- Moore’s Law: Old faithful + rising tide
- In the future Moore’s law will be less predictable
- Moore’s law may end around 2020 or 2022
- Silicon lattice will be a limiting factor in how small transistors can go
- Past 40 years the industry has marched towards homogeneous (x86, ARM, CMOS, etc.)
- Happening now: swing back towards heterogeneous tech base
- Current uses for increased transistor budget: more thread or cores.
- Future uses for transistors: specialized scenarios such as GPUs, crypto, image DSP, etc.
- Emergent compute fabric: constellation of specialized compute units
- Compound challenges: Distributed resources, concurrency, heterogeneity
- What is needed: revised OS abstraction/stack layers, revised role for compilers, reduced role of programmers
- System software is living in the past: OSes built for CPU+disk+network world
- OS thinks accelerators are I/O devices – Not good
- Coming soon: Creativity + algorithms
- Big data: can’t read it all efficiently. Big distributed systems: can’t know exactly what’s going on at every single moment
- Main tools: randomization: If you don’t know something for sure, make a guess
- Load balancing can work with hashing, but it does not scale effectively.
- Power of two: Pick two random bins, and place the data in the least loaded bin.
- Consistency: Past – Error prone transactions to ACID and strong consistency.
- Consistency future: weak consistency and massive scale.
- Transactional journey: DBMS (SQL) to NoSQL (Cassandra) to NewSQL (ZooKeeper, MongoDB)
- Today’s data technology inefficiency: a complex software stack
- VMware developed Corfu and it’s open source in GitHub
- Internet of things: Link the physical and virtual worlds
- Past 40 years: batch computing moving to interactive
- Coming soon: interactive moving to proactive computing
- Systems that anticipate our needs and act on our behalf
