[
https://issues.apache.org/jira/browse/IMPALA-11204?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17517322#comment-17517322
]
ASF subversion and git services commented on IMPALA-11204:
----------------------------------------------------------
Commit 85ddd27b640bf42a61d8af238938e16618db537e in impala's branch
refs/heads/master from stiga-huang
[ https://gitbox.apache.org/repos/asf?p=impala.git;h=85ddd27b6 ]
IMPALA-11204: Template implementation for OrcStringColumnReader::ReadValue
There are some checks in OrcStringColumnReader::ReadValue() that we can
determine outside the scope of this method. They should be optimized
since this is a critical method that will be executed for each row (and
for each string column). With these checks, the method is too complex to
be inlined in OrcBatchedReader::ReadValueBatch() by the compiler.
This patch templates OrcStringColumnReader::ReadValue() with two
parameters, one for the target slot type (i.e. STRING/CHAR/VARCHAR),
and the other one for whether the column is dictionary encoded. Also
adds an ALWAYS_INLINE marker to force inlining it.
OrcStringColumnReader::ReadValueBatch() will call a template version of
ReadValue() based on the slot type and the orc batch encoded state.
Ran a single node perf test on TPCH(30) on my dev box using 3 impalad
instances. There are some improvements and no significant regressions:
+----------+--------+-------------+------------+
| Query | Avg(s) | Base Avg(s) | Delta(Avg) |
+----------+--------+-------------+------------+
| TPCH-Q19 | 5.62 | 6.07 | I -7.41% |
| TPCH-Q6 | 2.56 | 2.78 | I -7.77% |
| TPCH-Q4 | 3.85 | 4.25 | I -9.42% |
| TPCH-Q12 | 4.25 | 4.99 | I -14.78% |
+----------+--------+-------------+------------+
Base commit: ff21728
File Format: orc/snap/block
Iterations: 30
Change-Id: I5e5f88c28059fb3d3ac1172e6d383d06ee3bedd5
Reviewed-on: http://gerrit.cloudera.org:8080/18366
Reviewed-by: Impala Public Jenkins <impala-public-jenk...@cloudera.com>
Tested-by: Impala Public Jenkins <impala-public-jenk...@cloudera.com>
> OrcStringColumnReader should be a template class
> ------------------------------------------------
>
> Key: IMPALA-11204
> URL: https://issues.apache.org/jira/browse/IMPALA-11204
> Project: IMPALA
> Issue Type: Improvement
> Components: Backend
> Reporter: Quanlong Huang
> Assignee: Quanlong Huang
> Priority: Critical
> Attachments: perf-report-tpch30-orc-snap.txt
>
>
> There are some duplicated checks in OrcStringColumnReader::ReadValue() which
> is executed for each row. They are static and should only be performed once
> per ORC stripe.
> {code:cpp}
> Status OrcStringColumnReader::ReadValue(int row_idx, Tuple* tuple, MemPool*
> pool) {
> if (IsNull(DCHECK_NOTNULL(batch_), row_idx)) {
> SetNullSlot(tuple);
> return Status::OK();
> }
> char* src_ptr;
> int src_len;
> if (batch_->isEncoded) {
> orc::EncodedStringVectorBatch* currentBatch =
> static_cast<orc::EncodedStringVectorBatch*>(batch_);
> orc::DataBuffer<int64_t>& offsets =
> currentBatch->dictionary->dictionaryOffset;
> int64_t index = currentBatch->index[row_idx];
> if (UNLIKELY(index < 0 || static_cast<uint64_t>(index) + 1 >=
> offsets.size())) {
> return Status(Substitute("Corrupt ORC file: $0. Index ($1) out of range
> [0, $2) in "
> "StringDictionary.", scanner_->filename(), index, offsets.size()));;
> }
> src_ptr = blob_ + offsets[index];
> src_len = offsets[index + 1] - offsets[index];
> } else {
> // The pointed data is now in blob_, a buffer handled by Impala.
> src_ptr = blob_ + (batch_->data[row_idx] - batch_->blob.data());
> src_len = batch_->length[row_idx];
> }
> int dst_len = slot_desc_->type().len;
> if (slot_desc_->type().type == TYPE_CHAR) {
> int unpadded_len = min(dst_len, static_cast<int>(src_len));
> char* dst_char = reinterpret_cast<char*>(GetSlot(tuple));
> memcpy(dst_char, src_ptr, unpadded_len);
> StringValue::PadWithSpaces(dst_char, dst_len, unpadded_len);
> return Status::OK();
> }
> StringValue* dst = reinterpret_cast<StringValue*>(GetSlot(tuple));
> if (slot_desc_->type().type == TYPE_VARCHAR && src_len > dst_len) {
> dst->len = dst_len;
> } else {
> dst->len = src_len;
> }
> dst->ptr = src_ptr;
> return Status::OK();
> }
> {code}
> This method is executed for each row. The complexity of it causes the
> compiler unable to inline it. I can see this method takes some portion of the
> time in TPC-H queries (e.g. 21% in Q12).
> Actually, the slot type is determined when we are creating the reader.
> Whether the orc vector batch is encoded is determined when we start reading a
> new stripe. If the string column is in dictionary encodings in the stripe and
> we set EnableLazyDecoding to true in RowReaderOptions, the orc vector batch
> will be encoded.
> [https://github.com/apache/orc/blob/21259815ae665f6a4beac29edb4fab52c2403e13/c%2B%2B/src/Reader.cc#L1205]
> [https://github.com/apache/orc/blob/21259815ae665f6a4beac29edb4fab52c2403e13/c%2B%2B/src/ColumnReader.cc#L1843-L1847]
> We can switch to a template implementation so the compile can remove these
> checks and inline the method.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-all-unsubscr...@impala.apache.org
For additional commands, e-mail: issues-all-h...@impala.apache.org
