-
Elastic Processing of Analytical Query Workloads on IaaS Clouds — Kllapi et al., 2015
- Key idea: Designed an elastic query processing engine for cloud infrastructure — dynamically scaling compute resources based on workload demand. Optimizes cost vs. latency trade-offs for analytical queries.
- 🔗 Download (search SIGMOD)
-
Exoshuffle: An Extensible Shuffle Architecture — Luan et al., 2023
- Key idea: Redesigned the shuffle stage of distributed data processing (the most expensive bottleneck in MapReduce/Spark). Achieves 2-3× speedup through better scheduling and data placement.
- 🔗 Download
-
DeCaf: Iterative Collaborative Processing over the Edge
- Key idea: Distributed edge computing framework that splits neural network inference across multiple edge devices, collaboratively processing data closer to the source.
- 🔗 Download (search SIGCOMM 2021)
-
Straggler-Robust Distributed Optimization in Parameter Server Networks — Atallah et al., 2021
- Key idea: Addressed slow workers (stragglers) in distributed training using coded computation — adding redundancy so the aggregation can proceed without waiting for all workers.
- (Also listed under Federated Learning)
- 🔗 Download
- Shuffle operations (redistributing data between map and reduce phases) are the dominant cost in distributed data processing.
- Elastic scaling (cloud) vs. fixed topology (edge) present different optimization challenges.
- Coded computation adds redundancy to tolerate slow or failed workers — a key technique for reliable distributed training.