You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@systemml.apache.org by mb...@apache.org on 2018/08/08 21:35:07 UTC
systemml git commit: [SYSTEMML-2487] Fix parfor parallelism rewrite
(awareness native libs)
Repository: systemml
Updated Branches:
refs/heads/master 7fb38b9b0 -> 9ae0a9167
[SYSTEMML-2487] Fix parfor parallelism rewrite (awareness native libs)
In case parfor does not consume all the available parallelism, we
propagate this parallelism down to individual operations with slight
(max 50%) over-provisioning. For example, if we have 80 vcores, and
parfor is assigned k=47, we still assign k=2 to individual operations.
However, with native BLAS or DNN operations this causes JVM crashes due
to SIGFPE with Intel MKL and too few memory buffers with OpenBLAS.
Hence, this patch improves the related parfor rewrite to select the
degree of parallelism with awareness of loaded native libraries.
Project: http://git-wip-us.apache.org/repos/asf/systemml/repo
Commit: http://git-wip-us.apache.org/repos/asf/systemml/commit/9ae0a916
Tree: http://git-wip-us.apache.org/repos/asf/systemml/tree/9ae0a916
Diff: http://git-wip-us.apache.org/repos/asf/systemml/diff/9ae0a916
Branch: refs/heads/master
Commit: 9ae0a916777dff0fa36b769f773aed6f64253f1c
Parents: 7fb38b9
Author: Matthias Boehm <mb...@gmail.com>
Authored: Wed Aug 8 14:36:15 2018 -0700
Committer: Matthias Boehm <mb...@gmail.com>
Committed: Wed Aug 8 14:36:54 2018 -0700
----------------------------------------------------------------------
.../controlprogram/parfor/opt/OptimizerRuleBased.java | 9 +++++++--
1 file changed, 7 insertions(+), 2 deletions(-)
----------------------------------------------------------------------
http://git-wip-us.apache.org/repos/asf/systemml/blob/9ae0a916/src/main/java/org/apache/sysml/runtime/controlprogram/parfor/opt/OptimizerRuleBased.java
----------------------------------------------------------------------
diff --git a/src/main/java/org/apache/sysml/runtime/controlprogram/parfor/opt/OptimizerRuleBased.java b/src/main/java/org/apache/sysml/runtime/controlprogram/parfor/opt/OptimizerRuleBased.java
index e7c7a6e..7adf2d8 100644
--- a/src/main/java/org/apache/sysml/runtime/controlprogram/parfor/opt/OptimizerRuleBased.java
+++ b/src/main/java/org/apache/sysml/runtime/controlprogram/parfor/opt/OptimizerRuleBased.java
@@ -82,6 +82,7 @@ import org.apache.sysml.runtime.controlprogram.caching.MatrixObject.UpdateType;
import org.apache.sysml.runtime.controlprogram.context.ExecutionContext;
import org.apache.sysml.runtime.controlprogram.context.SparkExecutionContext;
import org.apache.sysml.runtime.util.ProgramConverter;
+import org.apache.sysml.utils.NativeHelper;
import org.apache.sysml.runtime.controlprogram.parfor.ResultMergeLocalFile;
import org.apache.sysml.runtime.controlprogram.parfor.opt.CostEstimator.ExcludeType;
import org.apache.sysml.runtime.controlprogram.parfor.opt.CostEstimator.TestMeasure;
@@ -1335,7 +1336,7 @@ public class OptimizerRuleBased extends Optimizer
mhop.setMaxNumThreads(opsK); //set max constraint in hop
c.setK(opsK); //set optnode k (for explain)
//need to recompile SB, if changed constraint
- recompileSB = true;
+ recompileSB = true;
}
//for all other multi-threaded hops set k=1 to simply debugging
else if( h instanceof MultiThreadedHop ) {
@@ -1371,7 +1372,11 @@ public class OptimizerRuleBased extends Optimizer
//compute max remaining operations parallelism k with slight over-provisioning
//such that k * tmpK <= 1.5 * opsK; note that if parfor already exploits the
//maximum parallelism, this will not introduce any over-provisioning.
- return (int)Math.max(Math.round((double)opsK / tmpK), 1);
+ //(when running with native BLAS/DNN libraries, we disable over-provisioning
+ //to avoid internal SIGFPE and allocation buffer issues w/ MKL and OpenBlas)
+ return NativeHelper.isNativeLibraryLoaded() ?
+ (int) Math.max(opsK / tmpK, 1) :
+ (int) Math.max(Math.round((double)opsK / tmpK), 1);
}
///////