Github user paul-rogers commented on a diff in the pull request:
https://github.com/apache/drill/pull/717#discussion_r100708660
--- Diff:
exec/java-exec/src/main/java/org/apache/drill/exec/physical/impl/spill/RecordBatchSizer.java
---
@@ -0,0 +1,293 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+package org.apache.drill.exec.physical.impl.spill;
+
+import java.util.ArrayList;
+import java.util.List;
+
+import org.apache.drill.exec.expr.TypeHelper;
+import org.apache.drill.exec.memory.BaseAllocator;
+import org.apache.drill.exec.record.BatchSchema;
+import org.apache.drill.exec.record.MaterializedField;
+import org.apache.drill.exec.record.VectorAccessible;
+import org.apache.drill.exec.record.VectorWrapper;
+import org.apache.drill.exec.record.selection.SelectionVector2;
+import org.apache.drill.exec.vector.BaseDataValueVector;
+import org.apache.drill.exec.vector.FixedWidthVector;
+import org.apache.drill.exec.vector.NullableVarCharVector;
+import org.apache.drill.exec.vector.NullableVector;
+import org.apache.drill.exec.vector.ValueVector;
+import org.apache.drill.exec.vector.VarCharVector;
+
+import io.netty.buffer.DrillBuf;
+
+/**
+ * Given a record batch or vector container, determines the actual memory
+ * consumed by each column, the average row, and the entire record batch.
+ */
+
+public class RecordBatchSizer {
+ private static final org.slf4j.Logger logger =
org.slf4j.LoggerFactory.getLogger(RecordBatchSizer.class);
+
+ /**
+ * Column size information.
+ */
+ public static class ColumnSize {
+ public final MaterializedField metadata;
+
+ /**
+ * Assumed size from Drill metadata.
+ */
+ public int stdSize;
+ /**
+ * Actual memory consumed by all the vectors associated with this
column.
+ */
+ public int totalSize;
+ /**
+ * Actual average column width as determined from actual memory use.
This
+ * size is larger than the actual data size since this size includes
per-
+ * column overhead such as any unused vector space, etc.
+ */
+ public int estSize;
+
+ /**
+ * The size of the data vector backing the column. Useful for detecting
+ * cases of possible direct memory fragmentation.
+ */
+ public int dataVectorSize;
+ public int capacity;
+ public int density;
+ public int dataSize;
+
+ @SuppressWarnings("resource")
+ public ColumnSize(VectorWrapper<?> vw) {
+ metadata = vw.getField();
+ stdSize = TypeHelper.getSize(metadata.getType());
+
+ // Can't get size estimates if this is an empty batch.
+
+ ValueVector v = vw.getValueVector();
+ int rowCount = v.getAccessor().getValueCount();
+ if (rowCount == 0) {
+ return;
+ }
+ DrillBuf[] bufs = v.getBuffers(false);
+ for (DrillBuf buf : bufs) {
+ totalSize += buf.capacity();
+ }
+
+ // Capacity is the number of values that the vector could
+ // contain. This is useful only for fixed-length vectors.
+
+ capacity = v.getValueCapacity();
+
+ // Crude way to get the size of the buffer underlying simple
(scalar) values.
+ // Ignores maps, lists and other esoterica. Uses a crude way to
subtract out
+ // the null "bit" (really byte) buffer size for nullable vectors.
+
+ if (v instanceof BaseDataValueVector) {
+ dataVectorSize = totalSize;
+ if (v instanceof NullableVector) {
+ dataVectorSize -= bufs[0].getActualMemoryConsumed();
+ }
+ }
+
+ // Determine "density" the number of rows compared to potential
+ // capacity. Low-density batches occur at block boundaries, ends
+ // of files and so on. Low-density batches throw off our estimates
+ // for Varchar columns because we don't know the actual number of
+ // bytes consumed (that information is hidden behind the Varchar
+ // implementation where we can't get at it.)
+ //
+ // A better solution is to have each vector do this calc rather
+ // than trying to do it generically, but that increases the code
+ // change footprint and slows the commit process.
+
+ if (v instanceof FixedWidthVector) {
+ dataSize = stdSize * rowCount;
+ } else if ( v instanceof VarCharVector ) {
+ VarCharVector vv = (VarCharVector) v;
+ dataSize = vv.getOffsetVector().getAccessor().get(rowCount);
+ } else if ( v instanceof NullableVarCharVector ) {
+ NullableVarCharVector vv = (NullableVarCharVector) v;
+ dataSize =
vv.getValuesVector().getOffsetVector().getAccessor().get(rowCount);
+ } else {
+ dataSize = 0;
+ }
+ if (dataSize > 0) {
+ density = roundUp(dataSize * 100, dataVectorSize);
+ estSize = roundUp(dataSize, rowCount);
+ }
+ }
+
+ @Override
+ public String toString() {
+ StringBuilder buf = new StringBuilder();
+ buf.append(metadata.getName());
+ buf.append("(std col. size: ");
+ buf.append(stdSize);
+ buf.append(", actual col. size: ");
+ buf.append(estSize);
+ buf.append(", total size: ");
+ buf.append(totalSize);
+ buf.append(", vector size: ");
+ buf.append(dataVectorSize);
+ buf.append(", data size: ");
+ buf.append(dataSize);
+ buf.append(", row capacity: ");
+ buf.append(capacity);
+ buf.append(", density: ");
+ buf.append(density);
+ buf.append(")");
+ return buf.toString();
+ }
+ }
+
+ List<ColumnSize> columnSizes = new ArrayList<>();
+
+ /**
+ * Number of records (rows) in the batch.
+ */
+ private int rowCount;
+ /**
+ * Standard row width using Drill meta-data.
+ */
+ private int stdRowWidth;
+ /**
+ * Actual batch size summing all buffers used to store data
+ * for the batch.
+ */
+ private int totalBatchSize;
+ /**
+ * Actual row width computed by dividing total batch memory by the
+ * record count.
+ */
+ private int grossRowWidth;
+ /**
+ * Actual row width computed by summing columns. Use this if the
+ * vectors are partially full; prevents overestimating row width.
+ */
+ private int netRowWidth;
+ private boolean hasSv2;
+ private int sv2Size;
+ private int avgDensity;
+
+ public RecordBatchSizer(VectorAccessible va) {
+ rowCount = va.getRecordCount();
+ for (VectorWrapper<?> vw : va) {
+ measureField(vw);
+ }
+
+ if (rowCount > 0) {
+ grossRowWidth = roundUp(totalBatchSize, rowCount);
+ }
+
+ hasSv2 = va.getSchema().getSelectionVectorMode() ==
BatchSchema.SelectionVectorMode.TWO_BYTE;
+ if (hasSv2) {
+ @SuppressWarnings("resource")
+ SelectionVector2 sv2 = va.getSelectionVector2();
+ sv2Size = sv2.getBuffer().capacity();
+ grossRowWidth += sv2Size;
+ netRowWidth += 2;
+ }
+
+ int totalDensity = 0;
+ int usableCount = 0;
+ for (ColumnSize colSize : columnSizes) {
+ if ( colSize.density > 0 ) {
+ usableCount++;
+ }
+ totalDensity += colSize.density;
+ }
+ avgDensity = roundUp(totalDensity, usableCount);
+ }
+
+ public void applySv2() {
+ if (hasSv2) {
+ return;
+ }
+
+ sv2Size = BaseAllocator.nextPowerOfTwo(2 * rowCount);
+ grossRowWidth += roundUp(sv2Size, rowCount);
+ totalBatchSize += sv2Size;
+ }
+
+ private void measureField(VectorWrapper<?> vw) {
+ ColumnSize colSize = new ColumnSize(vw);
+ columnSizes.add(colSize);
+
+ stdRowWidth += colSize.stdSize;
+ totalBatchSize += colSize.totalSize;
+ netRowWidth += colSize.estSize;
+ }
+
+ public static int roundUp(int denom, int num) {
+ if(num == 0) {
+ return 0;
+ }
+ return (int) Math.ceil((double) denom / num);
+ }
+
+ public int rowCount() { return rowCount; }
+ public int stdRowWidth() { return stdRowWidth; }
+ public int grossRowWidth() { return grossRowWidth; }
+ public int netRowWidth() { return netRowWidth; }
+ public int actualSize() { return totalBatchSize; }
+ public boolean hasSv2() { return hasSv2; }
+ public int getAvgDensity() { return avgDensity; }
+
+ public static final int MAX_VECTOR_SIZE = 16 * 1024 * 1024; // 16 MiB
+
+ /**
+ * Look for columns backed by vectors larger than the 16 MiB size
+ * employed by the Netty allocator. Such large blocks can lead to
+ * memory fragmentation and unexpected OOM errors.
+ * @return if any column is oversize
+ */
+ public boolean checkOversizeCols() {
+ boolean hasOversize = false;
+ for (ColumnSize colSize : columnSizes) {
+ if ( colSize.dataVectorSize > MAX_VECTOR_SIZE) {
+ logger.warn( "Column is wider than 256 characters: OOM due to
memory fragmentation is possible - " + colSize.metadata.getPath() );
+ hasOversize = true;
--- End diff --
The idea is to report all the oversized columns, but only do so once (not
on every incoming batch.) That is, you'll find out that columns b, e, and h are
oversized, but only once per sort.
This is temporary until the memory fragmentation issue is fixed.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
