AMPC
summary
Introduces AMPC = MPC + distributed hash table and claims adaptive reads let many graph tasks drop from polylog rounds to constant rounds by doing short sequential RDMA lookups. Implements the system and shows real runs (up to 226 billion edges) with measurable speedups versus baseline MapReduce-style pipelines, relying on 1–3 microsecond lookups and read-only per-round fault tolerance. Notes limits: single-machine RAM still beats distributed I/O for some workloads, and capacity cliffs, skew, and write semantics remain open practical and theoretical issues.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:30Talk introduction
- 2:35Motivation for graph algorithms
- 7:32MapReduce and Pregel overview
- 19:48MPC model introduction
- 36:57Adaptive Massively Parallel Computation model
- 51:56AMPC implementation results
- 1:06:52Why distributed computation matters
- 1:12:54Future directions
related talks
14mArchTM: Architecture-Aware, High Performance Transaction for Persistent Memory
01h 12mData parallelism from a multicore perspective
05mAn Introduction to hpxMP...
12mSPHT: Scalable Persistent Hardware Transactions
14mNap: A Black-Box Approach to NUMA-Aware Persistent Memory Indexes
21mModel Checking C/C++ with Mixed-Size Accesses
01h 02m“Beyond the Issaquah Challenge: High-Performance Scalable Complex Updates"
52mSecrets of High Performance in Multicore Systems ☠️