jprdonnelly/impala
1
0
Fork 0
mirror of https://github.com/apache/impala.git synced 2026-01-10 18:00:14 -05:00
Code Issues Projects Releases Wiki Activity
10,410 Commits 49 Branches 63 Tags
3962ae1972e9e2d6592cdd5e305c021cf38c41bd
Commit Graph

4 Commits

Author SHA1 Message Date
Tim Armstrong
d4648e87b4 IMPALA-4356,IMPALA-7331: codegen all ScalarExprs 
Based on initial draft patch by Pooja Nilangekar.

Codegen'd expressions can be executed in two ways - either by
being called directly from a fully codegend function, or from
interpreted code via a function pointer (previously
ScalarFnCall::scalar_fn_wrapper_).

This change moves the function pointer from ScalarFnCall to its
base class ScalarExpr, so the full expr tree can be codegen'd, not
just the ScalarFnCall subtrees. The key refactoring and improvements
are:
* ScalarExpr::Get*Val() switches between interpreted and the codegen'd
  function pointer code paths in an inline function, avoiding a
  virtual function call to ScalarFnCal::Get*Val().
* Boilerplate logic is moved to ScalarExpr::GetCodegendComputeFn(),
  which calls a virtual function GetCodegenComputeFnImpl().
* ScalarFnCall's logic for deciding whether to interpret or codegen is
  better abstracted and exposed to ScalarExpr as IsInterpretable()
  and ShouldCodegen() methods.
* The ScalarExpr::codegend_compute_fn_ function pointer is only
  populated for expressions that are "codegen entry points". These
  include the roots of expr trees and non-root expressions
  where the parent expression calls Get*Val() from the
  pseudo-codegend GetCodegendComputeFnWrapper().
* ScalarFnCall is always initialised for interpreted execution.
  Otherwise the function pointer is needed for non-root expressions,
  e.g. to support ScalarExprEvaluator::GetConstantVal().
* Latent bugs/gaps for codegen of CollectionVal are fixed. CollectionVal
  is modified to use the StringVal memory layout to allow code sharing
  with StringVal. These fixes allowed simplification of
  IsNotEmptyPredicate codegen (from IMPALA-7657).

I chose to tackle two problems in one change - adding support for
generating codegen'd function pointers for all ScalarExprs, and adding
the "entry point" concept - to avoid a blow-up in the number of
codegen'd entry points that could lead to longer codegen times and/or
worse code because of inlining changes.

IMPALA-7331 (CHAR codegen support functions) is also fixed because
it was simpler to enable CHAR codegen within ScalarExpr than to carry
forward the exiting CHAR workarounds from ScalarFnCall. The
CHAR-specific codegen support required in the scalar expr subsystem is
very limited.  StringVal intermediates are used everywhere. Only
SlotRef actually operates on the different tuple layout, and the
required codegen support for SlotRef already exists for UDA
intermediates anyway.

Testing:
* Ran exhaustive tests.

Perf:
* Ran a basic insert benchmark, which went from 10.1s to 7.6s
  create table foo stored as parquet as
  select case when l_orderkey % 2 = 0 then 'aaa' else 'bbb' end
  from tpch30_parquet.lineitem;
* Ran a basic CHAR expr test:
  set num_nodes=1;
  set mt_dop=1;
  select count(*) from lineitem
  where cast(l_linestatus as CHAR(2)) = 'O ' and
        cast(l_returnflag as CHAR(2)) = 'N '
  The time spent in the scan went from 520ms to 220ms.
* Added perf regression test to tpcds-insert, similar to the manual
  benchmark.
* Ran single-node TPC-H with large and small scale factors, to estimate
  impact on execution perf and query startup time, respectively.

+----------+-----------------------+---------+------------+------------+----------------+
| Workload | File Format           | Avg (s) | Delta(Avg) | GeoMean(s) | Delta(GeoMean) |
+----------+-----------------------+---------+------------+------------+----------------+
| TPCH(30) | parquet / none / none | 6.84    | -0.18%     | 4.49       | -0.31%         |
+----------+-----------------------+---------+------------+------------+----------------+

+----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------+----------------+---------+--------+
| Workload | Query    | File Format           | Avg(s) | Base Avg(s) | Delta(Avg) | StdDev(%) | Base StdDev(%) | Iters | Median Diff(%) | MW Zval | Tval   |
+----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------+----------------+---------+--------+
| TPCH(30) | TPCH-Q20 | parquet / none / none | 2.58   | 2.47        |   +4.18%   |   1.29%   |   0.88%        | 5     |   +4.12%       | 2.31    | 5.81   |
| TPCH(30) | TPCH-Q17 | parquet / none / none | 4.81   | 4.61        |   +4.33%   |   2.18%   |   2.15%        | 5     |   +3.91%       | 1.73    | 3.09   |
| TPCH(30) | TPCH-Q21 | parquet / none / none | 26.45  | 26.16       |   +1.09%   |   0.37%   |   0.50%        | 5     |   +1.36%       | 2.02    | 3.94   |
| TPCH(30) | TPCH-Q9  | parquet / none / none | 15.92  | 15.75       |   +1.09%   |   2.87%   |   1.65%        | 5     |   +0.88%       | 0.29    | 0.73   |
| TPCH(30) | TPCH-Q12 | parquet / none / none | 2.38   | 2.35        |   +1.12%   |   1.64%   |   1.11%        | 5     |   +0.80%       | 1.15    | 1.26   |
| TPCH(30) | TPCH-Q14 | parquet / none / none | 2.94   | 2.91        |   +1.13%   |   7.68%   |   5.37%        | 5     |   -0.34%       | -0.29   | 0.27   |
| TPCH(30) | TPCH-Q18 | parquet / none / none | 18.10  | 18.02       |   +0.42%   |   2.70%   |   0.56%        | 5     |   +0.28%       | 0.29    | 0.34   |
| TPCH(30) | TPCH-Q8  | parquet / none / none | 4.72   | 4.72        |   -0.04%   |   1.20%   |   1.65%        | 5     |   +0.05%       | 0.00    | -0.04  |
| TPCH(30) | TPCH-Q19 | parquet / none / none | 3.92   | 3.93        |   -0.26%   |   1.08%   |   2.36%        | 5     |   +0.20%       | 0.58    | -0.23  |
| TPCH(30) | TPCH-Q6  | parquet / none / none | 1.27   | 1.27        |   -0.28%   |   0.22%   |   0.88%        | 5     |   +0.09%       | 0.29    | -0.68  |
| TPCH(30) | TPCH-Q16 | parquet / none / none | 2.64   | 2.65        |   -0.45%   |   1.65%   |   0.65%        | 5     |   -0.24%       | -0.58   | -0.57  |
| TPCH(30) | TPCH-Q22 | parquet / none / none | 3.10   | 3.13        |   -0.76%   |   1.47%   |   1.12%        | 5     |   -0.21%       | -0.29   | -0.93  |
| TPCH(30) | TPCH-Q2  | parquet / none / none | 1.20   | 1.21        |   -0.80%   |   2.26%   |   2.47%        | 5     |   -0.82%       | -1.15   | -0.53  |
| TPCH(30) | TPCH-Q4  | parquet / none / none | 1.97   | 1.99        |   -1.37%   |   1.84%   |   3.21%        | 5     |   -0.47%       | -0.58   | -0.83  |
| TPCH(30) | TPCH-Q13 | parquet / none / none | 11.53  | 11.63       |   -0.91%   |   0.46%   |   0.49%        | 5     |   -0.95%       | -2.02   | -3.08  |
| TPCH(30) | TPCH-Q10 | parquet / none / none | 5.13   | 5.21        |   -1.51%   |   2.24%   |   4.05%        | 5     |   -0.94%       | -0.58   | -0.73  |
| TPCH(30) | TPCH-Q5  | parquet / none / none | 3.61   | 3.66        |   -1.40%   |   0.66%   |   0.79%        | 5     |   -1.33%       | -1.73   | -3.05  |
| TPCH(30) | TPCH-Q7  | parquet / none / none | 19.42  | 19.71       |   -1.52%   |   1.34%   |   1.39%        | 5     |   -1.22%       | -1.44   | -1.76  |
| TPCH(30) | TPCH-Q3  | parquet / none / none | 5.08   | 5.15        |   -1.49%   |   1.34%   |   0.73%        | 5     |   -1.35%       | -1.44   | -2.20  |
| TPCH(30) | TPCH-Q15 | parquet / none / none | 3.42   | 3.49        |   -1.92%   |   0.93%   |   1.47%        | 5     |   -1.53%       | -1.15   | -2.49  |
| TPCH(30) | TPCH-Q11 | parquet / none / none | 1.15   | 1.19        |   -3.17%   |   2.27%   |   1.95%        | 5     |   -4.21%       | -1.15   | -2.41  |
| TPCH(30) | TPCH-Q1  | parquet / none / none | 9.26   | 9.63        |   -3.85%   |   0.62%   |   0.59%        | 5     |   -3.78%       | -2.31   | -10.25 |
+----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------+----------------+---------+--------+

Cluster Name: UNKNOWN
Lab Run Info: UNKNOWN
Impala Version:          impalad version 3.2.0-SNAPSHOT RELEASE ()
Baseline Impala Version: impalad version 3.2.0-SNAPSHOT RELEASE (2019-03-19)

+----------+-----------------------+---------+------------+------------+----------------+
| Workload | File Format           | Avg (s) | Delta(Avg) | GeoMean(s) | Delta(GeoMean) |
+----------+-----------------------+---------+------------+------------+----------------+
| TPCH(2)  | parquet / none / none | 0.90    | -0.08%     | 0.80       | -0.05%         |
+----------+-----------------------+---------+------------+------------+----------------+

+----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------+----------------+---------+-------+
| Workload | Query    | File Format           | Avg(s) | Base Avg(s) | Delta(Avg) | StdDev(%) | Base StdDev(%) | Iters | Median Diff(%) | MW Zval | Tval  |
+----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------+----------------+---------+-------+
| TPCH(2)  | TPCH-Q18 | parquet / none / none | 1.22   | 1.19        |   +1.93%   |   3.81%   |   4.46%        | 20    |   +3.34%       | 1.62    | 1.46  |
| TPCH(2)  | TPCH-Q10 | parquet / none / none | 0.74   | 0.73        |   +1.97%   |   3.36%   |   2.94%        | 20    |   +0.97%       | 1.88    | 1.95  |
| TPCH(2)  | TPCH-Q11 | parquet / none / none | 0.49   | 0.48        |   +1.91%   |   6.19%   |   4.64%        | 20    |   +0.25%       | 0.95    | 1.09  |
| TPCH(2)  | TPCH-Q4  | parquet / none / none | 0.43   | 0.43        |   +1.99%   |   6.26%   |   5.86%        | 20    |   +0.15%       | 0.92    | 1.03  |
| TPCH(2)  | TPCH-Q15 | parquet / none / none | 0.50   | 0.49        |   +1.82%   |   7.32%   |   6.35%        | 20    |   +0.26%       | 1.01    | 0.83  |
| TPCH(2)  | TPCH-Q1  | parquet / none / none | 0.98   | 0.97        |   +0.79%   |   4.64%   |   2.73%        | 20    |   +0.36%       | 0.77    | 0.65  |
| TPCH(2)  | TPCH-Q19 | parquet / none / none | 0.83   | 0.83        |   +0.65%   |   3.33%   |   2.80%        | 20    |   +0.44%       | 2.18    | 0.67  |
| TPCH(2)  | TPCH-Q14 | parquet / none / none | 0.62   | 0.62        |   +0.97%   |   2.86%   |   1.00%        | 20    |   +0.04%       | 0.13    | 1.42  |
| TPCH(2)  | TPCH-Q3  | parquet / none / none | 0.88   | 0.87        |   +0.57%   |   2.17%   |   1.74%        | 20    |   +0.29%       | 1.15    | 0.92  |
| TPCH(2)  | TPCH-Q12 | parquet / none / none | 0.53   | 0.53        |   +0.27%   |   4.58%   |   5.78%        | 20    |   +0.46%       | 1.47    | 0.16  |
| TPCH(2)  | TPCH-Q17 | parquet / none / none | 0.72   | 0.72        |   +0.15%   |   3.64%   |   5.55%        | 20    |   +0.21%       | 0.86    | 0.10  |
| TPCH(2)  | TPCH-Q21 | parquet / none / none | 2.05   | 2.05        |   +0.21%   |   1.99%   |   2.37%        | 20    |   +0.01%       | 0.25    | 0.30  |
| TPCH(2)  | TPCH-Q5  | parquet / none / none | 1.28   | 1.27        |   +0.24%   |   1.61%   |   1.80%        | 20    |   -0.02%       | -0.57   | 0.44  |
| TPCH(2)  | TPCH-Q13 | parquet / none / none | 1.27   | 1.27        |   -0.34%   |   1.69%   |   1.83%        | 20    |   -0.20%       | -1.65   | -0.61 |
| TPCH(2)  | TPCH-Q7  | parquet / none / none | 1.72   | 1.73        |   -0.55%   |   2.40%   |   1.69%        | 20    |   -0.03%       | -0.42   | -0.83 |
| TPCH(2)  | TPCH-Q8  | parquet / none / none | 1.27   | 1.28        |   -0.68%   |   3.10%   |   3.89%        | 20    |   -0.06%       | -0.54   | -0.62 |
| TPCH(2)  | TPCH-Q6  | parquet / none / none | 0.36   | 0.36        |   -0.84%   |   0.79%   |   3.51%        | 20    |   -0.07%       | -0.36   | -1.04 |
| TPCH(2)  | TPCH-Q2  | parquet / none / none | 0.65   | 0.65        |   -1.17%   |   4.76%   |   5.99%        | 20    |   -0.05%       | -0.25   | -0.69 |
| TPCH(2)  | TPCH-Q9  | parquet / none / none | 1.59   | 1.62        |   -2.01%   |   1.45%   |   5.12%        | 20    |   -0.16%       | -1.24   | -1.69 |
| TPCH(2)  | TPCH-Q20 | parquet / none / none | 0.68   | 0.69        |   -1.73%   |   4.35%   |   4.43%        | 20    |   -0.49%       | -1.74   | -1.25 |
| TPCH(2)  | TPCH-Q22 | parquet / none / none | 0.38   | 0.40        |   -2.89%   |   7.42%   |   6.39%        | 20    |   -0.21%       | -0.66   | -1.34 |
| TPCH(2)  | TPCH-Q16 | parquet / none / none | 0.59   | 0.62        |   -4.01%   |   6.33%   |   5.83%        | 20    |   -4.72%       | -1.39   | -2.13 |
+----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------+----------------+---------+-------+

Change-Id: I839d7a3a2f5e1309c33a1f66013ef11628c5dc11
Reviewed-on: http://gerrit.cloudera.org:8080/12797
Reviewed-by: Impala Public Jenkins <impala-public-jenkins@cloudera.com>
Tested-by: Impala Public Jenkins <impala-public-jenkins@cloudera.com>
 2019-05-15 22:34:28 +00:00
Lars Volker
a64cfc523e IMPALA-7032: Disable codegen for CHAR type null literals 
Analogous to IMPALA-6435, we have to disable codegen for CHAR type null
literals. Otherwise we will crash in
impala::NullLiteral::GetCodegendComputeFn().

This change adds a test to make sure that the crash is fixed.

Change-Id: I34033362263cf1292418f69c5ca1a3b84aed39a9
Reviewed-on: http://gerrit.cloudera.org:8080/10409
Reviewed-by: Lars Volker <lv@cloudera.com>
Tested-by: Lars Volker <lv@cloudera.com>
 2018-05-16 00:00:15 +00:00
aphadke
11d1784c0a IMPALA-6435: Disable codegen for CHAR literals. 
Currently we do not codegen CHAR types. This change checks
for CHAR literals in a expr and disables codegen.

Change-Id: I7e4e27350c53bc69ce412a004e392e7480214f73
Reviewed-on: http://gerrit.cloudera.org:8080/9102
Reviewed-by: Tim Armstrong <tarmstrong@cloudera.com>
Tested-by: Impala Public Jenkins
 2018-01-24 05:00:33 +00:00
Tim Armstrong
c4d284f3cc IMPALA-5483: Automatically disable codegen for small queries 
This is similar to the single-node execution optimisation, but applies
to slightly larger queries that should run in a distributed manner but
won't benefit from codegen.

This adds a new query option disable_codegen_rows_threshold that
defaults to 50,000. If fewer than this number of rows are processed
by a plan node per impalad, the cost of codegen almost certainly
outweighs the benefit.

Using rows processed as a threshold is justified by a simple
model that assumes the cost of codegen and execution per row for
the same operation are proportional. E.g. if x is the complexity
of the operation, n is the number of rows processed, C is a
constant factor giving the cost of codegen and Ec/Ei are constant
factor giving the cost of codegen'd and interpreted execution and
d, then the cost of the codegen'd operator is C * x + Ec * x * n
and the cost of the interpreted operator is Ei * x * n. Rearranging
means that interpretation is cheaper if n < C / (Ei - Ec), i.e. that
(at least with the simplified model) it makes sense to choose
interpretation or codegen based on a constant threshold. The
model also implies that it is somewhat safer to choose codegen
because the additional cost of codegen is O(1) but the additional
cost of interpretation is O(n).

I ran some experiments with TPC-H Q1, varying the input table size, to
determine what the cut-over point where codegen was beneficial was.
The cutover was around 150k rows per node for both text and parquet.
At 50k rows per node disabling codegen was very beneficial - around
0.12s versus 0.24s.  To be somewhat conservative I set the default
threshold to 50k rows. On more complex queries, e.g. TPC-H Q10, the
cutover tends to be higher because there are plan nodes that process
many fewer than the max rows.

Fix a couple of minor issues in the frontend - the numNodes_
calculation could return 0 for Kudu, and the single node optimization
didn't handle the case where for a scan node with conjuncts, a limit
and missing stats correctly (it considered the estimate still valid.)

Testing:
Updated e2e tests that set disable_codegen to set
disable_codegen_rows_threshold to 0, so that those tests run both
with and without codegen still.

Added an e2e test to make sure that the optimisation is applied in
the backend.

Added planner tests for various cases where codegen should and shouldn't
be disabled.

Perf:
Added a targeted perf test for a join+agg over a small input, which
benefits from this change.

Change-Id: I273bcee58641f5b97de52c0b2caab043c914b32e
Reviewed-on: http://gerrit.cloudera.org:8080/7153
Reviewed-by: Tim Armstrong <tarmstrong@cloudera.com>
Tested-by: Impala Public Jenkins
 2017-06-29 21:14:59 +00:00