ECHO: The 1st International Workshop on Edge-Cloud-HPC Operational Continuum

Camera-based smart-agriculture systems capture visual evidence of weeds, diseases, and pests, but multi-megapixel images are costly to transmit over low-power field networks. Many systems therefore perform on-device analysis and transmit only compact summaries, preventing remote visual verification. Image compression can reduce this burden, but must preserve information needed for downstream analysis. We present a preliminary task-level study using four annotated sugar-beet field images compressed with JPEG and JPEG 2000 over a 2–16x range, rerunning the same species-level weed detector on each reconstruction. Results show that lossy compression can substantially reduce image size while retaining useful detection performance, with JPEG 2000 consistently outperforming JPEG at the task level. We also find that similar PSNR values do not necessarily imply similar detection utility. However, payloads remain costly even at compression ratios that preserve detection quality, motivating agriculture-specific compression methods that more aggressively reduce data while protecting task-relevant visual information.

The autoregressive nature of large language model (LLM) decoding results in matrix multiplications with small token dimensions, which underutilize systolic arrays and exhibit low arithmetic intensity. Efficient execution of these workloads requires jointly improving processing-element utilization and coordinating data movement with computation under limited on-chip memory. We present PerfDecode, a performance-model-guided approach to accelerating LLM decoding on systolic arrays. PerfDecode reorganizes matrix multiplication execution to exploit idle processing elements and pipeline data movement with computation. At its core, an analytical performance model captures DMA transfer and launch overhead, systolic-array computation, reduction and write-back costs, and on-chip memory constraints to automatically determine blocking configurations for different matrix shapes. On AWS Trainium2, PerfDecode consistently outperforms tiled matmul across diverse decode operators in Llama-3-8B and Llama-3-70B, achieving geometric-mean speedups of 2.73x and 2.85x, respectively.

Edge deployment of LLMs requires multiple devices working together, since no single device matches a cloud server's range. Existing edge routers treat each query as one unit and use fixed, offline-trained policies that break down as workloads change. We present SPEAR (Semantic Partitioning and Edge-Adapted Routing), which routes work at the subtask level and keeps adapting after deployment. We split each query into subtasks, send each to an edge device by capability score and risk threshold, and fall back to a cloud model only when confidence is too low. We also update each device's LoRA parameters from cloud feedback, so edge models and the router keep improving without cloud-side retraining. On a testbed of Raspberry Pi 4s, Jetson Orins, and an RTX 4090 server, SPEAR reaches 94.7% routing accuracy, beating baselines like RouteLLM, with only 32.4 ms overhead; local adaptation raises Banking77 accuracy from 31.0% to 56.5% over five rounds.

Memory over-provisioning results in resource underutilization when HPC workloads run on Kubernetes. The default Vertical Pod Autoscaler (VPA) cannot anticipate phase-driven memory spikes for first-run HPC jobs. In this work, we present a reinforcement learning (RL) recommender that formulates vertical memory scaling as a Markov Decision Process and trains a Proximal Policy Optimization (PPO) agent on 3,353 real Prometheus traces. Evaluated on a live Google Kubernetes Engine (GKE) cluster using LAMMPS, graph analytics, in-memory analytics, and MLPerf 3D-UNet, the RL agent reclaims 31.6% of the available memory headroom and incurs at most one OOM event, while VPA's static provisioning does not adapt to workload phases and would have been insufficient to avoid OOM in 29 runs.

LLM compression for cloud and HPC settings assumes diverse, unpredictable workloads, but edge deployments typically serve a single fixed task: unlike cloud inference, which is bottlenecked by attention and KV-cache transfer, edge inference is bottlenecked by MLP layers. We further find that layer and weight importance carry task-specific sensitivity signatures: a layer or channel prunable for one task can be critical for another. Building on these insights, we present a task-aware compression framework, PRISM, for edge-cloud LLM serving. The cloud profiles per-task layer and weight sensitivity via causal ablation and AWQ-style quantization error, encoding results as an execution layer mask and a precision weight mask woven into a task-specialized inference DAG. On the edge, a sliding-chunk scheduler streams weights with the minimum resident window that fully overlaps prefetch with computation. Evaluated on SmolLM2 and Qwen2.5, our pipeline cuts peak memory by 2.9x and end-to-end latency by 26% versus a dense baseline.

The United Kingdom's AI Research Resource (AIRR) opens national supercomputers to a community familiar with always-reachable cloud services. With batch-scheduled HPC, a user requests a time-bounded allocation that is then reclaimed. Federated Learning (FL) exposes this tension, pairing a long-lived control plane with short-lived training. We characterize the resulting placement asymmetry: the coordination server sees only model updates and control metadata, so it can sit on neutral infrastructure, whereas the client control process is bound to the participating site's data and must remain within its trust boundary. The outbound-only connection this implies is also how edge clients connect, though the batch-scheduler mismatch is specific to HPC. We demonstrate a cross-silo federation spanning three independently administered systems (Isambard-AI, Isambard 3, and Zenith) using unmodified production FL frameworks (FLARE-wrapped Flower), measure the cost the client placement incurs, and propose Kubernetes-as-a-service as the in-boundary home for the persistent process.

Federated Split Learning (FSL) has emerged as an effective paradigm for mitigating the computational burden of edge clients in classical federated learning while preserving data privacy. However, applying FSL to large language model (LLM) fine-tuning introduces several critical challenges. In particular, cut-layer selection must be adaptively configured across clients to accommodate heterogeneous data distributions and device capabilities, which can significantly affect system performance. Moreover, reducing communication overhead during fine-tuning remains an important yet underexplored problem. We propose SplitFT, an edge-cloud collaborative federated split learning system for LLM fine-tuning. SplitFT allows clients to select different cut layers according to their computational resources and model performance, and further reduces communication costs by decreasing the LoRA rank at the cut layer. To better emulate real-world data heterogeneity, we introduce a length-based Dirichlet partitioning strategy. Extensive experiments on popular benchmarks demonstrate that SplitFT outperforms state-of-the-art methods in both fine-tuning efficiency and model performance.

Serverless computing offers HPC workflows elastic capacity, but only some tasks fit its runtime and memory limits, and finding them has required manual placement. We separate two questions that manual placement conflates: deployment-independent candidacy, which combines platform limits with task-level heuristics and is settled before execution, and deployment-dependent placement, which turns on how much workflow parallelism the local allocation absorbs. Our framework combines static DAG analysis with one bounded profiling run per task type to assign each a candidacy class, then splits candidates across local and serverless backends without per-task input. In a single-run campaign on Epigenomics, Seismology, and Cycles workflows, hybrid execution reaches the lowest observed makespan of three strategies, but the margin tracks how capacity-constrained each workflow is: 65.0% on Seismology against 4.9% and 9.0% on the others. Withholding the task types the analysis rejects lowers observed makespan a further 9.8–39.9%, and dispatch costs a median of 221 microseconds.