Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

IBM Runtimes Performance Observations with Apache Spark


Published on

In this talk presented at the Spark London meetup on the 23rd of November 2016 I have detailed our findings in IBM's Runtime Technologies department around Apache Spark. I share best practices we observed by profiling Spark on a variety of workloads I have covered and help Spark users to profile their own applications. I've also touched on how anybody can develop using fast networking capabilities (RDMA) and can achieve substantial performance speedups using GPUs.

Published in: Software
  • Be the first to comment

IBM Runtimes Performance Observations with Apache Spark

  1. 1. © 2016 IBM Corporation 1 ● Sharing observations from IBM Runtimes ● High level techniques and tools ● Writing efficient code ● Hardware accelerators RDMA for networking GPUs for computation Apache Spark performance Adam Roberts IBM Runtimes, Hursley, UK
  2. 2. © 2016 IBM Corporation 2 Workloads we're especially interested in ● HIBench ● SparkSqlPerf, all 100 TPC-DS queries ● Real customer applications ● PoCs and Spark demos
  3. 3. © 2016 IBM Corporation 3 What I will be covering ✔ Best practices for Java/Scala code ✔ Writing code that works well with a JIT compiler ✔ Profiling techniques you can use ✔ How to use RDMA for fast networking ✔ How to use GPUs for fast data processing ✔ How we can use the above to dramatically increase our Spark performance: get results faster ✔ Package for anyone to try
  4. 4. © 2016 IBM Corporation 4 What I won't be covering ● High-level application design decisions ● Avoiding the shuffle: knowing which Spark methods to use ● File systems, operating systems, and file types to use ● Conventional Spark options e.g. spark.shuffle.*, compression codecs, spark.memory.*, spark.rpc.*, spark.streaming.*, spark.dynamicAllocation.* ● Java options in depth: though a matching -Xms and -Xmx shows good results in Spark 2 (omitted by default in a PR) and we use the Kryo serializer
  5. 5. © 2016 IBM Corporation 5 Tooling we use, all freely available ● Health Center ● TPROF with Visual Performance Analyzer ● GCMV: garbage collection and memory visualizer ● MAT: diagnose and resolve memory leaks ● Linux perf tools ● Jenkins, Slack, Maven, ScalaTest, Eclipse, Intellij Community Edition
  6. 6. © 2016 IBM Corporation 6 Profiling Spark with Healthcenter -Xhealthcenter:level=headless
  7. 7. © 2016 IBM Corporation 7 Profiling Java with TPROF -agentlib:jprof=tprof
  8. 8. © 2016 IBM Corporation 8 Tips for performance in Java and Scala ● Locals are faster than globals Can prove closed set of storage readers / modifers Fields and statics slow; parameters and locals fast ● Constants are faster than variables Can copy constants inline or across memory caches Java’s final and Scala’s val are your friends ● private is faster than public private methods can't be dynamically redefined protected and “package private” just as slow as public ● Small methods (≤100 bytecodes) are good More opportunities to in-line them ● Simple is faster than complex Easier for the JIT to reason about the effects ● Limit extension points and architectural complexity when practical Makes call sites concrete.
  9. 9. © 2016 IBM Corporation 9 Scala has lots of features, not all of them are fast ● Understand the implementation of Scala language features – use them judiciously ● Reduce uncertainty for the compiler in your coding style: use type ascription, avoid ambiguous polymorphism ● Stick to common coding patterns - the JIT is tuned for them, as new workloads emerge the latest JITs will change too ● Focus on performance hotspots using the profiling tools I mentioned ● Too much emphasis on performance can compromise maintainability! ● Too much emphasis on maintainability can compromise performance!
  10. 10. © 2016 IBM Corporation 10 Idiomatic vs imperative Scala
  11. 11. © 2016 IBM Corporation 11 for (x <- 1 to 10) { println(“Value of x: “ + x) } val values = List(1,2,3,4,5,6) for (x <- values) { println(“Value of x: “ + x) } val x = 1 while (x <= 10) { println(“Value of x: “ + x) x = x + 1 } val values = List(1,2,3,4,5,6) var x = 0 while (x < values.length) { println(“Value of x: “ + values(x)) x = x + 1 } Scala for loops
  12. 12. © 2016 IBM Corporation 12 Takeaway is to avoid boxing/unboxing (involves object allocation) – avoid collections of type AnyRef! Know your types Convert to AnyRef with care
  13. 13. © 2016 IBM Corporation 13 ● Max heap size, initial heap size, quickstart can make a big difference - for Spark 2 we've noticed that a matching -Xms and -Xmx improves performance on HiBench and SparkSqlPerf ● O* JDK has a method size bytecode limit for the JIT, ours does not, if you do use O*JDK try -XX:DontCompileHugeMethods if you find certain queries become very slow ● Experiment then profile – spend your time in what's actually used the most, not nitpicking over barely used code paths! ● for environment variables ● spark-defaults.conf for Spark settings Observations with Java options
  14. 14. © 2016 IBM Corporation 14 ● The VM searches the JAR, loads and verifies bytecodes to internal representation, runs bytecode form directly ● After many invocations (or via sampling) code gets compiled at ‘cold’ or ‘warm’ level ● An internal, low overhead sampling thread is used to identify frequently used methods ● Methods may get recompiled at ‘hot’ or ‘scorching’ levels (for more optimizations) ● Transition to ‘scorching’ goes through a temporary profiling step cold hot scorching profiling interpreter warm Java's intermediate bytecodes are compiled as required and based on runtime profiling - code is compiled 'just in time' as required - dynamic compilation can determine the target machine capabilities and app demands The JIT takes a holistic view of the application, looking for global optimizations based on actual usage patterns, and speculative assumptions
  15. 15. © 2016 IBM Corporation 15 export IBM_JAVA_OPTIONS=”-Xint“ to run without it, see the difference for yourself What a difference a JIT makes...
  16. 16. © 2016 IBM Corporation 16 ● Using type ascription ● Avoiding ambiguities ● Preferring val/final and private ● Reducing non-obvious polymorphism ● Avoiding collections of AnyRef ● Avoiding JNI Writing JIT friendly code guidelines
  17. 17. © 2016 IBM Corporation 17 0 0.2 0.4 0.6 0.8 1 1.2 1.4 Java8 GA Java8 SR1 Java8 SR2 Java8 SR2 FP 10 Java 8 SR3 proto 1/Geometric mean of HiBench time on Linux on Z (32cores)and25Gheap IBM JDK8 SR3 (tuned) IBM JDK8 SR3 (out of the box) PageRank 160% 148% Sleep 187% 113% Sort 103% 147% WordCount 130% 146% Bayes 100% 91% Terasort 160% 131% 1/Geometric mean of HiBench time on zLinux 32 cores, 25G heap Improvements in successive IBM Java 8 releases Performance compared with OpenJDK 8 HiBench huge, Spark 2.0.1, Linux Power8 12 core * 8-way SMT 1.35x Can we tune a JDK to work well with Spark?
  18. 18. © 2016 IBM Corporation 18 Contributing back changes to Spark core [SPARK-18231]: optimising the SizeEstimator Hot methods in these classes with PageRank: ● [SPARK-18196]: optimising CompactBuffer ● [SPARK-18197]: optimising AppendOnlyMap ● [SPARK-18224]: optimising PartitionedPairBuffer Blog post for more details here
  19. 19. © 2016 IBM Corporation 19 Takeaways ● Profile Lots In Pre-production (PLIP) Our tools will help ● Not all Java implementations are the same ● Remember to focus on what's hot in the profiles! Make a change, rebuild and reprofile, repeat ● Many ways to achieve the same goal in Scala, use convenient code in most places and simple imperative code for what's critical
  20. 20. © 2016 IBM Corporation 20 We can only get so far writing fast code, I'll talk about RDMA for fast networking and how we can use GPUs for fast processing Beyond optimum code...
  21. 21. © 2016 IBM Corporation 21 ● Feature available in our SDK for Java: Java Sockets over RDMA ● Requires RDMA capable network adapter: you don't need Infiniband, can use RDMA over Converged Ethernet (RoCE) ● Investigating other RDMA implementations so we can avoid marshalling and data (de)serialization costs ● Breaking Sorting World Records with RDMA ● Getting started with RDMA Remote Direct Memory Access (RDMA)
  22. 22. © 2016 IBM Corporation 22 Spark VM Buffer Off Heap Buffer Spark VM Buffer Off Heap Buffer Ether/IB SwitchRDMA NIC/HCA RDMA NIC/HCA OS OS DMA DMA (Z-Copy) (Z-Copy) (B-Copy)(B-Copy) Acronyms: Z-Copy – Zero Copy B-Copy – Buffer Copy IB – InfiniBand Ether - Ethernet NIC – Network Interface Card HCA – Host Control Adapter ● Low-latency, high-throughput networking ● Direct 'application to application' memory pointer exchange between remote hosts ● Off-load network processing to RDMA NIC/HCA – OS/Kernel Bypass (zero-copy) ● Introduces new IO characteristics that can influence the Apache Spark transfer plan Spark node #1 Spark node #2
  23. 23. © 2016 IBM Corporation 23 TCP/IP RDMA RDMA exhibits improved throughput and reduced latency Our JVM makes RDMA available transparently via APIs (JsoR) or explicitly via jVerbs calls
  24. 24. © 2016 IBM Corporation 24 32 cores (1 master, 4 nodes x 8 cores/node, 32GB Mem/node), IBM Java 8 0 100 200 300 400 500 600 Spark HiBench TeraSort [30GB] Execution Time (sec) 556s 159s TCP/IP JSoR Elapsed time with 30 GB of data, 32 GB executor
  25. 25. © 2016 IBM Corporation 25 TPC-H benchmark 100Gb 30% improvement in database operations Shuffle-intensive benchmarks show 30% - 40% better performance with RDMA HiBench PageRank 3Gb 40% faster, lower CPU usage 32 cores (1 master, 4 nodes x 8 cores/node, 32GB Mem/node)
  26. 26. © 2016 IBM Corporation 26 Why? ● Faster computation of results or the ability to process more data in the same amount of time – we want to improve accuracy of systems and free up CPUs for boring work ● GPUs becoming available in servers and many modern computers for us to use ● Drivers and SDKs freely available Fast computation: Graphics Processing Units
  27. 27. © 2016 IBM Corporation 27 z13 BigInsights How popular is Java?
  28. 28. © 2016 IBM Corporation 28 AlphaGo: 1,202 CPUs, 176 GPUs Titan: 18,688 GPUs, 18,688 CPUs CERN and Geant: reported to be using GPUs Oak Ridge, IBM “the world's fastest supercomputers by 2017”: two, $325m Databricks: recent blog post mentions deep learning with GPUs and Spark Who's interested in GPUs?
  29. 29. © 2016 IBM Corporation 29 GPUs excel at executing many of the same operations at once (Single Instruction Multiple Data programming) We'll program using CUDA or OpenCL – like C and C++ and we'll write JNI code to access data in our Java world using the GPU We'll run code on computers that are shipped with graphics cards, there are free CUDA drivers for x86-64 Windows, Linux, and IBM's Power LE, OpenCL drivers, SDK and source also widely available CPUGPU
  30. 30. © 2016 IBM Corporation 30 Assume we have an integer array in CUDA C called myData Allocate space on the GPU (device side) using cudaMalloc, this returns a pointer we'll use later. Let's say we call this variable myDataOnGPU Copy myData from the host to your allocated space (myDataOnGPU) using cudaMemcpyHostToDevice Process your data on the GPU in a kernel (we use <<< and >>>) Copy the result back (what's at myDataOnGPU replaces myData on the host) using cudaMemcpyDeviceToHost How do we use a GPU?
  31. 31. © 2016 IBM Corporation 31 __global__ void addingKernel(int* array1, int* array2){ array1[threadIdx.x] += array2[threadIdx.x]; } __global__ : it's a function we can call on the host (CPU), it's available to be called from everywhere How is the data arranged and how can I access it? Sequentially, a kernel runs on a grid (blocks x threads) and it's how we can run many threads that work on different parts of the data int*? A pointer to integers we've copied to the GPU threadIdx.x? We use this as an index to our array, remember lots of threads run on the GPU. Access each item for our example using this
  32. 32. © 2016 IBM Corporation 32 ● Assume we have an integer array on the Java heap: myData ● Create a native method in Java or Scala ● Write .cpp or .c code with a matching signature for your native method ● In your native code, use JNI to get a pointer to your data ● With this pointer, we can figure out how much memory we need ● Allocate space on the GPU (device side): cudaMalloc, returns myDataOnTheGPU ● Copy myData to your allocated space (myDataOnTheGPU) using cudaMemcpyHostToDevice ● Process your data on the GPU in a kernel (look for <<< and >>>) ● Copy the result back (what's now at myDataOnTheGPU replaces myData on the host) using cudaMemcpyDeviceToHost ● Release the elements (updating your JNI pointer so the data in our JVM heap is now the result) How would we use a GPU with Java or Scala? Easier ways?
  33. 33. © 2016 IBM Corporation 33 Our option: 40,000,000 400,000,000 Ints sorted per second Array length 400m per sec 40m per sec Sorting throughput for ints 30,000 300,000 3,000,000 30,000,000 300,000,000 Details online here Making it simple: Java class library modification
  34. 34. © 2016 IBM Corporation 34 Our option: -Xjit:enableGPU Making it simple: Java JIT compiler modification Use an IntStream and specify our JIT option Primitive types can be used (byte, char, short, int, float, double, long)
  35. 35. © 2016 IBM Corporation 35 Measured performance improvement with a GPU using four programs using 1-CPU-thread sequential execution 160-CPU-thread parallel execution Experimental environment used IBM Java 8 Service Release 2 for PowerPC Little Endian Two 10-core 8-SMT IBM POWER8 CPUs at 3.69 GHz with 256GB memory (160 hardware threads in total) with one NVIDIA Kepler K40m GPU (2880 CUDA cores in total) at 876 MHz with 12GB global memory (ECC off) Performance of GPU enabled lambdas
  36. 36. © 2016 IBM Corporation 36 Name Summary Data size Data type MM A dense matrix multiplication: C = A.B 1024 x 1024 (1m) items double SpMM As above, sparse matrix 500k x 500k (250m) items double Jacobi2D Solve an equation using the Jacobi method 8192 x 8192 (67m) items double LifeGame Conway's Game of Life with 10k iterations 512 x 512 (262k) items byte
  37. 37. © 2016 IBM Corporation 37 This shows GPU execution time speedup amounts compared to what's in blue (1 CPU thread) and yellow (160 CPU threads) The higher the bar, the bigger the speedup!
  38. 38. © 2016 IBM Corporation 38 Similar to JCuda but provides a higher level abstraction, production ready and supported by us ● No arbitrary and unrestricted use of Pointer(long) ● Still feels like Java instead of C Write your kernel and compile it into a fat binary nvcc --fatbin Add your Java code import*; import*; Load your fat binary module = new Loader().loadModule("AdamDoubler.fatbin", device); Build and run as you would any other Java application Making it simple: CUDA4J API
  39. 39. © 2016 IBM Corporation 39 Only doubling integers; could be any use case where we're doing the same operation to lots of elements at once Full code listing at the end, Javadocs: search IBM Java 8 API * Tip: the offsets are byte offsets, so you'll want your index in Java * the size of the object! module = new Loader().loadModule("AdamDoubler.fatbin", device); kernel = new CudaKernel(module, "Cuda_cuda4j_AdamDoubler_Strider"); stream = new CudaStream(device); numElements = 100; myData = new int[numElements]; Util.fillWithInts(myData); CudaGrid grid = Util.makeGrid(numElements, stream); buffer1 = new CudaBuffer(device, numElements * Integer.BYTES); buffer1.copyFrom(myData); Parameters kernelParams = new Parameters(2).set(0, buffer1).set(1, numElements); kernel.launch(grid, kernelParams); buffer1.copyTo(myData); If our dynamically created grid dimensions are too big we need to break down the problem and use the slice* API: doChunkingProblem() Our kernel, compiles into AdamDoubler.fatbin
  40. 40. © 2016 IBM Corporation 40 ● Recommendation algorithms such as ● Alternating Least Squares ● Movie recommendations on Netflix ● Recommended purchases on Amazon ● Similar songs with Spotify ● Clustering algorithms such as ● K-means (unsupervised learning) ● Produce clusters from data to determine which cluster a new item can be categorised as ● Identify anomalies: transaction fraud or erroneous data ● Classification algorithms such as ● Logistic regression ● Create a model that we can use to predict where to plot the next item in a sequence ● Healthcare: predict adverse drug reactions based on known interactions between similar drugs Improving MLlib
  41. 41. © 2016 IBM Corporation 41 ● Under the covers optimisation, set the spark.mllib.ALS.useGPU property ● Full paper: ● Full implementation: Netflix 1.5 GB 12 threads, CPU 64 threads, CPU GPU Intel, IBM Java 8 676 seconds N/A 140 seconds Currently always sends work to a GPU regardless of size, remember we have limited device memory! 2x Intel(R) Xeon(R) CPU E5-2667 v2 @ 3.30GHz, 16 cores in the machine (SMT-2), 256 GB RAM vs 2x Nvidia Tesla K80Ms Also available for Power LE. Improving Alternating Least Squares
  42. 42. © 2016 IBM Corporation 42 We modified the existing ALS (.scala) implementation's computeFactors method ● Added code to check if spark.mllib.ALS.useGPU is set ● If set we'll then call our native method written to ue JNI (.cpp) ● Our JNI method calls native CUDA (.cu) method ● CUDA used to send our data to the GPU, calls our kernel, returns the results over JNI back to the Java heap ● Built with our Spark distribution and the shared library is included: ● Remember this will require the CUDA runtime (libcudart) and a capable GPU ALS.scala computeFactors CuMFJNIInterface.cpp
  43. 43. © 2016 IBM Corporation 43 We can send code to a GPU with APIs or if we make substantial changes to existing implementations, but we can also make our changes at a higher level to be more pervasive Input: user application using DataFrame or Datasets, data stored in Parquet format for now ✔ Spark with Tungsten. Uses UnsafeRow and, sun.misc.unsafe, idea is to bring Spark closer to the hardware than previously, exploit CPUA caches, improved memory and CPU efficiency, reduce GC times, avoid Java object overheads – good deep dive here ✔ Spark with Catalyst. Optimiser for Spark SQL APIs, good deep dive here, transforms a query plan (abstraction of a user's program) into an optimised version, generates optimised code with Janino compiler ✔ Spark with our changes: Java and core Spark class optimisations, optimised JIT Pervasive GPU opportunities for Spark
  44. 44. © 2016 IBM Corporation 44 Output: generated code able to leverage auto-SIMD and GPUs We want generated code that: ✔ has a counted loop, e.g. one controlled by an automatic induction variable that increases from a lower to an upper bound ✔ accesses data in a linear fashion ✔ has as few branches as possible (simple for the GPU's kernel) ✔ does not have external method calls or contains only calls that can be easily inlined These help a JIT to either use auto-SIMD capabilities or GPUs
  45. 45. © 2016 IBM Corporation 45 Problems 1) Data representation of columnar storage (CachedBatch with Array[Byte]) isn't commonly used 2) Compression schemes are specific to CachedBatch, limited to just several data types 3) Building in-memory cache involves a long code path -> virtual method calls, conditional branches 4) Generated whole-stage code -> unnecessary conversion from CachedBatch or ColumnarBatch to UnsafeRow Solutions 1) Use ColumnarBatch format instead of CachedBatch for the in-memory cache generated by the cache() method. ColumnarBatch and ColumnVector are commonly used data representations for columnar storage 2) Use a common compression scheme (e.g. lz4) for all of the data types in a ColumnVector 3) Generate code at runtime that is simple and specialized for building a concrete instance of the in- memory cache 4) Generate whole-stage code that directly reads data from columnar storage (1) and (2) increase code reuse, (3) improves runtime performance of executing the cache() method and (4) improves performance of user defined DataFrame and Dataset operations
  46. 46. © 2016 IBM Corporation 46 We propose a new columnar format: CachedColumnarBatch, that has a pointer to ColumnarBatch (used by Parquet reader) that keeps each column as OnHeapUnsafeColumnVector instead of OnHeapColumnVector. Not yet using GPUS! ● [SPARK-13805], merged into 2.0, performance improvement: 1.2x Get data from ColumnVector directly by avoiding a copy from ColumnVector to UnsafeRow when a program reads data in parquet format ● [SPARK-14098] will be merged into 2.2, performance improvement: 3.4x Generate optimized code to build CachedColumnarBatch, get data from a ColumnVector directly by avoiding a copy from the ColumnVector to UnsafeRow, and use lz4 to compress ColumnVector when df.cache() or ds.cache is executed ● [SPARK-15962], merged into 2.1, performance improvement: 1.7x Remove indirection at offsets field when accessing each element in UnsafeArrayData, reduce memory footprint of UnsafeArrayData
  47. 47. © 2016 IBM Corporation 47 ● [SPARK-16043], performance improvement: 1.2x Use a Scala primitive array (e.g. Array[Int]) instead of Array[Any] for avoiding boxing operations when putting a primitive array into GenericArrayData ● [SPARK-15985], merged into 2.1, performance improvement: 1.3x Eliminate boxing operations to put a primitive array into GenericArrayData when a Dataset program with a primitive array is ran ● [SPARK-16213], to be merged into 2.2, performance improvement: 16.6x Eliminate boxing operations to put a primitive array into GenericArrayData when a DataFrame program with a primitive array is ran ● [SPARK-17490], merged into 2.1, performance improvement: 2.0x Eliminate boxing operations to put a primitive array into GenericArrayData when a DataFrame program with a primitive array is used
  48. 48. © 2016 IBM Corporation 48 ● improving a commonly used API and contributing the code ● Ensuring generated code is in the right format for exploitation ● Making it simple for any Spark user to exploit hardware accelerators, be it GPU or auto-SIMD code for the latest processors ● We know how to build GPU based applications ● We can figure out if a GPU is available ● We can figure out what code to generate ● We can figure out which GPU to send that code to ● All while retaining Java safety features such as exceptions, bounds checking, serviceability, tracing and profiling hooks ● Assuming you have the hardware, add an option and watch performance improve: this is our goal What's in it for me?
  49. 49. © 2016 IBM Corporation 49 ● We provide an optimised JDK with Spark bundle that includes hardware offloading, profiling, a tuned JIT and is under constant development ● We can talk more about performance aspects, not covered FPGAs, CAPI flash, an improved serializer, GC optimisations, object layout, monitoring... ● Upcoming blog post at outlining the Catalyst related work ● Look out for more pull requests and involvement from IBM, we want to improve performance for everybody and maintain Spark's status ● Open to ideas and wanting to work in communities for everyone's benefit – we want to share our own observations and work with others to learn more Feedback and suggestions always welcome: Wrapping it all up...
  50. 50. © 2016 IBM Corporation Backup slides, code listing, legal information and disclaimers beyond this point
  51. 51. © 2016 IBM Corporation 51 CUDA core: part of the GPU, they execute groups of threads Kernel: a function we'll run on the GPU Grid: think of it as a CUBE of BLOCKS which lay out THREADS; our GPU functions (KERNELS) run on one of these, we need to know the grid dimensions for each kernel Threads: these do our computation, much more available than with CPUs Blocks: groups of threads Recommended reading: The nvidia-smi command tells you about your GPU's limits One GPU can have MANY CUDA cores, each CUDA core executes many threads
  52. 52. © 2016 IBM Corporation 52 CUDA grid: why is this important? To achieve parallelism: a layout of threads we can use to solve our big data problems Block dimensions? How many threads can run on a block Grid dimensions? How many blocks we can have threadIdx.x? (BLOCKS contain THREADS) Built in variable to get the current x coordinate of a given THREAD (can have an x, y, z coordinate too) blockIdx.x? (GRIDS contain BLOCKS) Built in variable to get the current x coordinate of a given BLOCK (can have an x, y, z coordinate too) Grid image is fully credited to
  53. 53. © 2016 IBM Corporation 53 For figuring out the dimensions we can use the following Java code, we want 512 threads and as many blocks as possible for the problem size int log2BlockDim = 9; int numBlocks = (numElements + 511) >> log2BlockDim; int numThreads = 1 << log2BlockDim; Size Blocks Threads 500 1 512 1,024 2 512 32,000 63 512 64,000 125 512 100,000 196 512 512,000 1,000 512 1,024,000 2,000 512
  54. 54. CUDA4J sample, part 1 of 3 import*; import*; public class Sample { private static final boolean PRINT_DATA = false; private static int numElements; private static int[] myData; private static CudaBuffer buffer1; private static CudaDevice device = new CudaDevice(0); private static CudaModule module; private static CudaKernel kernel; private static CudaStream stream; public static void main(String[] args) { try { module = new Loader().loadModule("AdamDoubler.fatbin", device); kernel = new CudaKernel(module, "Cuda_cuda4j_AdamDoubler_Strider"); stream = new CudaStream(device); doSmallProblem(); doMediumProblem(); doChunkingProblem(); } catch (CudaException e) { e.printStackTrace(); } catch (Exception e) { e.printStackTrace(); } } private static void doSmallProblem() throws Exception { System.out.println("Doing the small sized problem"); numElements = 100; myData = new int[numElements]; Util.fillWithInts(myData); CudaGrid grid = Util.makeGrid(numElements, stream); System.out.println("Kernel grid: <<<" + grid.gridDimX + ", " + grid.blockDimX + ">>>"); buffer1 = new CudaBuffer(device, numElements * Integer.BYTES); buffer1.copyFrom(myData); Parameters kernelParams = new Parameters(2).set(0, buffer1).set(1, numElements); kernel.launch(grid, kernelParams); int[] originalArrayCopy = new int[myData.length]; System.arraycopy(myData, 0, originalArrayCopy, 0, myData.length); buffer1.copyTo(myData); Util.checkArrayResultsDoubler(myData, originalArrayCopy); }
  55. 55. private static void doMediumProblem() throws Exception { System.out.println("Doing the medium sized problem"); numElements = 5_000_000; myData = new int[numElements]; Util.fillWithInts(myData); // This is only when handling more than max blocks * max threads per kernel // Grid dim is the number of blocks in the grid // Block dim is the number of threads in a block // buffer1 is how we'll use our data on the GPU buffer1 = new CudaBuffer(device, numElements * Integer.BYTES); // myData is on CPU, transfer it buffer1.copyFrom(myData); // Our stream executes the kernel, can launch many streams at once CudaGrid grid = Util.makeGrid(numElements, stream); System.out.println("Kernel grid: <<<" + grid.gridDimX + ", " + grid.blockDimX + ">>>"); Parameters kernelParams = new Parameters(2).set(0, buffer1).set(1, numElements); kernel.launch(grid, kernelParams); int[] originalArrayCopy = new int[myData.length]; System.arraycopy(myData, 0, originalArrayCopy, 0, myData.length); buffer1.copyTo(myData); Util.checkArrayResultsDoubler(myData, originalArrayCopy); } CUDA4J sample, part 2 of 3
  56. 56. private static void doChunkingProblem() throws Exception { // I know 5m doesn't require chunking on the GPU but this does System.out.println("Doing the too big to handle in one kernel problem"); numElements = 70_000_000; myData = new int[numElements]; Util.fillWithInts(myData); buffer1 = new CudaBuffer(device, numElements * Integer.BYTES); buffer1.copyFrom(myData); CudaGrid grid = Util.makeGrid(numElements, stream); System.out.println("Kernel grid: <<<" + grid.gridDimX + ", " + grid.blockDimX + ">>>"); // Check we can actually launch a kernel with this grid size try { Parameters kernelParams = new Parameters(2).set(0, buffer1).set(1, numElements); kernel.launch(grid, kernelParams); int[] originalArrayCopy = new int[numElements]; System.arraycopy(myData, 0, originalArrayCopy, 0, numElements); buffer1.copyTo(myData); Util.checkArrayResultsDoubler(myData, originalArrayCopy); } catch (CudaException ce) { if (ce.getMessage().equals("invalid argument")) { System.out.println("it was invalid argument, too big!"); int maxThreadsPerBlockX = device.getAttribute(CudaDevice.ATTRIBUTE_MAX_BLOCK_DIM_X); int maxBlocksPerGridX = device.getAttribute(CudaDevice.ATTRIBUTE_MAX_GRID_DIM_Y); long maxThreadsPerGrid = maxThreadsPerBlockX * maxBlocksPerGridX; // 67,107,840 on my Windows box System.out.println("Max threads per grid: " + maxThreadsPerGrid); long numElementsAtOnce = maxThreadsPerGrid; long elementsDone = 0; grid = new CudaGrid(maxBlocksPerGridX, maxThreadsPerBlockX, stream); System.out.println("Kernel grid: <<<" + grid.gridDimX + ", " + grid.blockDimX + ">>>"); while (elementsDone < numElements) { if ( (elementsDone + numElementsAtOnce) > numElements) { numElementsAtOnce = numElements - elementsDone; // Just do the remainder } long toOffset = numElementsAtOnce + elementsDone; // It's the byte offset not the element index offset CudaBuffer slicedSection = buffer1.slice(elementsDone * Integer.BYTES, toOffset * Integer.BYTES); Parameters kernelParams = new Parameters(2).set(0, slicedSection).set(1, numElementsAtOnce); kernel.launch(grid, kernelParams); elementsDone += numElementsAtOnce; } int[] originalArrayCopy = new int[myData.length]; System.arraycopy(myData, 0, originalArrayCopy, 0, myData.length); buffer1.copyTo(myData); Util.checkArrayResultsDoubler(myData, originalArrayCopy); } else { System.out.println(ce.getMessage()); } } } CUDA4J sample, part 3 of 3
  57. 57. CUDA4J kernel #include <stdint.h> #include <stdio.h> /** * 2D grid so we can have 1024 threads and many blocks * Remember 1 grid -> has blocks/threads and one kernel runs on one grid * In CUDA 6.5 we have cudaOccupancyMaxPotentialBlockSize which helps * * Let's say we have 100 ints to double, keeping it simple * Assume we want to run with 256 threads at once * For this size our kernel will be set up as follows * 1 grid, 1 block, 512 threads * blockDim.x is going to be 1 * threadIdx.x will remain at 0 * threadIdx.y will range from 0 to 512 * So we'll go from 1 to 512 and we'll limit access to how many elements we have */ extern "C" __global__ void Cuda_cuda4j_AdamDoubler(int* toDouble, int numElements){ int index = blockDim.x * threadIdx.x + threadIdx.y; if (index < numElements) { // Don't go out of bounds toDouble[index] *= 2; // Just double it } } extern "C" __global__ void Cuda_cuda4j_AdamDoubler_Strider(int* toDouble, int numElements){ int i = blockIdx.x * blockDim.x + threadIdx.x; if (i < numElements) { // don't go overboard toDouble[i] *= 2; } }
  58. 58. Lambda example, part 1 of 2 import; public class Lambda { private static long startTime = 0; // -Xjit:enableGPU is our JVM option public static void main(String[] args) { boolean timeIt = true; int numElements = 500_000_000; int[] toDouble = new int[numElements]; Util.fillWithInts(toDouble); myDoublerWithALambda(toDouble, timeIt); double[] toHalf = new double[numElements]; Util.fillWithDoubles(toHalf); myHalverWithALambda(toHalf, timeIt); double[] toRandomFunc = new double[numElements]; Util.fillWithDoubles(toRandomFunc); myRandomFuncWithALambda(toRandomFunc, timeIt); } private static void myDoublerWithALambda(int[] myArray, boolean timeIt) { if (timeIt) startTime = System.currentTimeMillis(); IntStream.range(0, myArray.length).parallel().forEach(i -> { myArray[i] = myArray[i] * 2; // Done on GPU for us }); if (timeIt) { System.out.println("Done doubling with a lambda, time taken: " + (System.currentTimeMillis() - startTime) + " milliseconds"); } }
  59. 59. private static void myHalverWithALambda(double[] myArray, boolean timeIt) { if (timeIt) startTime = System.currentTimeMillis(); IntStream.range(0, myArray.length).parallel().forEach(i -> { myArray[i] = myArray[i] / 2; // Again on GPU }); if (timeIt) { System.out.println("Done halving with a lambda, time taken: " + (System.currentTimeMillis() - startTime) + " milliseconds"); } } private static void myRandomFuncWithALambda(double[] myArray, boolean timeIt) { if (timeIt) startTime = System.currentTimeMillis(); IntStream.range(0, myArray.length).parallel().forEach(i -> { myArray[i] = myArray[i] * 3.142; // Double so we don't lose precision }); if (timeIt) { System.out.println("Done with the random func with a lambda, time taken: " + (System.currentTimeMillis() - startTime) + " milliseconds"); } } } Lambda example, part 2 of 2
  60. 60. Utility methods, part 1 of 2 import*; public class Util { protected static void fillWithInts(int[] toFill) { for (int i = 0; i < toFill.length; i++) { toFill[i] = i; } } protected static void fillWithDoubles(double[] toFill) { for (int i = 0; i < toFill.length; i++) { toFill[i] = i; } } protected static void printArray(int[] toPrint) { System.out.println(); for (int i = 0; i < toPrint.length; i++) { if (i == toPrint.length - 1) { System.out.print(toPrint[i] + "."); } else { System.out.print(toPrint[i] + ", "); } } System.out.println(); } protected static CudaGrid makeGrid(int numElements, CudaStream stream) { int numThreads = 512; int numBlocks = (numElements + (numThreads - 1)) / numThreads; return new CudaGrid(numBlocks, numThreads, stream); }
  61. 61. /* * Array will have been doubled at this point */ protected static void checkArrayResultsDoubler(int[] toCheck, int[] originalArray) { long errorCount = 0; // Check result, data has been copied back here if (toCheck.length != originalArray.length) { System.err.println("Something's gone horribly wrong, different array length"); } for (int i = 0; i < originalArray.length; i++) { if (toCheck[i] != (originalArray[i] * 2) ) { errorCount++; /* System.err.println("Got an error, " + originalArray[i] + " is incorrect: wasn't doubled correctly!" + " Got " + toCheck[i] + " but should be " + originalArray[i] * 2); */ } else { //System.out.println("Correct, doubled " + originalArray[i] + " and it became " + toCheck[i]); } } System.err.println("Incorrect results: " + errorCount); } } Utility methods, part 2 of 2
  62. 62. CUDA4J module loader import; import; import; import; import; import; public class Loader { private final CudaModule.Cache moduleCache = new CudaModule.Cache(); CudaModule loadModule(String moduleName, CudaDevice device) throws CudaException, IOException { CudaModule module = moduleCache.get(device, moduleName); if (module == null) { try (InputStream stream = getClass().getResourceAsStream(moduleName)) { if (stream == null) { throw new FileNotFoundException(moduleName); } module = new CudaModule(device, stream); moduleCache.put(device, moduleName, module); } } return module; } }
  63. 63. CUDA4J build script on Windows nvcc -fatbin "C:ibm8sr3gasdkbinjava" -version "C:ibm8sr3gasdkbinjavac" *.java "C:ibm8sr3gasdkbinjava" -Xmx2g Sample "C:ibm8sr3gasdkbinjava" -Xmx4g Lambda "C:ibm8sr3gasdkbinjava" -Xjit:enableGPU={verbose} -Xmx4g Lambda
  64. 64. Set the PATH to include the CUDA library. For example, set PATH=<CUDA_LIBRARY_PATH>;%PATH%, where the <CUDA_LIBRARY_PATH> variable is the full path to the CUDA library. The <CUDA_LIBRARY_PATH> variable is C:Program FilesNVIDIA GPU Computing ToolkitCUDAv7.5bin, which assumes CUDA is installed to the default directory. Note: If you are using Just-In-Time Compiler (JIT) based GPU support, you must also include paths to the NVIDIA Virtual Machine (NVVM) library, and to the NVDIA Management Library (NVML). For example, the <CUDA_LIBRARY_PATH> variable is C:Program FilesNVIDIA GPU Computing ToolkitCUDAv7.5bin;<NVVM_LIBRARY_PATH>;<NVML_LIBRARY_P ATH>. If the NVVM library is installed to the default directory, the <NVVM_LIBRARY_PATH> variable is C:Program FilesNVIDIA GPU Computing ToolkitCUDAv7.5nvvmbin. You can find the NVML library in your NVIDIA drivers directory. The default location of this directory is C:Program FilesNVIDIA CorporationNVSMI. From IBM's Java 8 docs Environment example, see the docs for details
  65. 65. Notices and Disclaimers Copyright © 2016 by International Business Machines Corporation (IBM). No part of this document may be reproduced or transmitted in any form without written permission from IBM. U.S. Government Users Restricted Rights - Use, duplication or disclosure restricted by GSA ADP Schedule Contract with IBM. Information in these presentations (including information relating to products that have not yet been announced by IBM) has been reviewed for accuracy as of the date of initial publication and could include unintentional technical or typographical errors. IBM shall have no responsibility to update this information. THIS document is distributed "AS IS" without any warranty, either express or implied. In no event shall IBM be liable for any damage arising from the use of this information, including but not limited to, loss of data, business interruption, loss of profit or loss of opportunity. IBM products and services are warranted according to the terms and conditions of the agreements under which they are provided. Any statements regarding IBM's future direction, intent or product plans are subject to change or withdrawal without notice. Performance data contained herein was generally obtained in a controlled, isolated environments. Customer examples are presented as illustrations of how those customers have used IBM products and the results they may have achieved. Actual performance, cost, savings or other results in other operating environments may vary. References in this document to IBM products, programs, or services does not imply that IBM intends to make such products, programs or services available in all countries in which IBM operates or does business. Workshops, sessions and associated materials may have been prepared by independent session speakers, and do not necessarily reflect the views of IBM. All materials and discussions are provided for informational purposes only, and are neither intended to, nor shall constitute legal or other guidance or advice to any individual participant or their specific situation. It is the customer’s responsibility to insure its own compliance with legal requirements and to obtain advice of competent legal counsel as to the identification and interpretation of any relevant laws and regulatory requirements that may affect the customer’s business and any actions the customer may need to take to comply with such laws. IBM does not provide legal advice or represent or warrant that its services or products will ensure that the customer is in compliance with any law.
  66. 66. Notices and Disclaimers (con’t) Information concerning non-IBM products was obtained from the suppliers of those products, their published announcements or other publicly available sources. IBM has not tested those products in connection with this publication and cannot confirm the accuracy of performance, compatibility or any other claims related to non-IBM products. Questions on the capabilities of non-IBM products should be addressed to the suppliers of those products. IBM does not warrant the quality of any third-party products, or the ability of any such third-party products to interoperate with IBM’s products. IBM expressly disclaims all warranties, expressed or implied, including but not limited to, the implied warranties of merchantability and fitness for a particular purpose. The provision of the information contained herein is not intended to, and does not, grant any right or license under any IBM patents, copyrights, trademarks or other intellectual property right. IBM, the IBM logo,, Bluemix, Blueworks Live, CICS, Clearcase, DOORS®, Enterprise Document Management System™, Global Business Services ®, Global Technology Services ®, Information on Demand, ILOG, LinuxONE™, Maximo®, MQIntegrator®, MQSeries®, Netcool®, OMEGAMON, OpenPower, PureAnalytics™, PureApplication®, pureCluster™, PureCoverage®, PureData®, PureExperience®, PureFlex®, pureQuery®, pureScale®, PureSystems®, QRadar®, Rational®, Rhapsody®, SoDA, SPSS, StoredIQ, Tivoli®, Trusteer®, urban{code}®, Watson, WebSphere®, Worklight®, X-Force® and System z® Z/OS, are trademarks of International Business Machines Corporation, registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. A current list of IBM trademarks is available on the Web at "Copyright and trademark information" at: Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. Databricks is a registered trademark of Databricks, Inc. Apache Spark, Apache Cassandra, Apache Hadoop, Apache Maven, Spark, Apache, any other Apache project mentioned here and the Apache product logos including the Spark logo are trademarks of The Apache Software Foundation