1. Progress Report
2016/12/28

2. ICPADS’16 Keynote
Database Meets Deep Learning: Challenges and Opportunities
IoT: Towards a Connected Era—Research Direction and Social Impacts
On Sensorless Sensing for the Internet of Everything
Theory and Optimization of Multi-core Memory Performance

3. Database Meets Deep Learning: Challenges and Opportunities
Deep Learning(DL) is difficult to:
Understand (its effectiveness)
Tune the hyper-parameters
Optimize (the training speed and memory usage)
Apply database techniques to DL
For efficiency optimization
For memory optimization
Why?
Database operations are “predictable”.

4. Efficiency Optimization
To improve the training speed of DL on a single device(GPU or CPU device)
All operations of one iteration create a dataflow graph.
Existing systems either do static(Theano and TensorFlow) or dynamic(MxNet) dependency analysis to parallelize operations without data dependencies.
Possible improvements:
When there are limited resources, e.g. executors(CUDA streams), there could be multiple ways of placing the operations onto the executors.
Runtime optimization by 1) collecting the cost(i.e. FLOPS) of each operation and hardware statistics 2) estimating the total cost of all plans.

5. Memory Optimization Optimize memory footprint for DL Memory is used by
Parameter values
Parameter gradients
Hidden layer values
Hidden layer gradients
GPUs and other new hardware have less memory than CPUs.
Memory pool and garbage collection
Release memory if it is not used in the rest of one iteration.
Frequent malloc/free is costly for GPUs. Memory pool could help.
Swap data between GPU and CPU memory
Statically done by users.
Dynamically done by system using the paging strategies?
Drop some hidden layer values and redo the computation to recover them.
Statically analyze the dataflow to decide which one to drop.
Dynamically determine which one to drop and recover it when necessary or in advance? Need to consider which layer to recovery and estimate the best time to recover.

7. Summary Deep learning is effective but difficult to optimize.
DB optimization techniques could be adapted for DL systems.
DL model can be applied to DB applications.
Find optimal query plan.
Apache SINGA:
A distributed platform for DL.
Optimizing speed, memory for single node.
Optimizing communication/consistency for distributed training.
DLaSS easy deployment and visualization.

8. IoT: Towards a Connected Era—Research Direction and Social Impacts
Network society = Ubiquitous society
From vertical integration to horizontal integration.
IoT can help:
Elderly care, traffic for elderly, local industry, agriculture revolution, smart city, …
Toward a better society with open data.

10. Core Technology Big data processing Edge Computing
Pushing the frontier of computing applications, data, and services away from centralized nodes to the logical extremes of a network.
Enables analytics and knowledge generation to occur at the source of the data.
Edge access network construction

11. Relay-by-Smartphone
Delivering Messages Using only Wifi without Celler Coverage 1) 2) 3) 4)

12. On Sensorless Sensing for the Internet of Everything
The barriers for long-term and large scale sensor networks and IoT:
1) Hardware cost and its maintenance
2) Energy
Possible solutions:
1) Crowd-sourcing and participatory sensing.
2) backscatter to solve the energy problem.

13. Sensorless Sensing Applications: Benefits: Intrusion detection
Healthcare monitoring
Human-Computer Interaction
Benefits:
Wireless sensing without wire.
Contactless sensing without wearable sensors.
Sensorless sensing without dedicated sensors.

14. Projects CitySee 2011-ongoing 签信通 2011-ongoing
GreenOrbs (绿野千传) 2009-ongoing
Trustworthy RFID (签若磐石) 2007-ongoing
OceanSense, Sensor Network for Sea Monitoring
SASA, Coal Mine Monitoring with Sensors

15. Theory and Optimization of Multi-core Memory Performance
Memory is multi-layered, hierarchical and shared.
Caching is dynamic sharing of fast memory.
Parallel programs / workloads
Large cache
Run-time mediation between demand and supply
Data-ranking / priority allocation
Similar to the role of an OS scheduler for a CPU
But cache intervenes at every load/store
“The science of multi-core cache sharing”

16. Locality

17. Rochester Elastic Cache Utility (RECU)
Trade off between Efficiency and Fairness.
Provides an algorithm to choose a cache partitioning for a set of programs in order to optimize the predicted cache performance.
With some fairness concessions based on a baseline of each program having an equal partition.
Strict baseline.

18. Miss Ratio Elasticity and Cache Space Elasticity
Elastic miss ratio baseline (EMB)
The program's predicted miss ratio may increase by up to some percentage compared to what it would be with the strict baseline.
Elastic cache space baseline (ECB)
The program may yield at most a given percentage of its cache space to peer programs.

19. Evaluation Results

20. Evaluation Results(Cont.)

21. Other Works

22. Summary Algorithm + Data Structure + Locality = Efficient Programs.
Locality science
Formal, universal relations between locality metrics.
Universal properties
Monotonicity, concavity, composition invariance and symmetry
Locality optimization
Minimize data movement between
Fast and slow memory
Small and large memory
Local and remote memory
Optimal program and cache organization