The Datacenter Needs an Operating...

transcript

The Datacenter Needs an Operating System

UC BERKELEY

Anthony D. Joseph

LASER Summer School September 2013

My Talks at LASER 2013

1.  AMP Lab introduction

2.  The Datacenter Needs an Operating System

3.  Mesos, part one

4.  Dominant Resource Fairness

5.  Mesos, part two

6.  Spark 2

Collaborators

•  Matei Zaharia

•  Benjamin Hindman

•  Andy Konwinski

•  Ali Ghodsi

•  Randy Katz

•  Scott Shenker

•  Ion Stoica 3

Machines: Background

Clusters of commodity servers have become a major computing platform in industry and academia (100’s – 10,000’s of machines)

Driven by data volumes outpacing the processing capabilities of single machines – big data and science

Democratized by cloud computing

Machines: Background Some have declared that “the datacenter is the new computer”

Our claim: this new computer increasingly needs an operating system

Not necessarily a new host OS, but a common software layer that manages resources and provides shared services for the whole datacenter, like an OS does for one host

Why Datacenters need an OS Growing diversity of applications » Computing frameworks: MapReduce, Dryad, Pregel,

Percolator, Dremel, MR Online, Spark » Storage systems: GFS, BigTable, Dynamo, SCADS » Web apps and supporting services

Pregel

Cassandra Hypertable

Why Datacenters need an OS Growing diversity of applications » Computing frameworks: MapReduce, Dryad, Pregel,

Percolator, Dremel, MR Online, Spark » Storage systems: GFS, BigTable, Dynamo, SCADS » Web apps and supporting services

Growing diversity of users » 200+ Hive users at Facebook, ��

running near-interactive ��ad hoc queries

Same reasons computers��needed one!

What Operating Systems Provide

Resource Sharing

Data Sharing Programming Abstractions

Debugging & Monitoring

time-sharing, virtual memory, …

ptrace, DTrace, top, …

files, pipes, IPC, … libraries, languages

What Operating Systems Provide

Resource Sharing

Most importantly: enables a highly interoperable software ecosystem

that we now take for granted

Example

A scientist analyzing data on one machine can pipe it through a variety of tools, write new tools that interface with these through standard APIs, and trace across the stack

In the future, the scientist should be able to launch a cluster on EC2 and do the same things: » Mix and combine a variety of apps & programming models » Write new parallel programs that talk to these » Get a unified interface for managing the cluster » Debug and trace across all these components

Today’s Datacenter OS

Hadoop MapReduce as common execution and resource sharing platform » Means jobs have to compile to MapReduce » Inter-user resource sharing, but at the level of MR jobs

Hadoop InputFormat API for data sharing – what happens with the next hot platform after Hadoop?

Abstractions for productivity programmers, but not for system builders

Difficult to debug, especially across layers

Other examples: » Amazon/Azure services » Google internal stack and Google Compute Engine » Hadoop YARN

Abstractions for productivity programmers, but not for system builders

Difficult to debug, especially across layers

Other examples: » Amazon/Azure services » Google internal stack and Google Compute Engine » Hadoop YARN

The problems motivating a datacenter OS are well recognized, but solutions are narrowly targeted

Can researchers take a longer-term view?

Tomorrow’s Datacenter OS

Resource Sharing

Resource Sharing To solve these interaction problems we would like to have a computer made simultaneously available to many users in a manner somewhat like a telephone exchange. Each user would be able to use a console at his own pace and without concern for the activity of others using the system.”

– Fernando J. Corbató, 1962

Today’s Resource Sharing Today, cluster apps are built to run independently��and assume they own a fixed set of nodes

Result: inefficient static partitioning

What’s the right interface for dynamic sharing?

0% 17% 33%

0% 17% 33% 0%

100% App 1

App 3 16

Resource sharing: » Lower-level interfaces for fine-grained sharing – Mesos

and Hadoop YARN are first steps in this direction » Optimization for a variety of metrics (e.g., energy) » Integration with network scheduling mechanisms (e.g.,

Seawall [NSDI ‘11], NOX, Orchestra) » Others: Azure Fabric Controller

Resource Sharing

Persistent data sharing – many design issues addressed » Placement/Locality » Reliability » Availability » Consistency » Bandwidth/Latency » Software versioning

Persistent data sharing: » Standard interfaces for cluster file systems, key-value

stores, etc. » Lineage instead of replication for reliability (Spark RDDs) » Application frameworks self-manage versioning

Many possibilities: » Amazon Elastic Block Store and S3 » HDFS » Azure storage services

Transient data sharing – many design issues addressed » Failures on either side » Consistency » Timeliness

Transient data sharing: » In-memory data sharing (e.g. Spark, DFS cache), and a

unified system to manage this memory – DFS cache for MapReduce cluster could serve 90% of jobs at Facebook (HotOS ’11) » Streaming data abstractions (analogous to pipes)

Many possibilities: » Amazon/Azure message queues » Percolator

Resource Sharing

Programming abstractions: » Many new distributed application programming models,

abstractions, and languages » Tools for programming for distributed coordination and

fault-tolerance (e.g., Apache Zookeeper) » New tools that can be used to build the next

MapReduce / BigTable in a week (e.g., BOOM) » Efficient implementations of communication primitives

(e.g. shuffle, broadcast)

Resource Sharing

Debugging and Monitoring facilities: » Tracing and debugging tools that work across the cluster

software stack (e.g. X-Trace, Dapper, Magpie, Hystrix) » Replay debugging that takes advantage of limited

languages / computational models » Unified monitoring infrastructure and APIs (e.g., Hystrix)

Putting it Together

A successful datacenter OS might let users: » Build a Hadoop-like software stack in a week using the

OS’s APIs, while gaining other benefits (e.g. cross-stack replay debugging) » Share data efficiently between independently written

apps and programming frameworks » Understand cluster behavior without having to log into

individual nodes » Dynamically share the cluster with other users

How Researchers can Help Focus on paradigms, not performance » Industry is tackling performance but lacks luxury to take

long-term view towards abstractions

Explore clean-slate approaches » Likelier to have greater impact here than in a “real” OS

because datacenter software changes quickly!

Bring cluster computing to non-experts » Most impactful (datacenter as the new workstation) » Much harder and more rewarding than big users

Berkeley Data Analytics Stack

Apache Spark

Shark BlinkDB

HDFS / Hadoop Storage / Tachyon

Apache Mesos / YARN Resource Manager

Spark Streaming

GraphX MLBase

Apache Mesos – Cluster Operating System

Efficiently shares resources among diverse parallel

applications Mesos slave

Mesos master

Dryad scheduler

Mesos slave

Hadoop executor

Mesos slave

Dryad executor

MPI scheduler

MPI executor

Hadoop scheduler

Dryad executor

MPI executor

0% 10% 20% 30% 40% 50% 60% 70% 80% 90%

1 31 61 91 121 151 181 211 241 271 301 331

Time (s)

Hadoop

Machines Make datacenter a real computer!

Node OS (e.g. Linux)

Node OS (e.g. Windows)

Datacenter “OS” (e.g., Apache Mesos)

•  Share datacenter between multiple cluster computing apps •  Provide new abstractions and services

AMP stack

Existing stack

Hive Support existing cluster computing apps

AMP stack

Existing stack

k SCADS

Hive PIQL

Support interactive and iterative data analysis (e.g., ML algorithms)

Consistency adjustable data store

Predictive & insightful query language

AMP stack

Existing stack

k SCADS

Applications, tools

Hive PIQL •  Advanced ML algorithms •  Interactive data mining •  Collaborative visualization

AMP stack

Existing stack

Milestones

2010: Mesos in Apache incubator

2010: Spark open sourced

2012: Shark (SQL) open sourced

Feb 2013: Spark Streaming alpha open sourced

Mar 2013: Tachyon alpha open sourced

Jun 2013: Spark entered Apache Incubator

Aug 2013: Machine Learning library for Spark 35

BDAS Users��(partial list)

BDAS Buzz

Big Data Landscape – Our Corner

MLbase Meet Up at Twitter ��(13 Aug 2013)

BDAS Contributors

70+ public contributors on GitHub » US, China, India, UK, Canada, Vietnam » Startups and large multinationals: Intel, Yahoo, Ooyala,

Quantifind, ClearStory, Palantir, Foursquare, Groupon …

Researchers Using BDAS

UC Berkeley

IBM Almaden

Cornell

Tsinghua

Purdue

What is fueling the traction?

Superior technologies J » Fast and expressive » It works!

Integration with existing Hadoop ecosystem » HDFS » HBase » Hive

BDAS Future Directions

Future data analytics need to support » Fast SQL » Approximate queries » Machine learning » GraphX » Streaming » Crowdsourcing!!!

Mix and match all of the above

http://ampcamp.berkeley.edu/3/ 43

Conclusion

Datacenters need an OS-like software stack for same reasons as single computers: manageability, efficiency, programmability, and thriving software ecosystem

Multiple DCOS already emerging in ad-hoc ways

Researchers can help by taking a long-term systems view towards these problems

My Talks at LASER 2013

1.  AMP Lab introduction

2.  The Datacenter Needs an Operating System

3.  Mesos, part one

4.  Dominant Resource Fairness

5.  Mesos, part two

6.  Spark 45

The Datacenter Needs an Operating...

Documents