Speed up R with parallel programming in the Cloud
Download as PPTX, PDF7 likes13,780 views
David M. Smith discusses the use of R for data science, particularly in production environments, highlighting the need for parallel programming to enhance performance. He introduces tools like the 'foreach' package for parallel processing and 'doazureparallel' for leveraging Azure cloud resources to run R code efficiently. Additionally, the document covers integrating R with Spark for larger datasets using the 'sparklyr' package and provides practical guidance on setting up clusters and optimizing R workflows.
![Packages and Containers
• Docker images used to spawn nodes
– Default: rocker/tidyverse:latest
– Lots of R packages pre-installed
• But this cross-validation also needs:
– xgboost, e1071
• Easy fix: add to cluster.json
19 @revodavid
{
"name": "davidsmi8caret",
"vmSize": "Standard_D2_v2",
"maxTasksPerNode": 8,
"poolSize": {
"dedicatedNodes": {
"min": 4,
"max": 4
},
"lowPriorityNodes": {
"min": 4,
"max": 4
},
"autoscaleFormula": "QUEUE"
},
"containerImage":
"rocker/tidyverse:latest",
"rPackages": {
"cran": ["xgboost","e1071"],
"github": [],
"bioconductor": []
},
"commandLine": []
}](https://image.slidesharecdn.com/speeduprwithparallelprogramming-180127192527/85/Speed-up-R-with-parallel-programming-in-the-Cloud-19-320.jpg)
Speed up R with parallel programming in the Cloud
- 1. Speeding up R with Parallel Programming in the Cloud David M Smith Developer Advocate, Microsoft @revodavid
- 2. Hi. I’m David. • Lapsed statistician • Cloud Developer Advocate at Microsoft – cda.ms/7f • Editor, Revolutions blog – cda.ms/7g • Twitter: @revodavid 2 @revodavid
- 3. What is R? • Widely used data science software • Used by millions of data scientists, statisticians and analysts • Most powerful statistical programming language • Flexible, extensible and comprehensive for productivity • Creates beautiful and unique data visualizations • As seen in New York Times, The Economist and FlowingData • Thriving open-source community • Leading edge of Statistics research 3 @revodavid
- 4. What aren’t you telling me about R? • R is single-threaded • R is an in-memory application • R is kinda quirky And yet, major companies use R for production data science on large databases. • Examples: blog.revolutionanalytics.com/applications/ 4 @revodavid
- 5. Secrets to using R in production • Don’t use R alone – Know what it’s good for! (And what it’s not good for.) – Use R as part of a production stack • Use modern R workflows – Hire great data scientists • Use R in conjunction with parallel/distributed data & compute architectures 5 @revodavid
- 6. Speeding up your R code • Vectorize 6 @revodavid • Vectorize • Moar megahertz! • Moar RAM! • Moar cores! • Moar computers! • Moar cloud!
- 7. Embarassingly Parallel Problems Easy to speed things up when: • Calculating similar things many times – Iterations in a loop, chunks of data, … • Calculations are independent of each other • Each calculation takes a decent amount of time Just run multiple calculations at the same time 7 @revodavid
- 8. 8 Is this Embarrassing? Embarrassingly Parallel Group-by Analyses Reporting Simulations Resampling / Bootstrapping Optimization / Search (somewhat) Prediction (scoring) Cross-Validation Backtesting Not Embarrassingly Parallel SQL operations (many) Matrix inverse Linear regression (training) Logistic Regression (training) Trees (training) Neural Networks (training) Time Series (most things) Train tracks
- 9. The Birthday Paradox What is the likelihood that there are two people in this room who share the same birthday? 9 @revodavid
- 10. Birthday Problem Simulator 10 pbirthdaysim <- function(n) { ntests <- 100000 pop <- 1:365 anydup <- function(i) any(duplicated( sample(pop, n, replace=TRUE))) sum(sapply(seq(ntests), anydup)) / ntests } bdayp <- sapply(1:100, pbirthdaysim) About 5 minutes (on this laptop)
- 11. library(foreach) Looping with the foreach package on CRAN – x is a list of results – each entry calculated from RHS of %dopar% Learn more about foreach: cda.ms/6Q 11 @revodavid x <- foreach (n=1:100) %dopar% pbirthdaysim(n)
- 12. Parallel processing with foreach • Change how processing is done by registering a backend – registerDoSEQ() sequential processing (default) – registerdoParallel() local cluster via library(parallel) – registerAzureParallel() remote cluster in Azure • Whatever you use, call to foreach does not change – Also: no need to worry about data, packages etc. (mostly) 12 @revodavid library(doParallel) cl <- makeCluster(2) # local cluster, 2 workers registerDoParallel(cl) bdayp <- foreach(n=1:100) %dopar% pbirthdaysim(n)
- 13. foreach + doAzureParallel • doAzureParallel: A simple R package that uses the Azure Batch cluster service as a parallel-backend for foreach github.com/Azure/doAzureParallel
- 14. Demo: birthday simulation 8-node cluster (compute-optimized D2v2 2-core instances) • specify VM class in cluster.json • specify credentials for Azure Batch and Azure Storage in credentials.json 14 @revodavid library(doAzureParallel) setCredentials("credentials.json") cluster <- makeCluster("cluster.json") registerDoAzureParallel(cluster) bdayp <- foreach(n=1:100) %dopar% pbirthdaysim(n) bdayp <- unlist(bdayp) cluster.json (excerpt): "name": "davidsmi8caret", "vmSize": "Standard_D2_v2", "maxTasksPerNode": 8, "poolSize": { "dedicatedNodes": { "min": 8, "max": 8 } 45 seconds (more than 6 times faster!)
- 15. Scale • From 1 to 10,000 VMs for a cluster • From 1 to millions of tasks • Your selection of hardware: – General compute VMs (A-Series / D- Series) – Memory / storage optimized (G-Series) – Compute Optimized (F-Series) – GPU enabled (N-Series) • Results from computing the mandelbrot set when scaling up: Local machine 5 parallel workers 10 parallel workers 20 parallel workers
- 16. Cross-validation with caret • Most predictive modeling algorithms have “tuning parameters” • Example: Boosted Trees – Boosting iterations – Max Tree Depth – Shrinkage • Parameters affect model performance • Try ‘em out: cross-validate 16 @revodavid grid <- data.frame( nrounds = …, max_depth = …, gamma = …, colsample_bytree = …, min_child_weight = …, subsample = …) )
- 17. Cross-validation in parallel • Caret’s train function will automatically use the registered foreach backend • Just register your cluster first: registerDoAzureParallel(cluster) • Handles sending objects, packages to nodes 17 @revodavid mod <- train( Class ~ ., data = dat, method = "xgbTree", trControl = ctrl, tuneGrid = grid, nthread = 1 )
- 18. caret speed-ups • Max Kuhn benchmarked various hardware and OS for local parallel – cda.ms/6V • Let’s see how it works with doAzureParallel 18 @revodavid Source: cda.ms/6V
- 19. Packages and Containers • Docker images used to spawn nodes – Default: rocker/tidyverse:latest – Lots of R packages pre-installed • But this cross-validation also needs: – xgboost, e1071 • Easy fix: add to cluster.json 19 @revodavid { "name": "davidsmi8caret", "vmSize": "Standard_D2_v2", "maxTasksPerNode": 8, "poolSize": { "dedicatedNodes": { "min": 4, "max": 4 }, "lowPriorityNodes": { "min": 4, "max": 4 }, "autoscaleFormula": "QUEUE" }, "containerImage": "rocker/tidyverse:latest", "rPackages": { "cran": ["xgboost","e1071"], "github": [], "bioconductor": [] }, "commandLine": [] }
- 20. 20 @revodavid ================================================ Id: job20180126022301 chunkSize: 1 enableCloudCombine: TRUE packages: caret; errorHandling: stop wait: TRUE autoDeleteJob: TRUE ================================================ Submitting tasks (1250/1250) Submitting merge task. . . Job Preparation Status: Package(s) being install Waiting for tasks to complete. . . | Progress: 13.84% (173/1250) | Running: 59 | Qu MY LAPTOP: 78 minutes THIS CLUSTER: 16 minutes (almost 5x faster)
- 21. How much does it cost? • Pay by the minute only for VMs used in cluster – No additional cost for the Azure Batch cluster service • Using D2v2 Virtual Machines – Ubuntu 16, 8Mb RAM, 2-core “compute optimized” • 17 minutes × 8 VMs @ $0.10 / hour – about 23 cents (not counting startup) … but why pay full price? 21 @revodavid
- 22. Low Priority Nodes • Low-Priority = (very) Low Costs VMs from surplus capacity – up to 80% discount • Clusters can mix dedicated VMs and low-priority VMs My Local R Session Azure Batch Low Priority VMs at up to 80% discount Dedicated VMs "poolSize": { "dedicatedNodes": { "min": 3, "max": 3 }, "lowPriorityNodes": { "min": 9, "max": 9 },
- 23. TL;DR: Embarrassingly Parallel • Install the foreach and doAzureParallel packages • Get Azure Batch and Azure Storage accounts – Need an account? http://azure.com/free • Set up Azure keys in credentials.json • Define your cluster size/type in cluster.json • Use registerAzureParallel to set up your job • Use foreach / %dopar% to loop in parallel • Worked example with code: cda.ms/7d 24 @revodavid
- 24. DISTRIBUTED DATA PROCESSING WITH SPARKLYR For when it’s not embarrassingly parallel: 25 @revodavid Azure
- 25. What is Spark? • Distributed data processing engine • Store and analyze massive volumes in a robust, scalable cluster • Successor to Hadoop • in-memory engine 100x faster than map-reduce • Highly extensible, with machine-learning capabilities • Supports Scala, Java, Python, R … • Managed cloud services available – Azure Databricks & HDInsight, AWS EMR, GCP Dataproc • Largest open-source data project • Apache project with 1000+ contributors 26 @revodavid
- 26. R and Spark: Sparklyr • sparklyr: R interface to Spark – open-source R package from RStudio • Move data between R and Spark • “References” to Spark Data Frames – Familiar R operations, including dplyr syntax – Computations offloaded to Spark cluster, and deferred until needed • CPU/RAM/Disk consumed in cluster, not by R • Interfaces to Spark ML algorithms 27 @revodavid spark.rstudio.com
- 27. Provisioning clusters for Sparklyr with aztk • aztk: Command-line interface to provision Spark-ready (and Sparklyr-ready) clusters in Batch – www.github.com/azure/aztk • Provision a Spark cluster in about 5 minutes – Choice of VM instance types – Use provided Docker instances (or your own) – Pay only for VM usage, by the minute – Optionally, use low-priority nodes to save costs • Tools to manage persistent storage • Easily connect to RStudio Server and Spark UIs from desktop 28 @revodavid
- 28. Launch and connect Provision a Spark cluster: aztk spark cluster create --id davidsmispark4 --size 4 Connect to the Spark cluster and map ports: aztk spark cluster ssh --id davidsmispark4 Launch RStudio http://localhost:8787 29 @revodavid
- 29. dplyr with Sparklyr 30 @revodavid Connect to the Spark cluster: library(sparklyr) cluster_url <- paste0("spark://", system("hostname -i", intern = TRUE), ":7077") sc <- spark_connect(master = cluster_url) Load in some data: library(dplyr) flights_tbl <- copy_to(sc, nycflights13::flights, "flights") Munge with dplyr: delay <- flights_tbl %>% group_by(tailnum) %>% summarise(count = n(), dist = mean(distance), delay = mean(arr_delay)) %>% filter(count > 20, dist < 2000, !is.na(delay)) %>% collect
- 30. Things to note • All of the computation take place in the Spark cluster – Computations are delayed until you need results – Behind the scenes, Spark SQL statements are being written for you • None of the data comes back to R – Until you call collect, when it becomes a tbl – It’s only at this point you have to worry about data size • This is all ordinary dplyr syntax 31 @revodavid
- 32. Machine Learning with SparklyR 33 @revodavid > m <- ml_linear_regression(delay ~ dist, data=delay_near) * No rows dropped by 'na.omit' call > summary(m) Call: ml_linear_regression(delay ~ dist, data = delay_near) Deviance Residuals:: Min 1Q Median 3Q Max -19.9499 -5.8752 -0.7035 5.1867 40.8973 Coefficients: Estimate Std. Error t value Pr(>|t|) (Intercept) 0.6904319 1.0199146 0.677 0.4986 dist 0.0195910 0.0019252 10.176 <2e-16 *** --- Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1 R-Squared: 0.09619 Root Mean Squared Error: 8.075 > SparklyR provides R interfaces to Spark’s distributed machine learning algorithms (MLlib) Computations happing in the Spark cluster, not in R
- 33. In summary • Embarrassingly parallel (small): foreach + local backend • Embarrassingly parallel (big): foreach + cluster backend – Create & use clusters in Azure with doAzureParallel • Big, distributed data: sparklyr – Create Spark clusters with sparklyr in Azure with azkt 34 @revodavid
- 34. Get your links here • Code for birthday problem simulation: cda.ms/7d • Using the foreach package: cda.ms/6Q • Get doAzureParallel: cda.ms/7w • Get aztk (for sparklyr): cda.ms/7x • Sparklyr: spark.rstudio.com • Free Azure account with $200 credit: cda.ms/7v 35 @revodavid Thank you!