OSDI '21 / USENIX ATC '21
2021-07-14..2021-07-16
-- online --
who it's for
systems researchers, machine learning systems engineers, storage engineers, cloud infrastructure engineers, networking engineers, operating system developers, security and privacy engineers, verification engineers
topics
serverless scheduling and placement; distributed hash table overlays; graph neural network training; in-storage ML accelerators; persistent memory (Optane) optimizations; RDMA and RoCE deployments; LSM compaction versus unclustered logs; huge page allocators (2 MB); kernel-level networking (Homa); mountable Merkle trees for enclaves; autoscaling inference runtimes; verified concurrent journaling
| title | author | length |
|---|---|---|
| Preview: Cloud Computing | Kshiteej Mahajan | 16m |
A set of systems and findings target real limits: DART replaces a centralized master with a distributed hash table overlay to eliminate queuing and long‑tail… | ||
| Preview: Graph NNs | Anand Iyer | 16m |
Overview of graph representation learning and GNNs with concrete system results: shallow encoders are transductive lookups while deep GNN encoders are… | ||
| Preview: Hardware and Offloads | Ketan Bhardwaj, Sudarsun Kannan, Michio Honda | 22m |
Four systems claim concrete wins: a prediction-based scheduler integrates ML into Azure VM placement and shows twice the oversubscription capacity without… | ||
| Preview: KV-Stores and Databases | Alex Conway | 18m |
Compares LSM trees and unclustered-index logs by showing when moving data helps: LSM compactions improve read locality at the cost of write amplification from… | ||
| Preview: ML Serving | Arpan Gujarati | 12m |
Three systems target cloud-scale DNN inference with concrete claims and prototypes: PET compiles tensor programs using partially equivalent transformations… | ||
| Preview: ML Training | Deepak Narayanan | 13m |
Survey of ML training papers summarizes concrete wins and techniques: the compute curve reaches 8 petaflop‑years for the largest GPT‑3 model; zero offload… | ||
| Preview: Networking | Akshay Narayan | 12m |
Explains that RDMA lets a NIC write directly into memory so the CPU does no packet processing, lowering CPU overhead but requiring fancy NICs and a carefully… | ||
| Preview: Operating Systems | Adil Ahmad | 21m |
Shows AVX vector code (256/512-bit AVX) forcing cores into lower clock states and proposes scheduler detection plus compensation for hyperthread victims… | ||
| Preview: Persistent Memory | Ian Neal | 17m |
A concise tour of modern persistent memory: history plus actionable system results and tools. Claims and numbers: Optane Series 100 ~300 ns random reads… | ||
| Preview: Security and Privacy | Chia-che Tsai | 17m |
Preview lists concrete systems and attacks: papers use fully homomorphic encryption and garbled circuits but move heavy state and page faults into novel… | ||
| Preview: Storage | Saurabh Kadekodi | 27m |
Survey of storage work covers ZNS device interfaces that cut the block‑interface tax and raise throughput/capacity; an enhanced ZNS offloads compaction to… | ||
| Preview: Verification (Correctness) | Chengcheng Wan, Lefan Zhang | 18m |
Five papers show concrete claims and artifacts: data-driven environment synthesis uses protocol traces to prune and order invariant candidates so fewer solver… | ||
| Preview: Verification (Correctness & Debugging) | Chengcheng Wan, Lefan Zhang | 18m |
Five systems present concrete results: Pilot applies Python metadata, bytecode edits and dynamic profiling to patch online services without restarts; a fuzzing… | ||
| nothing matches | ||