You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@systemml.apache.org by ni...@apache.org on 2018/11/12 13:22:27 UTC
systemml git commit: [SYSTEMML-1325] Fix a performance bug when using
GPU backend with JMLC
Repository: systemml
Updated Branches:
refs/heads/master beb1a1d19 -> c022f1a5a
[SYSTEMML-1325] Fix a performance bug when using GPU backend with JMLC
This commit simplifies the clearTemporaryMemory logic in GPU memory manager and performs aggressive cleanup to reduce the memory pressure.
Closes #842.
Project: http://git-wip-us.apache.org/repos/asf/systemml/repo
Commit: http://git-wip-us.apache.org/repos/asf/systemml/commit/c022f1a5
Tree: http://git-wip-us.apache.org/repos/asf/systemml/tree/c022f1a5
Diff: http://git-wip-us.apache.org/repos/asf/systemml/diff/c022f1a5
Branch: refs/heads/master
Commit: c022f1a5a4479c7cfd380190d20d81b7747c7b92
Parents: beb1a1d
Author: Niketan Pansare <np...@us.ibm.com>
Authored: Mon Nov 12 18:50:20 2018 +0530
Committer: Niketan Pansare <np...@us.ibm.com>
Committed: Mon Nov 12 18:50:20 2018 +0530
----------------------------------------------------------------------
.../apache/sysml/api/ScriptExecutorUtils.java | 4 ++-
.../instructions/gpu/context/GPUContext.java | 6 ++--
.../gpu/context/GPUMatrixMemoryManager.java | 38 ++++----------------
.../gpu/context/GPUMemoryManager.java | 38 ++++++++++++--------
.../instructions/gpu/context/GPUObject.java | 24 +++++++++++++
5 files changed, 61 insertions(+), 49 deletions(-)
----------------------------------------------------------------------
http://git-wip-us.apache.org/repos/asf/systemml/blob/c022f1a5/src/main/java/org/apache/sysml/api/ScriptExecutorUtils.java
----------------------------------------------------------------------
diff --git a/src/main/java/org/apache/sysml/api/ScriptExecutorUtils.java b/src/main/java/org/apache/sysml/api/ScriptExecutorUtils.java
index e2e76cb..0d072e5 100644
--- a/src/main/java/org/apache/sysml/api/ScriptExecutorUtils.java
+++ b/src/main/java/org/apache/sysml/api/ScriptExecutorUtils.java
@@ -287,6 +287,7 @@ public class ScriptExecutorUtils {
} finally { // ensure cleanup/shutdown
if (ConfigurationManager.isGPU() && !ec.getGPUContexts().isEmpty()) {
try {
+ HashSet<MatrixObject> outputMatrixObjects = new HashSet<>();
// -----------------------------------------------------------------
// The below code pulls the output variables on the GPU to the host. This is required especially when:
// The output variable was generated as part of a MLContext session with GPU enabled
@@ -302,12 +303,13 @@ public class ScriptExecutorUtils {
gpuObj.acquireHostRead(null);
}
}
+ outputMatrixObjects.add(((MatrixObject)data));
}
}
}
// -----------------------------------------------------------------
for(GPUContext gCtx : ec.getGPUContexts()) {
- gCtx.clearTemporaryMemory();
+ gCtx.clearTemporaryMemory(outputMatrixObjects);
}
} catch (Exception e1) {
exceptionThrown = true;
http://git-wip-us.apache.org/repos/asf/systemml/blob/c022f1a5/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUContext.java
----------------------------------------------------------------------
diff --git a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUContext.java b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUContext.java
index 3275099..20587b4 100644
--- a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUContext.java
+++ b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUContext.java
@@ -31,6 +31,8 @@ import static jcuda.runtime.JCuda.cudaGetDeviceCount;
import static jcuda.runtime.JCuda.cudaSetDevice;
import static jcuda.runtime.JCuda.cudaSetDeviceFlags;
+import java.util.HashSet;
+
import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.sysml.api.DMLScript.EvictionPolicy;
@@ -424,8 +426,8 @@ public class GPUContext {
memoryManager.clearMemory();
}
- public void clearTemporaryMemory() {
- memoryManager.clearTemporaryMemory();
+ public void clearTemporaryMemory(HashSet<MatrixObject> outputMatrixObjects) {
+ memoryManager.clearTemporaryMemory(outputMatrixObjects);
}
@Override
http://git-wip-us.apache.org/repos/asf/systemml/blob/c022f1a5/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMatrixMemoryManager.java
----------------------------------------------------------------------
diff --git a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMatrixMemoryManager.java b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMatrixMemoryManager.java
index f69d340..029af19 100644
--- a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMatrixMemoryManager.java
+++ b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMatrixMemoryManager.java
@@ -20,7 +20,6 @@ package org.apache.sysml.runtime.instructions.gpu.context;
import java.util.Collections;
import java.util.HashSet;
-import java.util.List;
import java.util.Set;
import java.util.stream.Collectors;
@@ -46,33 +45,6 @@ public class GPUMatrixMemoryManager {
gpuObjects.add(gpuObj);
}
- /**
- * Get list of all Pointers in a GPUObject
- * @param gObj gpu object
- * @return set of pointers
- */
- Set<Pointer> getPointers(GPUObject gObj) {
- Set<Pointer> ret = new HashSet<>();
- if(!gObj.isDensePointerNull() && gObj.getSparseMatrixCudaPointer() != null) {
- LOG.warn("Matrix allocated in both dense and sparse format");
- }
- if(!gObj.isDensePointerNull()) {
- // && gObj.evictedDenseArr == null - Ignore evicted array
- ret.add(gObj.getDensePointer());
- }
- if(gObj.getSparseMatrixCudaPointer() != null) {
- CSRPointer sparsePtr = gObj.getSparseMatrixCudaPointer();
- if(sparsePtr != null) {
- if(sparsePtr.rowPtr != null)
- ret.add(sparsePtr.rowPtr);
- else if(sparsePtr.colInd != null)
- ret.add(sparsePtr.colInd);
- else if(sparsePtr.val != null)
- ret.add(sparsePtr.val);
- }
- }
- return ret;
- }
/**
* list of allocated {@link GPUObject} instances allocated on {@link GPUContext#deviceNum} GPU
@@ -91,18 +63,22 @@ public class GPUMatrixMemoryManager {
Set<GPUObject> getGpuObjects(Set<Pointer> pointers) {
Set<GPUObject> gObjs = new HashSet<>();
for (GPUObject g : gpuObjects) {
- if (!Collections.disjoint(getPointers(g), pointers))
+ if (!Collections.disjoint(g.getPointers(), pointers))
gObjs.add(g);
}
return gObjs;
}
+ Set<GPUObject> getGpuObjects() {
+ return gpuObjects;
+ }
+
/**
* Return all pointers in the first section
* @return all pointers in this section
*/
Set<Pointer> getPointers() {
- return gpuObjects.stream().flatMap(gObj -> getPointers(gObj).stream()).collect(Collectors.toSet());
+ return gpuObjects.stream().flatMap(gObj -> gObj.getPointers().stream()).collect(Collectors.toSet());
}
/**
@@ -116,7 +92,7 @@ public class GPUMatrixMemoryManager {
return gpuObjects.stream().filter(
gObj -> (gObj.isLocked() == locked && gObj.isDirty() == dirty) ||
(gObj.mat.isCleanupEnabled() == isCleanupEnabled)).flatMap(
- gObj -> getPointers(gObj).stream()).collect(Collectors.toSet());
+ gObj -> gObj.getPointers().stream()).collect(Collectors.toSet());
}
/**
http://git-wip-us.apache.org/repos/asf/systemml/blob/c022f1a5/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMemoryManager.java
----------------------------------------------------------------------
diff --git a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMemoryManager.java b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMemoryManager.java
index d403ca7..ce22a7e 100644
--- a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMemoryManager.java
+++ b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUMemoryManager.java
@@ -39,6 +39,7 @@ import org.apache.sysml.conf.ConfigurationManager;
import org.apache.sysml.conf.DMLConfig;
import org.apache.sysml.hops.OptimizerUtils;
import org.apache.sysml.runtime.DMLRuntimeException;
+import org.apache.sysml.runtime.controlprogram.caching.MatrixObject;
import org.apache.sysml.runtime.instructions.gpu.GPUInstruction;
import org.apache.sysml.utils.GPUStatistics;
@@ -518,21 +519,28 @@ public class GPUMemoryManager {
/**
* Clears up the memory used by non-dirty pointers.
*/
- public void clearTemporaryMemory() {
- // To record the cuda block sizes needed by allocatedGPUObjects, others are cleared up.
- Set<Pointer> unlockedDirtyPointers = matrixMemoryManager.getPointers(false, true, false);
- Set<Pointer> temporaryPointers = nonIn(allPointers.keySet(), unlockedDirtyPointers);
+ public void clearTemporaryMemory(HashSet<MatrixObject> outputMatrixObjects) {
+ Set<Pointer> donotClearPointers = new HashSet<>();
+ // First clean up all GPU objects except:
+ // 1. Output matrix objects
+ // 2. GPU objects that are currently being used (i.e. locked)
+ // 3. Matrix object are
+ Set<GPUObject> allGPUObjects = new HashSet<>(matrixMemoryManager.getGpuObjects());
+ for (GPUObject gpuObj : allGPUObjects) {
+ boolean isOutput = outputMatrixObjects.contains(gpuObj.mat);
+ if(!isOutput && !gpuObj.isLocked() && gpuObj.mat.isCleanupEnabled()) {
+ gpuObj.clearData(null, gpuObj.getGPUContext().EAGER_CUDA_FREE);
+ }
+ else {
+ donotClearPointers.addAll(gpuObj.getPointers());
+ }
+ }
+
+ // Next, cleanup workspace and other temporary pointers
+ Set<Pointer> temporaryPointers = nonIn(allPointers.keySet(), donotClearPointers);
for (Pointer tmpPtr : temporaryPointers) {
guardedCudaFree(tmpPtr);
}
-
- // Also set the pointer(s) to null in the corresponding GPU objects to avoid double freeing pointers
- Set<GPUObject> gObjs = matrixMemoryManager.getGpuObjects(temporaryPointers);
- for (GPUObject g : gObjs) {
- g.jcudaDenseMatrixPtr = null;
- g.jcudaSparseMatrixPtr = null;
- removeGPUObject(g);
- }
}
/**
@@ -579,18 +587,18 @@ public class GPUMemoryManager {
if(gpuObj.isLocked()) {
numLockedGPUObjects++;
sizeOfLockedGPUObjects += gpuObj.getSizeOnDevice();
- numLockedPointers += matrixMemoryManager.getPointers(gpuObj).size();
+ numLockedPointers += gpuObj.getPointers().size();
}
else {
if(gpuObj.isDirty()) {
numUnlockedDirtyGPUObjects++;
sizeOfUnlockedDirtyGPUObjects += gpuObj.getSizeOnDevice();
- numUnlockedDirtyPointers += matrixMemoryManager.getPointers(gpuObj).size();
+ numUnlockedDirtyPointers += gpuObj.getPointers().size();
}
else {
numUnlockedNonDirtyGPUObjects++;
sizeOfUnlockedNonDirtyGPUObjects += gpuObj.getSizeOnDevice();
- numUnlockedNonDirtyPointers += matrixMemoryManager.getPointers(gpuObj).size();
+ numUnlockedNonDirtyPointers += gpuObj.getPointers().size();
}
}
}
http://git-wip-us.apache.org/repos/asf/systemml/blob/c022f1a5/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUObject.java
----------------------------------------------------------------------
diff --git a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUObject.java b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUObject.java
index edcd683..04af229 100644
--- a/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUObject.java
+++ b/src/main/java/org/apache/sysml/runtime/instructions/gpu/context/GPUObject.java
@@ -24,6 +24,8 @@ import static jcuda.runtime.JCuda.cudaMemset;
import static jcuda.runtime.cudaMemcpyKind.cudaMemcpyDeviceToDevice;
import static jcuda.runtime.cudaMemcpyKind.cudaMemcpyDeviceToHost;
+import java.util.HashSet;
+import java.util.Set;
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.atomic.LongAdder;
@@ -151,6 +153,28 @@ public class GPUObject {
jcudaSparseMatrixPtr = null;
}
+ Set<Pointer> getPointers() {
+ Set<Pointer> ret = new HashSet<>();
+ if(!isDensePointerNull() && getSparseMatrixCudaPointer() != null) {
+ LOG.warn("Matrix allocated in both dense and sparse format");
+ }
+ if(!isDensePointerNull()) {
+ // && evictedDenseArr == null - Ignore evicted array
+ ret.add(getDensePointer());
+ }
+ if(getSparseMatrixCudaPointer() != null) {
+ CSRPointer sparsePtr = getSparseMatrixCudaPointer();
+ if(sparsePtr != null) {
+ if(sparsePtr.rowPtr != null)
+ ret.add(sparsePtr.rowPtr);
+ else if(sparsePtr.colInd != null)
+ ret.add(sparsePtr.colInd);
+ else if(sparsePtr.val != null)
+ ret.add(sparsePtr.val);
+ }
+ }
+ return ret;
+ }
/**
* Convenience method to directly set the dense matrix pointer on GPU