Vincent Lannurien
d822cdaa6a
7.5 KiB
title, venue, year, date, tags, paper_count, draft
| title | venue | year | date | tags | paper_count | draft | |||
|---|---|---|---|---|---|---|---|---|---|
| CCGrid 2024 Digest | CCGrid | 2024 | 2024-05-06 |
|
10 | false |
10 papers selected.
Fair, Efficient Multi-Resource Scheduling for Stateless Serverless Functions with Anubis
Amit Samanta 0001, Ryan Stutsman
TL;DR — Anubis introduces a fair, multi-resource scheduler for stateless serverless functions that achieves efficiency without sacrificing isolation between tenants.
Why notable — Fairness in serverless resource allocation is an open problem as functions compete for heterogeneous resources (CPU, memory, I/O); Anubis provides a concrete, deployable answer. The work directly addresses a gap in production FaaS platforms where existing schedulers optimize for throughput but ignore per-tenant equity.
SLO-Power: SLO and Power-aware Elastic Scaling for Web Services
Mehmet Savasci, Abel Souza, Li Wu, David Irwin 0001 et al.
TL;DR — SLO-Power co-optimizes SLO compliance and power consumption during elastic scaling of web services, reducing energy use without violating latency targets.
Why notable — Jointly chasing SLOs and power budgets is a critical concern for sustainable cloud operations; this work shows the two objectives can be reconciled in a single scaling controller. The approach is directly applicable to cloud autoscaling stacks where energy cost and QoS guarantees are both first-class concerns.
HAPPIES: a History-Aware Efficient Cloud Resource Overcommitment System
Ziwei Huang 0003, Shibo Tang, Zihao Chang, Lin Tan et al.
TL;DR — HAPPIES uses historical utilization traces to safely overcommit cloud resources, improving cluster efficiency while bounding the risk of SLO violations.
Why notable — Resource overcommitment is a primary lever for improving datacenter utilization, and HAPPIES advances the state of the art by making overcommitment decisions history-aware rather than reactive. Its deployment-oriented design makes it immediately relevant to hyperscale cloud operators seeking higher bin-packing ratios.
COTuner: Joint Optimization of Resource Configuration and Software Parameters for Recurring Streaming Jobs on the Cloud
Hui Dou, Shanshan Zhu, Yuxuan Zhou 0005, Yiwen Zhang 0001 et al.
TL;DR — COTuner simultaneously tunes cloud resource allocation and application-level software knobs for recurring streaming jobs, reducing cost and latency together.
Why notable — Streaming workloads on the cloud are poorly served by tools that tune resources and software parameters in isolation; COTuner's joint search closes this gap. The focus on recurring jobs makes the approach practical, amortizing tuning cost over repeated executions in production pipelines.
DeepVM: Integrating Spot and On-Demand VMs for Cost-Efficient Deep Learning Clusters in the Cloud
Yoochan Kim, Kihyun Kim, Yonghyeon Cho, Jinwoo Kim et al.
TL;DR — DeepVM dynamically mixes spot and on-demand VM instances to build cost-efficient, fault-tolerant deep learning training clusters in the cloud.
Why notable — Training large models on cloud infrastructure is expensive, and spot instance preemptions are a major obstacle to reliability; DeepVM provides a principled integration strategy that achieves both cost savings and resilience. The approach is practically significant given the rapid growth of cloud-hosted AI training workloads.
Opportunistic Energy-Aware Scheduling for Container Orchestration Platforms Using Graph Neural Networks
Philipp Raith, Gourav Rattihalli, Aditya Dhakal, Sai Rahul Chalamalasetti et al.
TL;DR — A GNN-based scheduler for Kubernetes-style container orchestration platforms exploits opportunistic energy signals to reduce power consumption without degrading application performance.
Why notable — Applying graph neural networks to container scheduling captures the complex topology of cluster resources in a way that heuristic schedulers cannot, while the energy-awareness angle addresses the sustainability imperative facing cloud providers. The work bridges recent ML advances with production container orchestration.
Jingle: IoT-Informed Autoscaling for Efficient Resource Management in Edge Computing
Yixuan Wang, Abhishek Chandra, Jon B. Weissman
TL;DR — Jingle leverages real-time IoT device signals to drive predictive autoscaling decisions at the edge, reducing both over-provisioning and latency spikes.
Why notable — Edge autoscaling is hampered by the lack of load predictors tuned to IoT event patterns; Jingle fills this gap by treating IoT telemetry as a first-class input to the scaling loop. The result is a tighter edge-cloud integration model with direct relevance to smart-city and industrial IoT deployments.
XFBench: A Cross-Cloud Benchmark Suite for Evaluating FaaS Workflow Platforms
Varad Kulkarni, Nikhil Reddy, Tuhin Khare, Harini Mohan et al.
TL;DR — XFBench is a portable benchmark suite that evaluates FaaS workflow platforms across multiple cloud providers using representative workload patterns.
Why notable — The absence of standardized, cross-cloud benchmarks for serverless workflow platforms hinders fair comparison and informed vendor selection; XFBench addresses this directly with a reusable, community-shareable artifact. It covers diverse workflow shapes and exposes platform-specific performance cliffs that single-provider benchmarks miss.
Hades: A Context-Aware Active Storage Framework for Accelerating Large-Scale Data Analysis
Jaime Cernuda, Luke Logan, Ana Gainaru, Scott Klasky et al.
TL;DR — Hades pushes data transformations into the storage layer using context-aware active storage, dramatically reducing I/O traffic and accelerating large-scale scientific data analysis.
Why notable — As HPC datasets grow to petabyte scale, moving data to compute becomes the dominant bottleneck; Hades revives and modernizes the active-storage model with context awareness to match modern workflow patterns. The work is highly relevant to HPC-cloud convergence efforts where storage and compute are increasingly disaggregated.
Workflow Mini-Apps: Portable, Scalable, Tunable & Faithful Representations of Scientific Workflows
Ozgur O. Kilic, Tianle Wang 0001, Matteo Turilli, Mikhail Titov et al.
TL;DR — Workflow Mini-Apps are compact, parameterizable proxies that faithfully capture the performance behavior of full scientific workflows, enabling portable scheduling research without the overhead of running complete pipelines.
Why notable — Scientific workflow scheduling research is bottlenecked by the cost and complexity of running real applications at scale; mini-apps lower this barrier while preserving the key performance characteristics needed for valid scheduler evaluation. The methodology is immediately reusable by the broader distributed workflow community.