Skip to content

Commit

Permalink
Support for Spark Dataset added (#70)
Browse files Browse the repository at this point in the history 
* Spark 2.1 support

* Sample codegen on top of DS

* working sample

* v2

* Stable jcuda codegen

* Removed println

* Removed the hardcoded size for partitions

* code cleanup

* Pinned Memory

* Pinned Memory for GPUOUT

* Array support

* Array support 1

* Array Support

* support for const array

* made logical plan to produce & consume objects

* incorporated review comments (#46)

* incorporated review comments

* include GpuDSArrayMult.scala sample file

* GPU Caching Feature Added for GpuEnabler Dataset APIs (#49)

* incorporated review comments

* enable cache prelim

* push cache setting to codegen part prelims

* cached gpuPtrs handled in codegen part prelims

* uncacheGPU for dataset included prelims

* GPU dimensions support with provision for multiple stage execution prelims

* handle constants - prelims

* code package movement

* handle constants - prelims2

* add comments - prelims2

* api's simplified

* reorder gpuoutput in the args list

* variable can be both GPUINPUT & GPUOUTPUT

* variable can be both GPUINPUT & GPUOUTPUT bug fixes

* started with testcase addition for GPU operations on Dataset

* new testcases addition

* cache bug fixes & new testcases added

* bugs related to multithreading fixed

* code comments added

* code caching added for autogenerated code

* cleaned up examples

* cleaned up examples

* Alloc host memory if the cached data is part of output

* perf issue patched

* minor mistake

* include compare perf example

* Performance & other misc patches (#50)

* guava dependency wrong version fixed

* prelim commit 1

* bug fixes

* performance bug fixes

* patch for perfDebug sample prog (#53)

* API to load data into GPU (#56)

* patch for perfDebug sample prog

* Added loadGpu API

* resolve conflicts

* Added loadGpu API

* resolve conflicts

* samples invoke loadGpu

* Performance optimization on speed and memory (#59)

* auto caching added between GPU calls

* toggle autocache gpu

* logging added

* testcase added for gpuonly cache

* optimize mem alloc

* reduce mem requirement

* handle buffer underflow

* optimize code flow

* Support for Multi Dimension as input GPU Grid dimension (#60)

* Additional GPU kernels added (#62)

* auto caching added between GPU calls

* toggle autocache gpu

* logging added

* testcase added for gpuonly cache

* optimize mem alloc

* reduce mem requirement

* optimize code flow

* sample program for logistic reg added

* kmeans partial code drop

* stabilize kmeans example

* gpu memory limit check added

* stabilize kmeans example

* GpuKMeans example read from file

* GpuKMeans example dump cluster index

* minor patches

* introduce gpuptr meta info & bug fixes

* child gpuptrs cleanup bug fixed

* GpuKMeans variants added

* optimize loadGpu

* optimize loadGpu

* optimize loadGpu

* optimize loadGpu

* modify perfDebug to remove sleep

* code cleanup

* remove duplicate testcase

* #Issue63: Support for CUDA8.0 and related jcuda libraries (#64)

* changes for jcuda 8.0

* Address performance issue in map GPU functions (#67)

* Heterogeneous Environment Support (#69)

* Heterogeneous Environment Support

* changes for Heterogeneous env and null dataset assertion

* GPUEnabler version changed to 2.0.0

* GPUEnabler version changed to 2.0.0

* Update README.md
  • Loading branch information
@josiahsams @kmadhugit
josiahsams authored and kmadhugit committed Oct 23, 2017
1 parent 05236aa commit 159bc74
Show file tree
Hide file tree
Showing 689 changed files with 16,999 additions and 5,427 deletions.
70 changes: 41 additions & 29 deletions README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -7,42 +7,45 @@ The following capabilities are provided by this package,
* convert the data from partitions to a columnar format so that
it can be easily fed into the GPU kernel.
* provide support for caching inside GPU for optimized performance.
* New: Supports Spark Datasets starting ver. 2.0.0.


## Requirements

This package is compatible with Spark 1.5+ and scala 2.10
This package is compatible with Spark 1.5+ and scala 2.10+


| Spark Version | Scala Version | Compatible version of Spark GPU |
| ------------- |-----------------|----------------------|
| `2.1+` | `2.11` |`2.0.0` |
| `1.5+` | `2.10` |`1.0.0` |


## Linking

You can link against this library (for Spark 1.5+) in your program at the following coordinates:

Using SBT:

```
libraryDependencies += "com.ibm" %% "gpu-enabler_2.10" % "1.0.0"
libraryDependencies += "com.ibm" %% "gpu-enabler_2.11" % "2.0.0"
```

Using Maven:

```xml
<dependency>
<groupId>com.ibm</groupId>
<artifactId>gpu-enabler_2.10</artifactId>
<version>1.0.0</version>
<artifactId>gpu-enabler_2.11</artifactId>
<version>2.0.0</version>
</dependency>
```

This library can also be added to Spark jobs launched through `spark-shell` or `spark-submit` by using the `--packages` command line option.
For example, to include it when starting the spark shell:

```
$ bin/spark-shell --packages com.ibm:gpu-enabler_2.10:1.0.0
$ bin/spark-shell --packages com.ibm:gpu-enabler_2.11:2.0.0
```

Unlike using `--jars`, using `--packages` ensures that this library and its dependencies will be added to the classpath.
Expand All @@ -54,8 +57,8 @@ The `--packages` argument can also be used with `bin/spark-submit`.
* Support x86_64 and ppc64le
* Support OpenJDK and IBM JDK
* Support NVIDIA GPU with CUDA (we confirmed with CUDA 7.0)
* Support CUDA 7.0 and 7.5 (should work with CUDA 6.0 and 6.5)
* Support scalar variables in primitive scalar types and primitive array in RDD
* Support CUDA8.0, CUDA 7.0 and 7.5 (should work with CUDA 6.0 and 6.5)
* Support scalar variables in primitive scalar types and primitive array in Spark RDD & Dataset.

## Examples

Expand All @@ -64,6 +67,8 @@ The recommended way to load and use GPU kernel is by using the following APIs, w
The package comes with a set of examples. They can be tried out as follows,
`./bin/run-example GpuEnablerExample`

Sample programs can be found [here](https://github.com/IBMSparkGPU/GPUEnabler/blob/master/examples/src/main/scala/com/ibm/gpuenabler/).

The Nvidia kernel used in these sample programs is available for download
[here](https://github.com/IBMSparkGPU/GPUEnabler/blob/master/examples/src/main/resources/GpuEnablerExamples.ptx).
The source for this kernel can be found [here](https://github.com/IBMSparkGPU/GPUEnabler/blob/master/examples/src/main/resources/GpuEnablerExamples.cu).
Expand All @@ -73,34 +78,43 @@ The source for this kernel can be found [here](https://github.com/IBMSparkGPU/GP

```scala
// import needed for the Spark GPU method to be added
import com.ibm.gpuenabler.CUDARDDImplicits._
import com.ibm.gpuenabler.CUDAFunction
import com.ibm.gpuenabler.CUDADSImplicits._
import com.ibm.gpuenabler.DSCUDAFunction

// Load a kernel function from the GPU kernel binary
val ptxURL = SparkGPULR.getClass.getResource("/GpuEnablerExamples.ptx")

val mapFunction = new CUDAFunction(
"multiplyBy2", // Native GPU function to multiple a given no. by 2 and return the result
Array("this"), // Input arguments
Array("this"), // Output arguments
ptxURL)

val reduceFunction = new CUDAFunction(
"sum", // Native GPU function to sum the input argument and return the result
Array("this"), // Input arguments
Array("this"), // Output arguments
ptxURL)
val ptxURL = "/GpuEnablerExamples.ptx"

val mulFunc = DSCUDAFunction(
"multiplyBy2", // Native GPU function to multiple a given no. by 2 and return the result
Seq("value"), // Input arguments
Seq("value"), // Output arguments
ptxURL)

val dimensions = (size: Long, stage: Int) => stage match {
case 0 => (64, 256, 1, 1, 1, 1)
case 1 => (1, 1, 1, 1, 1, 1)
}
val gpuParams = gpuParameters(dimensions)

val sumFunc = DSCUDAFunction(
"suml",
Array("value"),
Array("value"),
ptxURL,
Some((size: Long) => 2),
Some(gpuParams), outputSize=Some(1))

// 1. Apply a transformation ( multiple all the values of the RDD by 2)
// (Note: Conversion of row based formatting to columnar format which is understandable
// by GPU is done internally )
// 2. Trigger a reduction action (sum up all the values and return the result)
val output = sc.parallelize(1 to n, 1)
.mapExtFunc((x: Int) => 2 * x, mapFunction)
.reduceExtFunc((x: Int, y: Int) => x + y, reduceFunction)

val output = ss.range(1, N+1, 1, 10)
.mapExtFunc(_ * 2, mulFunc)
.reduceExtFunc(_ + _, sumFunc)
```

### Java API
### Java API (Supported only on RDD)

```java
// import needed for the Spark GPU method to be added
Expand Down Expand Up @@ -164,6 +178,4 @@ Note:
## Testing
To run the tests, you should run `mvn test`.

## On-going work
* Leverage existing schema awareness in DataFrame/DataSet
* Provide new DataFrame/DataSet operators to call CUDA Kernels

11 changes: 8 additions & 3 deletions bin/run-example
View file
Original file line number Diff line number Diff line change
Expand Up @@ -31,14 +31,19 @@ else
exit 1
fi

EXAMPLE_MASTER=${MASTER:-"local[*]"}
if [ -n "$1" ]; then
EXAMPLE_ARGS="$@"
EXAMPLE_ARGS=`echo $EXAMPLE_ARGS| sed "s/ \+/|/g"`
fi


if [[ ! $EXAMPLE_CLASS == com.ibm.gpuenabler* ]]; then
EXAMPLE_CLASS="com.ibm.gpuenabler.$EXAMPLE_CLASS"
fi

echo "Executing : mvn -q scala:run -DmainClass=$EXAMPLE_CLASS -DaddArgs=\"$EXAMPLE_MASTER\""
echo "Executing : mvn -q scala:run -DmainClass=$EXAMPLE_CLASS -DaddArgs=\"$EXAMPLE_ARGS\""

echo $EXAMPLES_DIR
cd $EXAMPLES_DIR

mvn -q scala:run -DmainClass=$EXAMPLE_CLASS -DaddArgs="$EXAMPLE_MASTER"
mvn -q scala:run -DmainClass=$EXAMPLE_CLASS -DaddArgs="$EXAMPLE_ARGS"
3 changes: 3 additions & 0 deletions compile.sh
View file
Original file line number Diff line number Diff line change
Expand Up @@ -27,6 +27,9 @@ if [[ $CUDA != 0 ]]; then
elif [[ $CUDAVER == "7.0" ]]; then
echo "Identified CUDA version is 7.0a"
CUDAVER="jcuda70a"
elif [[ $CUDAVER == "8.0" ]]; then
echo "Identified CUDA version is 8.0"
CUDAVER="jcuda80"
else
echo "Not a supported version. Installation will fallback to default"
fi
Expand Down
Loading

0 comments on commit 159bc74

Please sign in to comment.