitioning, bandwidth partitioning and prefetch throttling

Reducing the average memory access time is crucial for improving the performance of applications running on multi-core architectures. With workload consolidation this becomes increasingly challenging due to shared resource contention. Techniques for partitioning of shared resources - cache and bandwidth - and prefetching throttling have been proposed to mitigate contention and reduce the average memory access time. However, existing proposals only employ a single or a subset of these techniques and are therefore not able to exploit the full potential of coordinated management of cache, bandwidth and prefetching. Our characterization results show that application performance, in several cases, is sensitive to prefetching, cache and bandwidth allocation. Furthermore, the results show that managing these together provides higher performance potential during workload consolidation as it enables more resource trade-offs. In this paper, we propose CBP a coordination mechanism for dynamically managing prefetching throttling, cache and bandwidth partitioning, in order to reduce average memory access time and improve performance. CBP works by employing individual resource managers to determine the appropriate setting for each resource and a coordinating mechanism in order to enable inter-resource trade-offs. Our evaluation on a 16-core CMP shows that CBP, on average, improves performance by 11% compared to the state-of-the-art technique that manages cache partitioning and prefetching and by 50% compared to the baseline without cache partitioning, bandwidth partitioning and prefetch throttling.

cs.AR↗

arXiv2020

An Adaptive Performance-oriented Scheduler for Static and Dynamic Heterogeneity

With the emergence of heterogeneous hardware paving the way for the post-Moore era, it is of high importance to adapt the runtime scheduling to the platform's heterogeneity. To enhance adaptive and responsive scheduling, we introduce a Performance Trace Table (PTT) into XiTAO, a framework for elastic scheduling of mixed-mode parallelism. The PTT is an extensible and dynamic lightweight manifest of the per-core latency that can be used to guide the scheduling of both critical and non-critical tasks. By understanding the per-task latency, the PTT can infer task performance, intra-application interference as well as inter-application interference. We run random Direct Acyclic Graphs (DAGs) of different workload categories as a benchmark on NVIDIA Jetson TX2 chip, achieving up to 3.25x speedup over a standard work-stealing scheduler. To exemplify scheduling adaption to interference, we run DAGs with high parallelism and analyze the scheduler's response to interference from a background process on an Intel Haswell (2650v3) multicore workstation. We also showcase the XiTAO's scheduling performance by porting the VGG-16 image classification framework based on Convolutional Neural Networks (CNN).

cs.DC↗