arXiv subjects

Nuwan Jayasena

Publications and source records attributed to Nuwan Jayasena.

10 recordsLinked to original sources

arXiv2026

Characterizing LLM Kernel Access and Memory Interaction in Multi-Partition NUMA GPUs

Large language model (LLM) workloads motivate multi-partition GPUs as a path to scaling compute and memory capacity, but their non-uniform memory access characteristics and inter-partition communication can amplify contention and degrade locality, leading to suboptimal kernel latency. To address this, we analyze performance-critical LLM kernel implementations spanning weight projection, mixture-of-experts, and attention variants of state-of-the-art serving engines to present a characterization of data access patterns in multi-partition GPUs. First, we introduce memory trace analysis methodology to derive workgroup-level data access and sharing behavior, then evaluate the locality implications on latency using a cycle-level simulator. Using these tools, we categorize LLM kernel operands into three inter-workgroup sharing patterns (global, partial, or private) and show that the required optimization strategies differ across categories, from simple per-workgroup pinning to subgroup-aware co-scheduling. Our findings highlight the need for placement-aware kernel programming and smarter architectural support for work and data locality in multi-partition GPUs.

cs.AR↗

arXiv2026

The $qs$ Inequality: Quantifying the Double Penalty of Mixture-of-Experts at Inference

Mixture-of-Experts (MoE) models deliver high quality at low training FLOPs, but this efficiency often vanishes at inference. We identify a double penalty that structurally disadvantages MoE architectures during decoding: first, expert routing fragments microbatches and reduces weight reuse; second, massive resident expert pools reduce high-bandwidth memory (HBM) headroom for the KV cache. This phenomenon, formalized as reuse fragmentation, pushes feed-forward networks (FFNs) into a bandwidth-bound regime, especially at long context lengths. We introduce the $qs$ inequality, a predictive criterion that identifies when MoE is structurally disadvantaged relative to a quality-matched dense model. This criterion unifies sparsity ($s$), the fraction of parameters activated per token, and the quality-equivalence factor ($q$), the size multiplier required for a dense model to match MoE performance. Our evaluation across frontier models including DeepSeek-V3, Qwen3-235B, Grok-1, and Switch-C demonstrates that this fragmentation is a general architectural phenomenon. For DeepSeek-V3 at 128k context, this results in a 4.5x throughput advantage for a quality-matched dense baseline. Crucially, massive architectures like Switch-C can become infeasible on cluster sizes where a quality-matched dense model remains viable. Our results suggest that training-time FLOP efficiency is an incomplete proxy for inference-time performance in long-context serving. They also indicate that MoE may be best viewed as a training-time optimization, with distillation into dense models as a possible path toward inference-efficient deployment.

cs.LG↗

arXiv2026

RAPID-Serve: Resource-efficient and Accelerated P/D Intra-GPU Disaggregation

Two widely adopted techniques for LLM inference serving systems today are hybrid batching and disaggregated serving. A hybrid batch combines prefill and decode tokens of different requests in the same batch to improve resource utilization and throughput at the cost of increased latency per token. In contrast, disaggregated serving decouples compute-bound prefill and bandwidth-bound decode phases to optimize for service level objectives (SLOs) at the cost of resource under-utilization and KV-cache transfer overheads. To address the limitations of these techniques, we propose RAPID-Serve: a technique to concurrently execute prefill and decode on the same GPU(s) to meet latency SLOs while maintaining high throughput and efficient resource utilization. Furthermore, we propose Adaptive Resource Management for runtime compute resource allocation, optionally leveraging CU masking (a fine-grained Compute Unit partitioning feature on AMD Instinct\textsuperscript{TM} GPUs). RAPID-Serve provides up to 4.1x (average 1.7x) unconstrained throughput improvement and 32x and higher (average 4.9x) throughput improvement under SLO constraints, showing it as an effective strategy compared to the state-of-the-art approaches, particularly in resource-constrained environments.

cs.DC↗

arXiv2024

Global Optimizations & Lightweight Dynamic Logic for Concurrency

Modern accelerators like GPUs are increasingly executing independent operations concurrently to improve the device's compute utilization. However, effectively harnessing it on GPUs for important primitives such as general matrix multiplications (GEMMs) remains challenging. Although modern GPUs have significant hardware and software support for GEMMs, their kernel implementations and optimizations typically assume each kernel executes in isolation and can utilize all GPU resources. This approach is highly efficient when kernels execute in isolation, but causes significant resource contention and slowdowns when kernels execute concurrently. Moreover, current approaches often only statically expose and control parallelism within an application, without considering runtime information such as varying input size and concurrent applications -- often exacerbating contention. These issues limit performance benefits from concurrently executing independent operations. Accordingly, we propose GOLDYLOC, which considers the global resources across all concurrent operations to identify performant GEMM kernels, which we call globally optimized (GO)-Kernels. Moreover, GOLDYLOC introduces a lightweight dynamic logic which considers the dynamic execution environment for available parallelism and input sizes to execute performant combinations of concurrent GEMMs on the GPU. Overall, GOLDYLOC improves performance of concurrent GEMMs on a real GPU by up to 2$\times$ (18% geomean per workload) and provides up to 2.5$\times$ (43% geomean per workload) speedups over sequential execution.

cs.AR↗

arXiv2024

T3: Transparent Tracking & Triggering for Fine-grained Overlap of Compute & Collectives

Large Language Models increasingly rely on distributed techniques for their training and inference. These techniques require communication across devices which can reduce scaling efficiency as the number of devices increases. While some distributed techniques can overlap, and thus, hide this communication with independent computations, techniques such as Tensor Parallelism (TP) inherently serialize communication with model execution. One approach to hide this serialized communication is to interleave it with the producer operation (of the communicated data) in a fine-grained manner. However, this fine-grained interleaving of communication and computation in software can be difficult. Furthermore, as with any concurrent execution, it requires compute and memory resources to be shared between computation and communication, causing resource contention that reduces overlapping efficacy. To overcome these challenges, we propose T3 which applies hardware-software co-design to transparently overlap serialized communication while minimizing resource contention with compute. T3 transparently fuses producer operations with the subsequent communication via a simple configuration of the producer's output address space and requires minor software changes. At the hardware level, T3 adds a lightweight track and trigger mechanism to orchestrate the producer's compute, and communication. It further uses compute-enhanced memories for communication's attendant compute. As a result, T3 reduces resource contention, and efficiently overlaps serialized communication with computation. For important Transformer models like T-NLG, T3 speeds up communication-heavy sublayers by 30% geomean (max 47%) and reduces data movement by 22% geomean (max 36%). Furthermore, T3's benefits persist as models scale: geomean 29% for sublayers in $\sim$500-billion parameter models, PALM and MT-NLG.

cs.AR↗

arXiv2024

Inclusive-PIM: Hardware-Software Co-design for Broad Acceleration on Commercial PIM Architectures

Continual demand for memory bandwidth has made it worthwhile for memory vendors to reassess processing in memory (PIM), which enables higher bandwidth by placing compute units in/near-memory. As such, memory vendors have recently proposed commercially viable PIM designs. However, these proposals are largely driven by the needs of (a narrow set of) machine learning (ML) primitives. While such proposals are reasonable given the the growing importance of ML, as memory is a pervasive component, %in this work, we make there is a case for a more inclusive PIM design that can accelerate primitives across domains. In this work, we ascertain the capabilities of commercial PIM proposals to accelerate various primitives across domains. We first begin with outlining a set of characteristics, termed PIM-amenability-test, which aid in assessing if a given primitive is likely to be accelerated by PIM. Next, we apply this test to primitives under study to ascertain efficient data-placement and orchestration to map the primitives to underlying PIM architecture. We observe here that, even though primitives under study are largely PIM-amenable, existing commercial PIM proposals do not realize their performance potential for these primitives. To address this, we identify bottlenecks that arise in PIM execution and propose hardware and software optimizations which stand to broaden the acceleration reach of commercial PIM designs (improving average PIM speedups from 1.12x to 2.49x relative to a GPU baseline). Overall, while we believe emerging commercial PIM proposals add a necessary and complementary design point in the application acceleration space, hardware-software co-design is necessary to deliver their benefits broadly.

cs.AR↗

arXiv2023

Computation vs. Communication Scaling for Future Transformers on Future Hardware

Scaling neural network models has delivered dramatic quality gains across ML problems. However, this scaling has increased the reliance on efficient distributed training techniques. Accordingly, as with other distributed computing scenarios, it is important to understand how will compute and communication scale relative to one another as models scale and hardware evolves? A careful study which answers this question can better guide the design of future systems which can efficiently train future large models. Accordingly, this work provides a comprehensive multi-axial (algorithmic, empirical, hardware evolution) analysis of compute vs. communication (Comp-vs.-Comm) scaling for future Transformer models on future hardware. First, our algorithmic analysis shows that compute generally enjoys an edge over communication as models scale. However, since memory capacity scales slower than compute, these trends are being stressed. Next, we quantify this edge by empirically studying how Comp-vs.-Comm scales for future models on future hardware. To avoid profiling numerous Transformer models across many setups, we extract execution regions and project costs using operator models. This allows a spectrum (hundreds) of future model/hardware scenarios to be accurately studied ($<$15% error), and reduces profiling costs by 2100$\times$. Our experiments show that communication will be a significant portion (40-75%) of runtime as models and hardware evolve. Moreover, communication which is hidden by overlapped computation in today's models often cannot be hidden in future, larger models. Overall, this work highlights the increasingly large role communication will play as models scale and discusses techniques and upcoming technologies that can help address it.

cs.AR↗

arXiv2021

Demystifying BERT: Implications for Accelerator Design

Transfer learning in natural language processing (NLP), as realized using models like BERT (Bi-directional Encoder Representation from Transformer), has significantly improved language representation with models that can tackle challenging language problems. Consequently, these applications are driving the requirements of future systems. Thus, we focus on BERT, one of the most popular NLP transfer learning algorithms, to identify how its algorithmic behavior can guide future accelerator design. To this end, we carefully profile BERT training and identify key algorithmic behaviors which are worthy of attention in accelerator design. We observe that while computations which manifest as matrix multiplication dominate BERT's overall runtime, as in many convolutional neural networks, memory-intensive computations also feature prominently. We characterize these computations, which have received little attention so far. Further, we also identify heterogeneity in compute-intensive BERT computations and discuss software and possible hardware mechanisms to further optimize these computations. Finally, we discuss implications of these behaviors as networks get larger and use distributed training environments, and how techniques such as micro-batching and mixed-precision training scale. Overall, our analysis identifies holistic solutions to optimize systems for BERT-like models.

cs.AR↗

arXiv2020

SeqPoint: Identifying Representative Iterations of Sequence-based Neural Networks

The ubiquity of deep neural networks (DNNs) continues to rise, making them a crucial application class for hardware optimizations. However, detailed profiling and characterization of DNN training remains difficult as these applications often run for hours to days on real hardware. Prior works exploit the iterative nature of DNNs to profile a few training iterations. While such a strategy is sound for networks like convolutional neural networks (CNNs), where the nature of the computation is largely input independent, we observe in this work that this approach is sub-optimal for sequence-based neural networks (SQNNs) such as recurrent neural networks (RNNs). The amount and nature of computations in SQNNs can vary for each input, resulting in heterogeneity across iterations. Thus, arbitrarily selecting a few iterations is insufficient to accurately summarize the behavior of the entire training run. To tackle this challenge, we carefully study the factors that impact SQNN training iterations and identify input sequence length as the key determining factor for variations across iterations. We then use this observation to characterize all iterations of an SQNN training run (requiring no profiling or simulation of the application) and select representative iterations, which we term SeqPoints. We analyze two state-of-the-art SQNNs, DeepSpeech2 and Google's Neural Machine Translation (GNMT), and show that SeqPoints can represent their entire training runs accurately, resulting in geomean errors of only 0.11% and 0.53%, respectively, when projecting overall runtime and 0.13% and 1.50% when projecting speedups due to architectural changes. This high accuracy is achieved while reducing the time needed for profiling by 345x and 214x for the two networks compared to full training runs. As a result, SeqPoint can enable analysis of SQNN training runs in mere minutes instead of hours or days.

cs.DC↗

arXiv2017

CODA: Enabling Co-location of Computation and Data for Near-Data Processing

Recent studies have demonstrated that near-data processing (NDP) is an effective technique for improving performance and energy efficiency of data-intensive workloads. However, leveraging NDP in realistic systems with multiple memory modules introduces a new challenge. In today's systems, where no computation occurs in memory modules, the physical address space is interleaved at a fine granularity among all memory modules to help improve the utilization of processor-memory interfaces by distributing the memory traffic. However, this is at odds with efficient use of NDP, which requires careful placement of data in memory modules such that near-data computations and their exclusively used data can be localized in individual memory modules, while distributing shared data among memory modules to reduce hotspots. In order to address this new challenge, we propose a set of techniques that (1) enable collections of OS pages to either be fine-grain interleaved among memory modules (as is done today) or to be placed contiguously on individual memory modules (as is desirable for NDP private data), and (2) decide whether to localize or distribute each memory object based on its anticipated access pattern and steer computations to the memory where the data they access is located. Our evaluations across a wide range of workloads show that the proposed mechanism improves performance by 31% and reduces 38% remote data accesses over a baseline system that cannot exploit computate-data affinity characteristics.

cs.AR↗