How to have a research career in industry Rebecca Isaacs, Research - - PowerPoint PPT Presentation

How to have a research career in industry

Rebecca Isaacs, Research Scientist at Google SOSP Diversity Workshop October 28, 2017

These are my personal opinions and not necessarily those of my employer

1

Outline

• 1. The state of systems research in industry
• 2. How do we do research in my group at

Google?

• 3. Ways to succeed
• 4. Program committees and other forms of

public service

2

Who is this talk for?

• Students considering career choices
• Academics considering a switch to

industry

○ Sabbaticals can be a great way to test the water

• The sceptics who don’t believe research

gets done in industry

3

In the golden days of industry research labs...

“Zen and the Art of Research Management”, Naughton & Taylor

• 6. Do not pay too much attention to relevance,

deliverables and other concepts beloved of Senior Management.

• 7. Remember that creative people are like hearts -

they go where they are appreciated. They can be inspired or led, but not managed.

4

Systems research today at Google

“Google’s Hybrid Approach to Research”, Spector et al. The goal of research at Google is to bring significant, practical benefits to our users, and to do so rapidly, within a few years at most. Because of the time-frame and effort involved, Google’s approach to research is iterative and usually involves writing production, or near-production, code from day one.

5

A few influential systems from the research labs

• The Alto, Ethernet, Grapevine (Xerox PARC)
• The Firefly, early cache coherence protocols,

AltaVista, Paxos, Autonet, Practical RISC, DCPI (DEC SRC/WRL)

• MSR: Differential privacy, HoloLens,

Singularity, Dryad/DryadLINQ

• … and many more...

6

A few influential systems from Google

• MapReduce
• The Google file system
• Bigtable
• Spanner
• TensorFlow
• and many more

7

Different… but similar

• Both approaches have produced great

research

• The employer recognizes and carries the

risk

○ True innovation can - and should sometimes - fail

• The importance of the PhD pipeline and

academic research is recognized

8

Why choose an industrial research career?

• Impact

○ Your ideas, your designs, your code can change the world… literally

• Real problems and real data

○ Interesting and intellectually challenging

• Career path flexibility

○ SWE, SRE, PM, management, ...

9

Another view

making the system work… in production… at scale researcher defining the abstractions, designing the protocol, analyzing the algorithm

10

Another view

researcher defining the abstractions, designing the protocol, analyzing the algorithm making the system work… in production… at scale

11

Comparing research styles (slightly tongue-in-cheek)

12

Industry Academia Collaboration

Large teams Small number of PIs, students, cross-disciplinary opportunities

Politics

Multiple stakeholders, may have competing goals “Us vs the world”

Motivation

Solve a specific problem, eg improve user experience, or reduce memory footprint Apply a tasteful solution to an intellectually pleasing problem / follow the funding

Environment

Constrained implementation and deployment options Flexible (but must be cheap) / dictated by funding

Evaluation

By experience, likely at scale Micro-benchmarks, synthetic or limited real-world data

Desired output

Impact, tangible advances in the state-of-the-art Publications, prestige

A case study of research at Google

The Network Infrastructure Group “creates the networks that power Google” https://research.google.com/teams/netsys/

13

Leading by example

at Google

14

NetInfra focus and publishing

Congestion control Data center networks SDN Network management Wide area networks

BBR: Congestion-based congestion control, CACM Feb 2017 Carousel: Scalable Traffic Shaping at End Hosts, Sigcomm 2017 Taking the Edge off with Espresso: Scale, Reliability and Programmability for Global Internet Peering, Sigcomm 2017 B4: Experience with a Globally Deployed Software Defined WAN, Sigcomm 2013 BwE: Flexible, hierarchical bandwidth allocation for WAN distributed computing, Sigcomm 2016 Thinking about Availability in Large Service Infrastructures, HotOS 2017 Evolve or Die: High-Availability Design Principles Drawn from Google's Network Infrastructure, Sigcomm 2016 Jupiter rising: a decade of clos topologies and centralized control in Google's datacenter network, Sigcomm 2016

15

My current project

• Goals

○ Improve our understanding of how changes in the network affect applications ○ Know whether applications are getting the best possible network performance

• Approach

○ A network telemetry system at the intersection of the RPC layer and the transport protocol ○ Instrumentation across all of Google, from data center to WAN

16

Network operators vs users

User: My pager went off 4 times overnight due to latency SLO violations. The network is broken! Operator: No hardware issues. Utilization and load balancing looks good. There were some discard spikes on a few switches. Reconciling these views is often a hard problem, especially at scale, but critical for Google!

17

Challenges

• Data collection

○ What should we measure? At what resolution? ○ What information should we store, and in what form?

• Tools and techniques to exploit the data

○ Automated analysis: e.g. anomaly detection, A/B comparison, critical path analysis ○ Performance debugging via fine-grained inspection

• Handling the (big) data

○ We collect telemetry on millions of RPCs per second ○ Our daily processing job requires ~1PB temp storage and persists ~10TB / day

18

Succeeding as a researcher in industry

Choose your problems wisely

• There is no glory in solving a problem that nobody cares about

Build a network

• Know who to talk to

Be adaptable and play to your strengths

• Critical thinking and analysis
• Awareness of the big picture
• Knowledge about state-of-the-art
• Presentation and writing skills

19

Succeeding as a systems researcher in industry

Be prepared to “shovel”

• Implementation, code review, testing, on-call duties, documentation, rollout

schedules, pre-existing code, poor designs, ugly abstractions, ...

Be aware of the priorities of your team and your organization

• Pragmatism rather than idealism

Interact outside your team

• Understand others’ problems and constraints

20

Staying in touch: external community service

Can be personally rewarding and professionally advantageous. However, employers will typically underestimate the amount of time and energy external service

• demands. You may be “encouraged” to do it, but

not necessarily given enough time at work.

21

Program committees: why?

Stay visible in the research community Have visibility of the research community Networking opportunities: hiring, interns, being invited to give talks at diversity workshops

22

Program committees: why?

Keep your skills sharp

• Critical thinking
• Opportunities to learn

○ Experimental method, motivating a problem etc.

• Abstract problem solving

○ “Just write the code” can be a seductive way to avoid thinking too hard

• Articulating technical arguments

○ Writing (for the authors) ○ Speaking (in the PC meeting)

23

Program committees: how?

It’s by invitation, my advisor doesn’t help, nobody knows me…

Advice

• Let people know that you want to do it

○ Take on other volunteer tasks for conferences

• When asked, do it well

○ Even shadow PCs matter ○ Even small workshops matter ○ Even journals that you believe nobody reads matter

24

What about the diversity card?

PC Chairs will usually try to put together a diverse PC. This does not mean that you are invited solely because of your diversity. If you want to be invited again, make sure to do a good job.

25

Conclusion

Systems research in industry can be tough, the real world is… well,… real but potentially tremendously impactful and rewarding.

26