Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Scaling Science
Search
Matt Wood
November 21, 2012
Science
3
400
Scaling Science
Introducing five principles for reproducibility.
Matt Wood
November 21, 2012
Tweet
Share
More Decks by Matt Wood
See All by Matt Wood
Field Notes from Expeditions in the Cloud
mza
2
270
A Platform for Big Data
mza
6
650
The Data Lifecycle
mza
5
410
Provision Throughput Like a Boss
mza
0
360
Impact of Cloud Computing: Life Sciences
mza
2
760
Latency's Worst Nightmare: Performance Tuning Tips and Tricks
mza
4
1k
Under the Covers of DynamoDB
mza
4
870
From Analytics to Intelligence: Amazon Redshift
mza
9
900
High Performance Web Applications
mza
6
550
Other Decks in Science
See All in Science
20240420 Global Azure 2024 | Azure Migrate でデータセンターのサーバーを評価&移行してみる
olivia_0707
2
800
最適化超入門
tkm2261
11
2.9k
【健康&筋肉と生産性向上の関連性】 【Google Cloudを企業で運用する際の知識】 をお届け
yasumuusan
0
200
20231211ベクトル解析の計算
kamakiri1225
0
150
ICRA2024 速報
rpc
3
4.4k
効果検証入門に物申してみた_JapanR_2023
s1ok69oo
6
4.8k
ultraArmをモニター提供してもらった話
miura55
0
150
Machine Learning for Materials (Lecture 3)
aronwalsh
0
900
東大・松尾研主催 LLM Summer 2023 コンペ解法 (11位 – 20位枠での優秀賞)
hayataka88
0
270
Pandas 2 vs Polars vs Dask (PyDataGlobal 2023 December)
ianozsvald
0
660
Machine Learning for Materials (Lecture 2)
aronwalsh
0
640
はじめてのバックドア基準:あるいは、重回帰分析の偏回帰係数を因果効果の推定値として解釈してよいのか問題
takehikoihayashi
2
250
Featured
See All Featured
Learning to Love Humans: Emotional Interface Design
aarron
269
39k
Agile that works and the tools we love
rasmusluckow
325
20k
Docker and Python
trallard
37
2.9k
[RailsConf 2023] Rails as a piece of cake
palkan
35
4.4k
Put a Button on it: Removing Barriers to Going Fast.
kastner
58
3.3k
How GitHub Uses GitHub to Build GitHub
holman
471
290k
Faster Mobile Websites
deanohume
303
30k
Large-scale JavaScript Application Architecture
addyosmani
506
110k
個人開発の失敗を避けるイケてる考え方 / tips for indie hackers
panda_program
78
15k
What the flash - Photography Introduction
edds
65
11k
Easily Structure & Communicate Ideas using Wireframe
afnizarnur
189
16k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
18
1.2k
Transcript
Scaling Science
[email protected]
Dr. Matt Wood
Hello
Science
Beautiful, unique.
Impossible to re-create
Snowflake Science
Reproducibility
Reproducibility scales science
Reproduce. Reuse. Remix.
Value++
None
How do we get from here to there? 5PRINCIPLES REPRODUCIBILITY
OF
1. Data has Gravity 5 PRINCIPLES REPRODUCIBILITY OF
Increasingly large data collections
1000 Genomes Project: 200Tb
Challenging to obtain and manage
Expensive to experiment
Large barrier to reproducibility
Data size will increase
Data integration will increase
Data dependencies will increase
Move data to the users
Move data to the users X
Move tools to the data
Place data where it can consumed by tools
Place tools where they can access data
None
None
None
Canonical source
None
More data, more users, more uses, more locations
Cost
Force multiplier
Cost
Complexity
Cost and complexity kill reproducibility
Utility computing
Availability
Pay-as-you-go
Flexibility
Performance
CPU + IO
Intel Xeon E5 NVIDIA Tesla GPUs
240 TFLOPS
90 - 120k IOPS on SSDs
Performance through productivity
Cost
On-demand access
Reserved capacity
100% Reserved capacity
100% Reserved capacity On-demand
100% Reserved capacity On-demand
Spot instances
Utility computing enhanced reproducibility
None
2. Ease of use is a pre-requisite 5 PRINCIPLES REPRODUCIBILITY
OF
http://headrush.typepad.com/creating_passionate_users/2005/10/getting_users_p.html
Help overcome the suck threshold
Easy to embrace and extend
Choose the right abstraction for the user
$ ec2-run-instances
$ starcluster start
None
Package and automate
Package and automate Amazon machine images, VM import
Package and automate Amazon machine images, VM import Deployment scripts,
CloudFormation, Chef, Puppet
Expert-as-a-service
None
None
1000 Genomes Cloud BioLinux
None
Your HiSeq data Illumina BaseSpace
Architectural freedom
Freedom of abstraction
3. Reuse is as important as reproduction 5 PRINCIPLES REPRODUCIBILITY
OF
Seven Deadly sins of Bioinformatics: http://www.slideshare.net/dullhunk/the-seven-deadly-sins-of-bioinformatics
Seven Deadly sins of Bioinformatics: http://www.slideshare.net/dullhunk/the-seven-deadly-sins-of-bioinformatics
Infonauts are hackers
They have their own way of working
The ‘Big Red Button’
Fire and forget reproduction is a good first step, but
limits longer term value.
Monolithic, one-stop-shop
Work well for intended purpose
Challenging to install, dependency heavy
Di cult to grok
Inflexible
Infonauts are hackers: embrace it.
Small things. Loosely coupled.
Easier to grok
Easier to reuse
Easier to integrate
Lower barrier to entry
Scale out
Build for reuse. Be remix friendly. Maximize value.
4. Build for collaboration 5 PRINCIPLES REPRODUCIBILITY OF
Workflows are memes
Reproduction is just the first step
Bill of materials: code, data, configuration, infrastructure
Full definition for reproduction
Utility computing provides a playground for bioinformatics
Code + AMI + custom datasets + public datasets +
databases + compute + result data
Code + AMI + custom datasets + public datasets +
databases + compute + result data
Code + AMI + custom datasets + public datasets +
databases + compute + result data
Code + AMI + custom datasets + public datasets +
databases + compute + result data
Package, automate, contribute.
Utility platform provides scale for production runs
Drug discovery on 50k cores: Less than $1000
5. Provenance is a first class object 5 PRINCIPLES REPRODUCIBILITY
OF
Versioning becomes really important
Especially in an active community
Doubly so with loosely coupled tools
Provenance metadata is a first class entity
Distributed provenance
1. Data has gravity 2. Ease of use is a
pre-requisite 3. Reuse is as important as reproduction 4. Build for collaboration 5. Provenance is a first class object 5PRINCIPLES REPRODUCIBILITY OF
None
Thank you aws.amazon.com @mza
[email protected]