This is an automated email from the ASF dual-hosted git repository.
mboehm7 pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/systemds.git
The following commit(s) were added to refs/heads/master by this push:
new 07486b8 [SYSTEMDS-3042] Performance matrix multiplication w/ short lhs
07486b8 is described below
commit 07486b84f725ce697d99340c3511d4bf7745ebd1
Author: Matthias Boehm <[email protected]>
AuthorDate: Sun Jun 27 15:51:14 2021 +0200
[SYSTEMDS-3042] Performance matrix multiplication w/ short lhs
While vector-matrix multiplication was already well-optimized, there was
unnecessary inefficiency with short left hand sides (e.g., left hand
sides with 2<x<=16 rows) where parallelization switched to rows in the
lhs and thus did not fully exploit the available degree of parallelism
and not utilize the available memory bandwidth wisely (redundant scans).
Instead, we now do a more careful selection when to parallelize over the
common dimension.
On a machine with 112 vcores and 750GB mem, the following script shows
performance improvements of ~10x for several important sizes:
X = rand(rows=1e6, cols=1e3); #8GB
for(i in seq(1,16)) {
y = rand(rows=i, cols=1e6)
t1 = time();
for(j in 1:100)
q = y %*% X;
print("matrix multiplication ["+i+"]: "+(time()-t1)/1e11)
}
print(sum(q))
Results [in seconds]:
matrix multiplication [1]: 0.12487190379
matrix multiplication [2]: 0.12739698981
matrix multiplication [3]: 1.14 -> 0.12981205165
matrix multiplication [4]: 1.19 -> 0.13536354042
matrix multiplication [5]: 1.21 -> 0.14075080461
matrix multiplication [6]: 1.30 -> 0.14592031624
matrix multiplication [7]: 1.25 -> 0.15245149887
matrix multiplication [8]: 1.23 -> 0.1576228154
matrix multiplication [9]: 1.27 -> 0.16084557115
matrix multiplication [10]: 1.23 -> 0.16499888784
matrix multiplication [11]: 1.23 -> 0.15556115622
matrix multiplication [12]: 1.23 -> 0.13199431345
matrix multiplication [13]: 1.25 -> 0.1411010925
matrix multiplication [14]: 1.36 -> 0.14866302508
matrix multiplication [15]: 1.24 -> 0.16179848953
matrix multiplication [16]: 1.28 -> 0.17091530956
---
src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixMult.java | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git
a/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixMult.java
b/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixMult.java
index 3ff1bb9..dd88ec9 100644
--- a/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixMult.java
+++ b/src/main/java/org/apache/sysds/runtime/matrix/data/LibMatrixMult.java
@@ -3880,10 +3880,11 @@ public class LibMatrixMult
private static boolean checkParMatrixMultRightInputRows( MatrixBlock
m1, MatrixBlock m2, int k ) {
//parallelize over rows in rhs matrix if number of rows in
lhs/output is very small
+ double jvmMem = InfrastructureAnalyzer.getLocalMaxMemory();
return (m1.rlen==1 && LOW_LEVEL_OPTIMIZATION && m2.clen>1 &&
!(m1.isUltraSparse()||m2.isUltraSparse()))
|| (m1.rlen<=16 && LOW_LEVEL_OPTIMIZATION && m2.clen>1
&& m2.rlen > m1.rlen
&& ( !m1.isUltraSparse() && !m2.sparse )
//dense-dense / sparse/dense
- && (long)k * 8 * m1.rlen * m2.clen <
MEM_OVERHEAD_THRESHOLD );
+ && (long)k * 8 * m1.rlen * m2.clen <
Math.max(MEM_OVERHEAD_THRESHOLD,0.01*jvmMem) );
}
private static boolean checkParMatrixMultRightInputCols( MatrixBlock
m1, MatrixBlock m2, int k, boolean pm2r ) {
