[FLINK-14153][ml] Add to BLAS a method that performs DenseMatrix and SparseVector multiplication. #9732

xuyang1706 · 2019-09-20T12:46:26Z

What is the purpose of the change

Previously there is "gemv" method in BLAS that performs multiplications between DenseMatrix and DenseVector. Here we add another one that performs multiplications between DenseMatrix and SparseVector.

Brief change log

Add gemv method to BLAS.
Add BLASTest that provides test cases for BLAS.

Verifying this change

This change added tests and can be verified as follows:

run test case pass

Does this pull request potentially affect one of the following parts:

Dependencies (does it add or upgrade a dependency): (yes)
The public API, i.e., is any changed class annotated with @Public(Evolving): (no)
The serializers: (no)
The runtime per-record code paths (performance sensitive): (no)
Anything that affects deployment or recovery: JobManager (and its components), Checkpointing, Yarn/Mesos, ZooKeeper: (no)
The S3 file system connector: (no)

Documentation

Does this pull request introduce a new feature? (yes)
If yes, how is the feature documented? (JavaDocs)

…SparseVector multiplication.

flinkbot · 2019-09-20T12:49:04Z

Thanks a lot for your contribution to the Apache Flink project. I'm the @flinkbot. I help the community
to review your pull request. We will use this comment to track the progress of the review.

Automated Checks

Last check on commit c247440 (Wed Dec 04 14:48:48 UTC 2019)

Warnings:

No documentation files were touched! Remember to keep the Flink docs up to date!
This pull request references an unassigned Jira ticket. According to the code contribution guide, tickets need to be assigned before starting with the implementation work.

_{Mention the bot in a comment to re-run the automated checks.}

Review Progress

❓ 1. The [description] looks good.
❓ 2. There is [consensus] that the contribution should go into to Flink.
❓ 3. Needs [attention] from.
❓ 4. The change fits into the overall [architecture].
❓ 5. Overall code [quality] is good.

Please see the Pull Request Review Guide for a full explanation of the review process.

The Bot is tracking the review progress through labels. Labels are applied according to the order of the review items. For consensus, approval by a Flink committer of PMC member is required

Bot commands

The @flinkbot bot supports the following commands:

@flinkbot approve description to approve one or more aspects (aspects: description, consensus, architecture and quality)
@flinkbot approve all to approve all aspects
@flinkbot approve-until architecture to approve everything until architecture
@flinkbot attention @username1 [@username2 ..] to require somebody's attention
@flinkbot disapprove architecture to remove an approval you gave earlier

flinkbot · 2019-09-20T12:50:23Z

CI report:

6b34d9e : SUCCESS Build
dbf3b11 : SUCCESS Build
74c52fb : SUCCESS Build
c247440 : SUCCESS Build

Bot commands

The @flinkbot bot supports the following commands:

@flinkbot run travis re-run the last Travis build

walterddr

thanks for the contribution @xuyang1706 . looks good to me overall. I left a few comments. please kindly take a look if they make sense.

walterddr · 2019-10-18T04:08:12Z

flink-ml-parent/flink-ml-lib/src/main/java/org/apache/flink/ml/common/linalg/BLAS.java

+	 */
+	public static void gemv(double alpha, DenseMatrix matA, boolean transA,
+							SparseVector x, double beta, DenseVector y) {
+		if (transA) {


Can we create a transposePreconditionChecker? feel like this would be use multiple place/times with duplicate code in such as gemv, gemm, etc.

Yes, you are right. We have created a method "gemvDimensionCheck". However, the dimension check in "gemm" and "gemv" are totally different, so "gemm" is left untouched.

walterddr · 2019-10-18T04:17:25Z

flink-ml-parent/flink-ml-lib/src/test/java/org/apache/flink/ml/common/linalg/BLASTest.java

+	private SparseVector spv2 = new SparseVector(3, new int[]{0, 2}, new double[]{1, 3});
+
+	@Test
+	public void testGemvDense() throws Exception {


why is the dense version not tested until now?

We just added more test cases.

walterddr · 2019-10-18T04:20:32Z

flink-ml-parent/flink-ml-lib/src/test/java/org/apache/flink/ml/common/linalg/BLASTest.java

+	private SparseVector spv2 = new SparseVector(3, new int[]{0, 2}, new double[]{1, 3});
+
+	@Test
+	public void testGemvDense() throws Exception {


missing validator exception test cases:

invalid dimension

invalid dimension after transpose

Thanks, we just added the test cases.

walterddr · 2019-10-18T04:25:54Z

flink-ml-parent/flink-ml-lib/src/main/java/org/apache/flink/ml/common/linalg/BLAS.java

+			for (int i = 0; i < x.indices.length; i++) {
+				int index = x.indices[i];
+				double value = alpha * x.values[i];
+				F2J_BLAS.daxpy(m, value, matA.data, index * m, 1, y.data, 0, 1);


I am not familiar with BLAS internal performance. is this faster than directly coding it up ?
the reason why I am asking is that: there's two step involved. (scal and daxpy). may have duplicate mem access ?

There are two reasons I use BLAS here.

BLAS is a mature linear algebra libarary, which pays a lot of attension to data locality, thus it is more cache friendly than naive implementation. Usually we gain a lot of permformace improvement on level-2/level-3 BLAS routine through calling native (JNI) BLAS, while F2J BLAS is better in level-1 BLAS routines.

In the case here, y is first scaled by b, then each columns of Ax is added to y. It is inevitable that y would be visited more than once.

that makes sense. I am thinking that the assumption is for example BLAS can do some sort of SIMD/MIMD optimization based on the data locality so that it can save register/cache loadings and invalidations.
It is good to call out the rationale here by having some sort of inline comment:

// relying on the native implementation of BLAS for performance.

If there's any performance issue later, we can always avoid duplicate register loading by doing the multiplication the addition and variable assignment at the same line similar to

y.data[i] = beta * y.data[i] + alpha * s;

xuyang1706 · 2019-10-18T11:23:55Z

thanks for the contribution @xuyang1706 . looks good to me overall. I left a few comments. please kindly take a look if they make sense.

Thanks @walterddr for your comments and advices. I have added more test cases and precondition checkers.

walterddr

Hi @xuyang1706 thanks for the prompt update. I think overall the patch looks good.

I only have 2 higher level questions regarding the usage of BLAS (which on a level not only relates to this PR but also related to previously committed codes). Please kindly take a look. Thanks -Rong

walterddr · 2019-10-20T23:40:03Z

flink-ml-parent/flink-ml-lib/src/main/java/org/apache/flink/ml/common/linalg/BLAS.java

+	 * y[yOffset:yOffset+n] += a * x[xOffset:xOffset+n] .
+	 */
+	public static void axpy(int n, double a, double[] x, int xOffset, double[] y, int yOffset) {
+		F2J_BLAS.daxpy(n, a, x, xOffset, 1, y, yOffset, 1);


A higher level question:
Maybe it is good to clarify when to use F2J_BLAS vs NATIVE_BLAS.

I found a very interesting read from SPARK's mailing list and seems like there are some considerations regarding this.

Thanks, I have clarified in the inline doc that we should use F2J_BLAS for level-1 routines and NATIVE_BLAS for level-2 and level-3 routines. This is also the practices adopted by SparkML.

The read from SPARK's mailing list you give here indeed shows the pitfalls when using NATIVE_BLAS. It makes clear that the underlying native BLAS libarary should not use multithreading. Fortunately, the default library uses a BLAS version provided by http://www.netlib.org, which is a single-threaded version.

walterddr · 2019-10-20T23:42:12Z

flink-ml-parent/flink-ml-lib/src/main/java/org/apache/flink/ml/common/linalg/BLAS.java

 		final int n = matA.numCols();
 		final int lda = matA.numRows();
 		final String ta = transA ? "T" : "N";
 		NATIVE_BLAS.dgemv(ta, m, n, alpha, matA.getData(), lda, x.getData(), 1, beta, y.getData(), 1);


any reason why DenseVector uses NATIVE_BLAS while the SparseVector uses F2J_BLAS

I think at least on a specific level (0,1,2,3 or up) we should probably only use one specific BLAS version unless specific reason comes up (IMO it should be some very strong justifications)

FYI: I am not sure whether this is related. Some suggestions on stack shows that there are some performance considerations coming from latest development from the JIT compiler

We consistently use F2J_BLAS for level-1 routines such as scal/axpy/asum, and use NATIVE_BLAS for level-2/level-3 routines such as gemv, gemm. As for the gemv case here, we use NATIVE_BLAS for the dense case. But for the sparse case, the BLAS library is not directly applicable because it is a library for dense linear algebra. So we implement gemv for SparseVector by hand, using F2J_BLAS to do axpy(level-1 routine) during the course.

xuyang1706 · 2019-10-21T12:37:53Z

Hi @xuyang1706 thanks for the prompt update. I think overall the patch looks good.

I only have 2 higher level questions regarding the usage of BLAS (which on a level not only relates to this PR but also related to previously committed codes). Please kindly take a look. Thanks -Rong

Thanks for your questions and discuss, @walterddr. It is better to declare our rules for using F2J_BLAS and NATIVE_BLAS. We have added clear claims in the JavaDoc. Use F2J_BLAS for level-1 routines and NATIVE_BLAS for level-2 and level-3 routines.
Regards -Xu

walterddr

Thanks @xuyang1706 for the explanation. overall it looks good to me. I think we just need to fix the comments and make sure that the explanation also goes into the code.

walterddr · 2019-10-23T04:21:50Z

flink-ml-parent/flink-ml-lib/src/main/java/org/apache/flink/ml/common/linalg/BLAS.java

+	// For level-1 routines, we use Java implementation.
 	private static final com.github.fommil.netlib.BLAS NATIVE_BLAS = com.github.fommil.netlib.BLAS.getInstance();
+
+	// For level-2 and level-3 routines, we use the native BLAS.
 	private static final com.github.fommil.netlib.BLAS F2J_BLAS = com.github.fommil.netlib.F2jBLAS.getInstance();


comment and code doesn't align.

the field name is NATIVE_BLAS for the level-1,

the field name is F2J_BLAS for level-2,3.
any naming conversion problem?

Thanks, we have corrected the java doc.

walterddr · 2019-10-23T04:24:13Z

flink-ml-parent/flink-ml-lib/src/main/java/org/apache/flink/ml/common/linalg/BLAS.java

+				double value = alpha * x.values[i];
+				F2J_BLAS.daxpy(m, value, matA.data, index * m, 1, y.data, 0, 1);


could you add the explanation you added in this PR comment into the actual code comments? I think it helps others to understand this code in the future.

Thanks, we did it.

xuyang1706 · 2019-10-24T10:20:15Z

Thanks @xuyang1706 for the explanation. overall it looks good to me. I think we just need to fix the comments and make sure that the explanation also goes into the code.

Thanks for your carefully review, @walterddr . We have refined the JavaDoc.

walterddr · 2019-10-24T16:00:33Z

@flinkbot run travis

walterddr

Thanks for the quick reply/update @xuyang1706 . I think the patch looks good overall.
I will made some minor refinement and merge it soon.

walterddr · 2019-10-24T21:49:25Z

flink-ml-parent/flink-ml-lib/src/main/java/org/apache/flink/ml/common/linalg/BLAS.java

+	// For level-1 routines, we use Java implementation.
 	private static final com.github.fommil.netlib.BLAS NATIVE_BLAS = com.github.fommil.netlib.BLAS.getInstance();
+
+	// For level-2 and level-3 routines, we use the native BLAS.


nit: use /* */

Yes, please refine it when merging. Thanks.

xuyang1706 · 2019-10-25T02:16:43Z

Thanks for the quick reply/update @xuyang1706 . I think the patch looks good overall.
I will made some minor refinement and merge it soon.

Thanks for your kindly help @walterddr.

walterddr · 2019-10-25T22:40:13Z

@flinkbot run travis

[FLINK-14153][ml] Add to BLAS a method that performs DenseMatrix and …

6b34d9e

…SparseVector multiplication.

rmetzger added the review=description? label Sep 20, 2019

rmetzger added the component=Library/MachineLearning label Sep 20, 2019

walterddr self-assigned this Oct 17, 2019

walterddr reviewed Oct 18, 2019

View reviewed changes

[hotfix] add more comments and test cases

dbf3b11

walterddr reviewed Oct 20, 2019

View reviewed changes

[hotfix] add more comments in BLAS

74c52fb

walterddr approved these changes Oct 23, 2019

View reviewed changes

[hotfix] add more comments

c247440

xuyang1706 mentioned this pull request Oct 24, 2019

[FLINK-14154][ml] Add the class for multivariate Gaussian Distribution. #9733

Closed

walterddr approved these changes Oct 24, 2019

View reviewed changes

walterddr closed this in 23a34b8 Oct 25, 2019

		double value = alpha * x.values[i];
		F2J_BLAS.daxpy(m, value, matA.data, index * m, 1, y.data, 0, 1);

[FLINK-14153][ml] Add to BLAS a method that performs DenseMatrix and SparseVector multiplication. #9732

[FLINK-14153][ml] Add to BLAS a method that performs DenseMatrix and SparseVector multiplication. #9732

Uh oh!

Conversation

xuyang1706 commented Sep 20, 2019

What is the purpose of the change

Brief change log

Verifying this change

Does this pull request potentially affect one of the following parts:

Documentation

Uh oh!

flinkbot commented Sep 20, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Automated Checks

Review Progress

Uh oh!

flinkbot commented Sep 20, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

CI report:

Uh oh!

walterddr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

walterddr Oct 18, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

xuyang1706 commented Oct 18, 2019

Uh oh!

walterddr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

xuyang1706 commented Oct 21, 2019

Uh oh!

walterddr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

xuyang1706 commented Oct 24, 2019

Uh oh!

walterddr commented Oct 24, 2019

Uh oh!

walterddr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

flinkbot commented Sep 20, 2019 •

edited

Loading

flinkbot commented Sep 20, 2019 •

edited

Loading

walterddr Oct 18, 2019 •

edited

Loading