Link failures cap LEO capacity scalability at O(1/n)
Under a Markov model of ISL states, scalability rises to an optimum then falls toward zero as satellite count grows.
· “Capacity Scalability of LEO Constellations With Dynamic Link Failures”
cs.DC
Distributed, Parallel, and Cluster Computing
Covers fault-tolerance, distributed algorithms, stabilility, parallel computation, and cluster computing. Roughly includes material in ACM Subject Classes C.1.2, C.1.4, C.2.4, D.1.3, D.4.5, D.4.7, E.1.
sort pith recommended most recent
Static tile taskflow overlaps MoE compute and comms on NPUs
HyperParallel-MoE converts serial operator runs into a fixed heterogeneous schedule, cutting Dispatch-to-Combine latency up to 1.58x inside
· “HyperParallel-MoE: Multi-Core Interleaved Scheduling for Fast MoE Training on Ascend NPUs”
EdgeServing cuts SLO violations for multi-DNN edge serving
A stability score and early-exit choices expand the scheduler's options to lower deadline misses and P95 latency on shared GPUs.
· “EdgeServing: Deadline-Aware Multi-DNN Serving at the Edge”
Simulation platform tests datacenter power flexibility for grid coordination
OpenG2G connects real AI service measurements to grid models so users can compare controllers and measure impacts of model choices.
· “OpenG2G: A Simulation Platform for AI Datacenter-Grid Runtime Coordination”
Consensus-embedded checks give order-execute chains 10.6x throughput
Flexible per-contract endorsement is added by folding checks into consensus, avoiding extra messages and abort spikes in contested workloads
· “Back to the Future: Rethinking Endorsement in Order-Execute Blockchains”
Async GPU kernels speed up sparse matrix multiplies by up to 6x
BCSR and WCSR formats overlap TMA transfers with WGMMA computation to beat cuSPARSE on benchmarks and deliver 2.66x end-to-end LLM prefill.
· “AsyncSparse: Accelerating Sparse Matrix-Matrix Multiplication on Asynchronous GPU Architectures”
Dynamic grouping plus TEE cuts blockchain consensus messages
T-RBFT runs efficient Raft inside groups and TEE-BFT across them, raising throughput for permissioned chains.
Adaptive quantization cuts mobile LLM cold starts by 4x
EdgeFlow lowers precision on less critical weights to speed loading from flash while preserving accuracy on phone NPUs.
Disaggregating LoRA triples request rate under latency limits
Separating adapter execution from base models cuts memory costs and lets more requests meet strict response time targets.
· “InfiniLoRA: Disaggregated Multi-LoRA Serving for Large Language Models”
Sketch-based GPU solver handles 5000-asset portfolios in seconds
Nesterov-accelerated projection with subspace embeddings preserves accuracy while cutting full-model runtime from over a minute to under 3 s
Twin-field QKD secures blockchain with linear scaling
Hybrid architecture uses MDI star topology to integrate information-theoretic security into distributed consensus.
General adversaries make consensus pay an n^{1/d} tax
Projective-geometry adversary families lift the Ω(Ln) baseline to Ω(L n^{1+1/d}) bits for agreement and broadcast.
· “Multivalued Consensus: General Adversaries Require More Communication”
1-PLS of cost p yields t-PLS of cost p/t up to log n
Near-resolution of the tradeoff conjecture shows how larger neighborhoods reduce label sizes in proof labeling schemes for general and minor
· “Near-Resolution of the Tradeoff Conjecture in Distributed Proof Labeling Schemes”
Ringmaster LMO recovers optimal async time complexity for LMO
Extending delay-thresholding to LMO momentum handles heterogeneous workers and matches best known bounds in smooth cases.
· “Ringmaster LMO: Asynchronous Linear Minimization Oracle Momentum Method”
LCL complexity on trees shifts without exact n knowledge
Oblivious nodes or polynomial bounds on n let randomness help more problems and create unnatural complexity classes.
· “The Distributed Complexity Landscape on Trees Depends on the Knowledge About the Network Size”
3-process broadcast consensus gets exact error floor
For message-success probability p, no one-round algorithm can beat 2p²q+q³ for p≤2/3, nor q above it.
Adaptive clustering lifts wireless delivery by up to 9.8 points
Grouping peers by reliability and coding within and across clusters cuts latency 10-23% and raises retention up to 30%.
RNN load balancer cuts multi-GPU tissue-sim imbalance to 3.5%
Trained on synthetic workloads, it migrates 7.7x fewer slices than reactive balancing at similar runtimes.
· “Scalable Multi-GPU Simulation of 3D Multicellular Growth with RNN-Based Workload Balancing”
Learned aggregation matches FedAvg and Ditto on non-IID FL
Server and client agents jointly learn aggregation weights and personalization without pretraining.
Cloud docs alone can synthesize emulators that beat hand-built rivals
A neurosymbolic pipeline reads provider documentation, generates state-machine code, and aligns it against the real cloud.
Distributed AI training time predicted within 15% error
A lightweight model maps sharded and geo-distributed transformer training to deployment plans, cutting cost 21.7% at 99% of peak throughput.
· “ShardMeter: Sharded and Geo-Distributed Training Without the Guesswork”
8.6 GiB over budget, an 80B MoE still emits byte-exact output
Demanded weights beat the 34 GiB envelope; all 64 logit rows still match a zero-cache oracle byte for byte.
· “Memory-Sovereign Inference: Output-Exact Execution Beyond Full Residency”
Critic-guided GAN lifts molecule QED by more than 10%
A federated graph GAN that scores molecules on user-defined metrics keeps datasets private while targeting drug-likeness.
Turn off thread affinity: imbalanced jobs scale better on hybrid CPUs
On chips with fast and slow cores, letting the OS scheduler move threads beats pinning them for uneven workloads.
· “Effects of Hybrid CPU and Cache Architectures on Parallel HPC and Cloud Applications”
Microservice timing is emergent: component guarantees won't compose
Local delays in service-based vehicles spread through shared resources and control loops that timing models overlook.
· “A Survey of Timing Variability in Microservice-Based Software-Defined Vehicles”
Sleeping networks still self-heal at low awake cost
Coloring, independent sets, and matchings each need only O(Delta) or O(Delta^2) awake rounds per node.
CED-EF reaches centralized speed with one compressed vector
Even biased Top-k sparsification keeps the centralized O(σ/√(NK)) rate, after a transient N^3/(δ^4Δ^4).
· “CED-EF: Compressed Exact Diffusion with Error Feedback for Multi-Agent Learning”
RISC-V board beats Xeon CPU on Nekbone kernel at 7x lower power
Nekbone's AX kernel hits 242.97 GFLOPS on Tenstorrent's Wormhole at about 7x less power than the Xeon CPU.
· “Exploring spectral element methods on the Tenstorrent RISC-V accelerator”
Delta prefetching runs oversized LLMs 12x faster on edge
Models that exceed resident memory fetch only the newly active neurons per token, holding 92–96% of dense accuracy.
· “NeuroPrefetcher: Storage-Aware Sparse LLM Inference via Delta Prefetching”
Share latent codes, not model weights, to cut FL bandwidth 99%
GCA matches FedAvg accuracy while making server-side data extraction measurably harder.
One shared model for every honest node
Consensus on which updates to aggregate removes divergence and blunts poisoning without a central server.
· “Model-Consistent Byzantine-Resilient Decentralized Federated Learning for Collaborative Missions”
Bigger GPU clusters may need less interconnect bandwidth
Barrier waits from matrix-math timing variation grow with cluster size and cap the payoff of faster GPU links.
· “Understanding the Synchronization Tax in GPU Scale-Up Domains”
Concurrent objects can hide data—until domains go unbounded
Epistemic logic sets the bar: bounded registers stay leak-free; unbounded max registers cannot.
Exact payment-channel path choice now sub-100ms on 12,000-node networks
A quadtree pruning scheme finds the optimal virtual payment path in milliseconds without sacrificing exactness.
· “Scalable Exact Path Selection via Structure-Aware Search for Virtual Payment Channels”
Address preallocation cuts 80,000-pod deploy from 2 hours to 8 minutes
Prefix and ENI preallocation match overlay speed while keeping VPC-native pod addresses and flow-log visibility.
· “Fleet-Scale Pod Deployment with VPC-Native Networking in Managed Kubernetes”
Deeper splits boost privacy but destroy fine-tuning quality
The cut layers that maximize speed and privacy are exactly where model quality collapses, for every aggregation method.
· “Unveiling the Depth-Performance Dilemma in Split-Federated Fine-tuning of LLMs”
Map of 30+ consensus protocols shows liveness lags
Safety proofs dominate formal verification; liveness, code-level assurance, and scale remain the field's weak points.
· “Systematization of Knowledge: Formal Verification of Consensus Protocols”
Edge AI energy cut 36% by per-request model and CPU choice
The method picks the cheapest model and CPU allocation per request while meeting deadlines and quality targets.
· “PRISM: Predictive Runtime In-place Scaling and Model Selection for Edge Microservices”
Do-nothing test finds a 27-swap route on a radius-10 chip
Two benchmark numbers separate uniform trust from best-route luck on tested quantum hardware.
Multi-GPU FIFO queue on four A100s: up to 3.92x speedup
Broker-queue design over NVSHMEM also lifts Bellman-Ford to 3.03x max speedup on 10 graphs.
· “A Concurrent Queue System for Multi-GPU Platforms: Application to Bellman-Ford SSSP”
Sparse graph cuts point-cloud registration latency 2–3×
FlashReg keeps recall on par with TurboReg while halving peak tensor memory on embedded GPUs.
Tsunami waveform ensembling drops to 24 ms per rupture
FFT convolution on 64 GPUs cuts the source-to-waveform map to two matvecs, making 10,000-scenario ensembles a 4-minute job.
Power-aware serving rides caps to 1.64x baseline goodput
A KKT-driven runtime spends contracted latency slack on memory-bound stages, keeping tails near 1.3x nominal.
· “PowerSlider: Exploiting Phase Asymmetry for LLM Serving under Demand Response”
CoT stage cues guide expert caching for 1.33x faster MoE inference
By grouping experts per reasoning stage, SAEM cuts GPU-CPU transfers and boosts throughput 1.33x on average.
Thermo-FL keeps 21% accuracy under attack
Device heat throttles edge AI; Thermo-FL trims workload when hot and filters poisoned sparse updates.
· “Thermo-FL: Thermal-Aware Robust Federated Fine-Tuning of Large Language Models for Edge AI”
Batched lock-free repair makes GPU vector indexes stream-friendly
GrAND updates Vamana and CAGRA in place, beating SVFusion and FreshDiskANN-GPU by 2.2x–8.7x and up to 25x.
· “GrAND: GPU-based Dynamic Graph Indexes for Approximate Nearest Neighbour Search”
Plan first, tune later: LLM kernels reach 92.4% validity
Workload-aware choice among PyTorch ops, CUDA libraries, and custom CUDA beats fixed spaces under an 18-candidate budget.
· “HIERA: Workload-Aware Planning Across Implementation Spaces for GPU Kernel Optimization”
Python core beats Fortran+OpenACC by 20-30 percent
GT4Py and DaCe generate optimized GPU code from one Python source, with zero copy overhead in the legacy Fortran model.
Even 15% malicious peers break backdoor defenses in decentralized FL
A 13-defense benchmark finds Byzantine-robust methods fail on heterogeneous data and depend heavily on network topology.
· “BackDFL: A Unified Benchmark For Backdoor Attacks and Defenses In Decentralized Federated Learning”
Heat and energy, not radiation, cap orbital AI compute
Why half a year of CubeSat telemetry says orbital schedulers should plan around heat, sunlight, and battery state
One triangular solve replaces per-node state snapshots
For gated-delta hybrids, verifying a draft tree no longer needs a state snapshot per node, and freed HBM buys throughput where memory binds.
· “TreeWY: Speculative Verification for Gated DeltaNet Hybrids”
Pure Fortran runs a production solver on NVIDIA
POT3D scales across multi-GPU nodes using only standard 'do concurrent' loops and unified memory.
Network bandwidth rules Earth-observation storage up to 10 Gbps
Same servers at 1, 10 and 100 Gbps show throughput tracks link rate below 10 Gbps, then memory topology takes over.
Solana rug pulls predicted from first five minutes of trades
A 6.4M-token study shows liquidity and trade features alone carry early warning signal—no contract code needed.
· “Catching the Rug: Early Prediction of Fraudulent Memecoins on Solana via Machine Learning”
Every retrieved answer comes with a replayable audit proof
A committee's private scores are bound into digests any client can replay to check ranking.
· “TrustRAG: Blockchain-Enhanced RAG via Committee-Based Credibility Scoring”
Ten parallel patterns, one API: up to 260x GPU speedup
CPU workers reach 11.6x at 16 threads; WebGPU tops 260x on compute-bound workloads.
· “ParaWeb: Parallel Programming Patterns for Web Development”
Planned prefix affinity lifts LLM serving 2.3x over baselines
KV-cache hits jump from 64% to 93% at a 3.5-second tail SLO without changing model execution.
· “CacheRoute: Planned Prefix-Affinity Routing for Large-Scale LLM Serving”
Profiling that targets the fleet decision saves 6.9% GPU time
FleetSieve measures only cells that could shift the allocation, stopping once conservative and optimistic plans agree.
· “FleetSieve: Decision-Critical Profiling for SLO-Aware LLM Fleet Configuration”
LLM agents earn their cost only when demand surges open headroom
A controlled test shows the LLM layer gains a small but significant edge only when a mid-run surge creates headroom
30% context retention cuts edge RAG GPU energy by up to 53%
Measurements show generation dominates edge-RAG cost and mild compression loses energy, so rates should be set at runtime.
· “From Retrieved Context to Runtime Control: Adaptive Compression for Edge-based RAG”
Lazy image pulls can corrupt model files while health stays green
A 196-second window hides real read failures behind Ready and health checks; cache sizing is the fix.
Two AI PCs serve two users at 1.79x one machine's pace
Sharding a model across idle Intel laptops plus speculative decoding beats the unsplit baseline on the same hardware.
· “Pre-Compiled Pipeline Shards for Distributed LLM Inference on Intel AI PC Fleets”
Three tiers defend V2X cars inside a 100-millisecond window
A proposed architecture splits each safety-message verdict among the car, roadside edge, and cloud, biasing ambiguity toward escalation.
· “Autonomous Cyber Defense in Connected Vehicles: A Multi-Agent Approach to V2X Security”
Reusing pipeline pieces cuts federated data-sharing runtime by a third
Shared transformations and cached intermediates reduce duplication, energy, and wall-clock time in a federated trial.
· “Enabling Reuse for Data-Sharing Pipelines in Federated Environments”
LayerNorm-parameter shifts expose poisoned clients in federated LLMs
A server-only screen of normalization-weight changes beats six baselines, cutting test perplexity by up to 18.6%.
Autoscaler cuts SLO violations by 56% while using fewer replicas
Gating scaling decisions on forecast confidence and bottleneck location trims replica counts by up to a third.
· “SLO-Scaler: Uncertainty-Aware SLO-Driven Autoscaling for Microservices”
Event-driven microservices cut estimated energy 41% in big-model runs
For one large environmental workflow, microservices beat the monolith on energy despite running longer.
· “When Do Microservices Save Energy? Evidence from Environmental Simulation Workflows”
LLM cache restore pinned at 500 MiB
Probing the full prefix, installing in fixed windows: local restore staging stays flat as external state grows 32x.
· “Bounded-State Restoration: Decoupling Local Restore Capacity from External LLM State”
10-number code keeps robot formations converging when Wi-Fi fails
In overloaded Wi-Fi tests, uncompressed DMPC received 0.00% of messages; the LSTM code received 98% and converged.
· “Communication Reduction via Semantic-Based Encoding in DMPC Using LSTMs”
Only 5 of 19 high-altitude platform functions are flight-proven
A strict evidence rule shows the real edge is persistence at close range, not altitude.
Signed state deltas cut LLM triage tokens by 88.8%
A hash-chained telemetry ledger also cuts wire payloads 96.4% and lets agents prove an event never happened.
· “Agent-Native Telemetry: Verifiable State-Delta Evidence for Autonomous Operations”
Bandwidth ratio puts 753B models on a single GPU
FreeToken divides expert cache misses between PCIe and CPU by two measured bandwidths, reaching 1.5–2.3x faster edge decode.
· “FreeToken: Efficient Edge-Native MoE Serving with Bandwidth-Adaptive Execution”
Work-stealing pool triples labeling throughput under skewed load
Single-file scheduler matches static sharding at zero skew and recovers all 2,000 tasks when half the workers die.
Base-plus-log trick cuts linear-attention decode writes 7.83x
Deferring dense recurrent-state updates to periodic merges speeds decode kernels up to 1.86x without changing the model.
· “DeltaLog: Deferred Materialization of Recurrent States for Linear Attention Decoding”
Fusing outlier and quantized paths speeds LLM decoding up to 4x
One kernel reuses activation tiles on chip, beating unfused outlier-aware baselines by up to 1.5x.
· “FlashQuant: Sparse-Dense Fusion for Memory-Efficient Outlier-Aware LLM Inference”
A query alone can start the KV-cache sweep
Reading it one feed-forward early overlaps both devices — at no measurable training cost.
· “Q-First: Most of Attention Needs Only the Query in Disaggregated LLM Decoding”
eAVID halves dispersed storage once all N nodes confirm
Each node drops one of its two fragments on its own once every node has confirmed, and the message stays recoverable.
· “eAVID: Asynchronous Verifiable Information Dispersal with Post-Dissemination Pruning”
Latency cut 2.9 times by computing only where it matters
Planner-critical objects keep their recall; the rest loses compute, not accuracy.
· “MM-BEV: Enhancing Timeliness by Computing Where and When it Matters”
TERRA trains an 11.4B ocean model on 96 GPUs
Hybrid parallelism and rollout-aware memory offloading cut peak GPU memory by up to 52 percent.
Scheduler that switches token budgets beats any fixed batch size
P-PAS uses a large prefill budget under light load and a small one under pressure, beating static choices.
· “P-PAS: Prefill-Pressure Adaptive Scheduling for Long-Context LLM Serving”
