トップ差分一覧 Farm ソース検索ヘルプ PDF RSS ログイン

Diary/2019-4-14

昨日は昼間いい感じに目が覚めててよかっと思ったものの，
夕方激しい睡魔におそわれ睡眠...からの深夜に起床．
ワークショップの一部がハンズオン的な感じだったこともあって
割と一日有意義に起きてられてよかった．

What are architectural and research challenges for realizing specialized edge computing?

two-tier: cloud <-> IoT device
three-tier: cloud <-> Edge node <-> IoT device, cloud <-> cloudlet <-> IoT device,
- cloudlet: edge data center, distributed edge cloud
two-tier specialized edge: edgenode(+ VPU/TPU) <-> IoT devices
- specialized edge nodes
- accelerate specific workloads
- "server-class" performance
- Little/no cloud reliance
two-tier specialized edge variants: edgenode(+ VPU/TPU) <-> IoT devices(with accelerator)
- tensorflow board, GAP8 IoT processor
Research challenge
- specialized edge can only run a single class of application
  - lower hardware reuse across application classes(no multi-tenancy)
  - multiple specialized hardware configurations needed to support different application classes
  - - increases hardware costs and management complexity
Cloud vs. Edge economics
- cloud: greater multiplexing benefits
- edge: smaller number of servers per site - lower smoothing: reduces multiplexing benefits
- lower ecnomy of scale for edge clouds
Hardware Heterogeneity
split application processing
- application needs to be distributed across tiers - what function to put where?
Challenges
- greater hardware complexity
- greater application complexity
Macroprogramming
- origins in sensor networks (circa 2005)
- specify aggregate system behavior rather than device behavior
  - hides hardware diversity from programmers
  - write onece, run anyware
cf. http://lass.cs.umass.edu/

Video analytis towards vision zero

Democratize ideo analytics
- real-time, low-cost, accurate
video analytics at scale with approximation, NSI17, SIGCOMM18, SEC18 OSDI18
- video pipeline optimizer sigcomm18
- resource manager nsdi17
- edge/cloud executer sec18
- camera manager ipsn18
- video event store osdi18
low-cost ingestion: chaper CNN
- cheap CNNs are less accurate
- cheap CNNs can achieve high recall with small top-K results
- -> solution: top-k approximate index
low-latency query: redundancy elimination
cf. https://www.microsoft.com/en-us/research/project/live-video-analytics/
cf. https://github.com/antriv/MLADS_FALL_GAN_2017/tree/master/ppt