1. Ali Ghodsi UC Berkeley & KTH & SICS alig@cs.berkeley.edu
IoT Meets the Cloud
Ali Ghodsi
UC Berkeley & KTH & SICS

2. Cloud Computing?
Larry Ellison, CEO of Oracle Corporation “The computer industry is the only industry that is more fashion-driven than women's fashion. Maybe I'm an idiot, but I have no idea what anyone is talking about. What is it? It's complete gibberish. It's insane. When is this idiocy going to stop?”
Richard M. Stallman, President of FSF “It’s stupidity. It’s worse than stupidity: it’s a marketing hype campaign. Somebody is saying this is inevitable — and whenever you hear somebody saying that, it’s very likely to be a set of businesses campaigning to make it true.”
My claim:
Cloud computing is inevitable for the Internet-of-Things

3. Most of the Computation on the Cloud Already!
Mobile Applications
Proximity sensor, ambient light, digital compass, gyroscope, accelerometer, dual microphone, multi-touch touchscreen
bluetooth
Most of the Computation on the Cloud Already!

4. Do we need the cloud for IoT?
Device deluge
3 billion smart phones
Another 40 billion IoT devices
Devices will be challenged
Limited storage
Limited processing
Limited communication
Limited energy
12kbit/s,
P2P
Clouds needed for IoT, just as for phones and desktops

5. What is the cloud? Datacenter Computing Thousands of servers
Co-located storage
Routers and switches
Backup power supplies
Cooling
Talk of the size... Half a mill.... Amazon

6. Why do we need datacenters?
Multi-core Computing
Processing speed stagnation
Increased parallelism
Supercomputer not sufficient
Parallel computing quintessential to cloud computing
Request-level parallelism
Parallel algorithms (MapReduce, Indexing …)
Typical machine, 12 disk drives, 16 cores, several NICs, 160 GB memory

7. Why do we need datacenters? (2)
Economy of scale
Reduce server cost
Reduce cooling cost
Reduce power cost
Clouds are efficient
PUE = total_facility_power/ equipment_power ~ 1.2
Energy economy-of-scale
Commodity servers
Workload consolidation
Energy: common cooling, near cheap places and good climate for cooling, common infrastructure, buildings, backup power
Commodity: reliability

8. Workload Consolidation
Data replicated over commodity machines
Pioneered by Inktomi
Interactive and latency sensitive jobs
User facing applications e.g. search queries, tweets, …
Millisecond SLOs
Batch-jobs
Building search indexes …
Analytics of trends, business data …
AV/spam filtering …

9. Workload Consolidation (2)
Interactive and batch on same machines
Virtualization of computation e.g. migration, hardware agnosticism
Isolation of workloads e.g. meet SLO guarantees
Automatic fault-handling e.g. through replication

10. Transformation of Computing
Datacenter as a computer
Programs timeshare thousands of servers
Twitter

11. Berkeley Vision
Create an “Operating System Kernel” for the Datacenter Computer
First step with Mesos (mesosproject.org)
Twitter

12. Today’s Cloud Frameworks
Dryad
Pregel
Frameworks simplify distributed programming
Programming models
Hide failures, synchronization, delay variance
Each framework runs on a dedicated cluster/partition

13. One Framework Per Cluster Challenges
Inefficient resource usage
E.g., Hadoop cannot use available resources from IoT FW cluster
No opportunity for stat. multiplexing
Hard to share data
Copy or access remotely, expensive
Hard to cooperate
E.g., Not easy for IoT FW to use data generated by Hadoop
Hadoop
IoT FW
Hadoop
IoT FW
Need to run multiple frameworks on the same cluster

14. Solution: Mesos Common resource sharing layer Mesos Uniprograming
abstracts (“virtualizes”) resources to frameworks
enable diverse frameworks to share cluster
Hadoop
IoT FW
Hadoop
IoT FW
Mesos
Our solution is Mesos, a resource sharing layer that abstracts resources to framework, and this way allows over which diverse frameworks cab run, by abstracting resources to framework
By doing so we go from uniprograming or one framework per cluster to multiprogramming where we share entire cluster among diverse frameworks.
Uniprograming
Multiprograming

15. IoT Framework Diversity
Today’s frameworks tailored for specific application domains
MapReduce for indexing and filtering
Pregel for graph algorithms
IoT problem domain highly diverse
Existing frameworks poor fit for IoT

16. New IoT Frameworks for Clouds
IoT framework requirements
Efficient device tag matching and filtering
Online stream processing of IoT data
Offline storage and batch processing of IoT data
Goal: Build first cloud framework for IoT

17. IoT Framework Applications
Real time stream processing of data
Security, safety, health applications
Locating people, devices, objects
Real time. Supply-chain, real time trigger the ordering of new supplies when inventory reaches low
Real time. Enough-smoke sensors over an area goes off, trigger emergency system

18. IoT Framework Applications (2)
Batch processing of big data
Learning trends, patterns, anomalies
Collaborative filtering/recommendation
Computing global device statistics
Batch. Retailers in a region can collaborate to get aggregate trend of undersupply of a product category in a city.
Batch. Aggregate users behavior to recommend to others

19. Summary
Dichotomy:
Challenged IoT vs Powerful Clouds
”nerves”—sensors, actuators—collect and send data to the ”brain”—the datacenter
Datacenter is the new super computer
Will need to multiplex between many IoT FW
Need IoT-tailored frameworks to aid IoT services