Distributed machine learning 101 using apache spark from the browser
What is Distributed Computing, Why we use Apache Spark
-
Upload
andy-petrella -
Category
Technology
-
view
351 -
download
0
Transcript of What is Distributed Computing, Why we use Apache Spark
BigData, newborn technologies evolving fast. Why Apache Spark
outruns Apache Hadoop
Andy Petrella, NextlabXavier Tordoir, SilicoCloud
Andy
@Noootsab, I am@NextLab_be owner@SparkNotebook creator@Wajug co-driver@Devoxx4Kids organizerMaths & CSData lover: geo, open, massiveFool
Who are we?
Xavier
@xtordoirSilicoCloud-> Physics
-> Data analysis -> genomics
-> scalable systems-> ...
So what...Part I
● What○ distributed resources○ data○ managers
● Why:○ fastest○ smartest○ biggest
● How:○ Map Reduce○ Limitations○ Extensions
PART II● Spark
○ Model○ Caching and lineage○ Master and Workers○ Core example
● Beyond Processing○ Streaming○ SQL○ GraphX○ MLlib○ Example
● Use cases○ Parallel batch processing of
timeseries○ ADAM
Part I: The Distributed Age
What is a distributed environmentComputations needs three kind of resources:● CPU ● MEM● Data storage
However, it’s hard to extent each of them at will on a single machine
What is a distributed environmentLacking of one of these will result in higher response time or reduced accuracy.Unfortunately, it doesn’t matter how parallelized is the algorithm or optimized are the computations
If the solution can’t be inside, it must be outside.
What is a distributed environment
Distributed File SystemYou have 100 nodes in your cluster, but only 1 dataset.Will you replicate it on all nodes?
Extended case: your dataset is 1 Zettabyte (10⁹Tb)?
Lonesome solution:● split the file on nodes● axing the algorithm to access local data subsets
HDFS towards TachyonHadoop Distributed File SystemImplements GoogleFSStore and read files splitted and replicated on nodes1Zb file = 8E12 x 128Mb files
IOPs are expensive and require more CPU clocks than DRAM accessHence... Tachyon: memory-centric distributed file system
Nodes will fail, jobs cannotWe need resilience
Management
Resources are generally fewer than required by algorithm.We need scheduling
The requirements are fluctuatingWe need elasticity
Mesos and MarathonMesos: High available cluster managerNodes: attach or remove them on the flyNodes are offering resources -- Applications accept themNode crash: the application restarts the assigned tasks
Marathon: Meta application on MesosApplication crash: automatically restarted on different node
Why: for everybody and now ?
Fastest:1. Time to result2. Near real time processing
Runtime is smaller, Dev lifecyle is shorter→ no synchronization-hell
It can even be really interactive → consoles or notebooks tools.
Why for everybody and now
Why for everybody and nowNo bottlenecks → new-coming data are readily available for processing
Opens the doors for online models!
Why for everybody and nowSmartest: train more and more models, ensembling lots of them is no more a problem
More complex modelling can be tackled if required
Why for everybody and nowAccessing an higher level of accuracy is tricky and might require lots and lots of models.
Running a model takes quite some time, specially if the data has to be read every single time.
Example: Netflix contest winner (AT&T labs) ensembled 500 models to gain 10% accuracy.Although in 2009 it wasn’t possible to use it in production, today this could change.
Why for everybody and nowBiggest: no need for sampling big datasets
……
That’s it!
How!?Google papers stimulated the open software community, hence competitive tools now exist.
In the area of computation in distributed environment, there are two disruptive papers:● Google’s Mapreduce● Berkeley’s Spark
How!?MapReduce (Google white paper 2004):
Programming model for distributed data intensive computations
Helps dealing with parallelization, fault-tolerance, data distribution, load balancing
Functions:Map ≅ transform data to key value pairs
Reduce ≅ aggregate key value pairs per key (e.g. sum, max, count)
Mappers and Reducers are sent to data location (nodes)
How!?
Map
Reduce: apply a binary associative operator on all elements
Image from RxJava: https://github.com/ReactiveX/RxJava/wiki/Transforming-Observables
How!?
Hadoop implementation has some limitations
Mappers and Reducers ship functions to data while java is not a functional language
⇒ Composability is difficult and more IO/network operations are required
Iterative algorithms (e.g. stochastic gradient) have to read data at each step (while data has not changed, only parameters)
How!?
How!?MapReduce on steroids
I) Functional paradigm:- process built lazily based on simple concepts- Map and Reduce are two of them
II) Cache data in memory. No more IO.
So what...Part I
● What○ distributed resources○ data○ managers
● Why:○ fastest○ smartest○ biggest
● How:○ Map Reduce○ Limitations○ Extensions
PART II● Spark
○ Model○ Caching and lineage○ Master and Workers○ Core example
● Beyond Processing○ Streaming○ SQL○ GraphX○ MLlib○ Example (notebook)
● Use cases○ Parallel batch processing of
timeseries○ ADAM
Part II: Spark to the Rescue
RDDsThink of an RDD[T] as an immutable, distributed collection of objects of type T
• Resilient => Can be reconstructed in case of failure• Distributed => Transformations are parallelizable
operations• Dataset => Data loaded and partitioned across cluster
nodes (executors)
RDD[T]Data distribution hierarchy:- RDD[T]- Elements
[ x1, x2 ]
[ x10 ]
[ x8,x5,x6 ]
[ x11 ]
[ x14,x13 ]
[ x9,x16 ]
[ x3 ]
[ x7,x12 ]
[ x15 ]
[ x17,x4 ]
Executor 1
- Executors- Partitions
Executor 2 Executor 3 Executor 4
Execution
Execution is split in fundamental units: Tasks
Tasks running in parallel are grouped in Stages
Execution
Core1Task0(read/process/write)
Task0(read/process/write)
Task0(read/process/write)
Core2Task1(read/process/write)
Task1(read/process/write)
Task1(read/process/write)
Core3Task2(read/process/write)
Task2(read/process/write)
Task2(read/process/write)
Stage2 Stage1 Stage0
Master and Workers
Spark StreamingWhen you have big fat streams behaving as one single collection
t
DStream[T]
RDD[T] RDD[T] RDD[T] RDD[T] RDD[T]
DStreams: Discretized Streams (= Sequence of RDDs)
Spark SQL
Mapping: RDD -> “table”, Element Field -> “column”
MLLib: Distributed ML
Classification● linear SVM, logistic regression, classification trees, naive Bayes Models
Regression● SVM, regression trees, linear regression (regularized)
Clustering & dimensionality reduction● singular value decomposition, PCA, k-means clustering
“The library to teach them all”
GraphX
Connecting the dots
Graph processing at scale. > Take edges > Link nodes > Combine/Send messages
Use cases examples
- Parallel batch processing of time series- Bayesian Network in financial market- IoT platform (Lambda architecture)- OpenStreetMap cities topologies classification- Markov Chain in Land Use/Land Cover prediction- Genomics: ADAM
Genomics
Biological systems are very complexOne human sequence is 60Gb
ADAMCredits: AmpLab (UC Berkeley)
Stratification using 1000Genomes
http://www.1000genomes.org/
ref: http://upload.wikimedia.org/wikipedia/en/e/eb/Genetic_Variation.jpg
Machine Learning model
Clustering: KMeans
ref: http://en.wikipedia.org/wiki/K-means_clustering
Machine Learning modelMLLib, KMeans
MLLib: ● Machine Learning Algorithms● Data structures (e.g. Vector)
Mashupprediction
Sample [NA20332] is in cluster #0 for population Some( ASW)
Sample [NA20334] is in cluster # 2 for population Some( ASW)
Sample [HG00120] is in cluster # 2 for population Some( GBR)
Sample [NA18560] is in cluster # 1 for population Some( CHB)
Mashup
#0 #1 #2
GBR 0 0 89ASW 54 0 7CHB 0 97 0
Cluster40 m3.xlarge160 cores + 600G
Eggo project (public genomics data in ADAM format on s3)
We…1000genomes in ADAM format on S3. Open Source GA4GH Interop services implementationMachine learning on 1000genomes
Genomic data and distributed computing