4 Commits

Author	SHA1	Message	Date
Tim Armstrong	852e1bb728	IMPALA-3931: arbitrary fixed-size uda intermediate types ... Make many builtin aggregate functions use fixed-length intermediate types: * avg() * ndv() * stddev(), variance(), etc * distinctpc(), distinctpcsa() sample(), appx_median(), histogram() and group_concat() actually allocate var-len data so aren't changed. This has some major benefits: * Spill-to-disk works properly with these aggregations. * Aggregations are more efficient because there is one less pointer indirection. * Aggregations use less memory, because we don't need an extra 12-byte StringValue for the indirection. Adds a special-purpose internal type FIXED_UDA_INTERMEDIATE. The type is represented in the same way as CHAR - a fixed-size array of bytes, stored inline in tuples. However, it is not user-visible and does not support CHAR semantics, i.e. users can't declare tables, functions, etc with the type. The pointer and length is passed into aggregate functions wrapped in a StringVal. Updates some internal codegen functions to work better with the new type. E.g. store values directly into the result tuple instead of via an intermediate stack allocation. Testing: This change only affects builtin aggregate functions, for which we have test coverage already. If we were to allow wider use of this type, it would need further testing. Added an analyzer test to ensure we can't use the type for UDAs. Added a regression test for spilling avg(). Added a regression test for UDA with CHAR intermediate hitting DCHECK. Perf: Ran TPC-H locally. TPC-H Q17, which has a high-cardinality AVG(), improved dramatically. +----------+-----------------------+---------+------------+------------+----------------+ | Workload | File Format | Avg (s) | Delta(Avg) | GeoMean(s) | Delta(GeoMean) | +----------+-----------------------+---------+------------+------------+----------------+ | TPCH(60) | parquet / none / none | 18.44 | -17.54% | 11.92 | -5.34% | +----------+-----------------------+---------+------------+------------+----------------+ +----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------------+-------+ | Workload | Query | File Format | Avg(s) | Base Avg(s) | Delta(Avg) | StdDev(%) | Base StdDev(%) | Num Clients | Iters | +----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------------+-------+ | TPCH(60) | TPCH-Q12 | parquet / none / none | 18.40 | 17.64 | +4.32% | 0.77% | 1.09% | 1 | 5 | | TPCH(60) | TPCH-Q22 | parquet / none / none | 7.07 | 6.90 | +2.36% | 0.28% | 0.30% | 1 | 5 | | TPCH(60) | TPCH-Q3 | parquet / none / none | 12.37 | 12.11 | +2.10% | 0.18% | 0.15% | 1 | 5 | | TPCH(60) | TPCH-Q7 | parquet / none / none | 42.48 | 42.09 | +0.93% | 2.45% | 0.80% | 1 | 5 | | TPCH(60) | TPCH-Q6 | parquet / none / none | 3.18 | 3.15 | +0.89% | 0.67% | 0.76% | 1 | 5 | | TPCH(60) | TPCH-Q19 | parquet / none / none | 7.24 | 7.20 | +0.50% | 0.95% | 0.67% | 1 | 5 | | TPCH(60) | TPCH-Q10 | parquet / none / none | 13.37 | 13.30 | +0.50% | 0.48% | 1.39% | 1 | 5 | | TPCH(60) | TPCH-Q5 | parquet / none / none | 7.47 | 7.44 | +0.36% | 0.58% | 0.54% | 1 | 5 | | TPCH(60) | TPCH-Q11 | parquet / none / none | 2.03 | 2.02 | +0.06% | 0.26% | 1.95% | 1 | 5 | | TPCH(60) | TPCH-Q4 | parquet / none / none | 5.48 | 5.50 | -0.27% | 0.62% | 1.12% | 1 | 5 | | TPCH(60) | TPCH-Q13 | parquet / none / none | 22.11 | 22.18 | -0.31% | 0.18% | 0.55% | 1 | 5 | | TPCH(60) | TPCH-Q15 | parquet / none / none | 8.45 | 8.48 | -0.32% | 0.40% | 0.47% | 1 | 5 | | TPCH(60) | TPCH-Q9 | parquet / none / none | 33.39 | 33.66 | -0.81% | 0.75% | 0.59% | 1 | 5 | | TPCH(60) | TPCH-Q21 | parquet / none / none | 71.34 | 72.07 | -1.01% | 1.84% | 1.79% | 1 | 5 | | TPCH(60) | TPCH-Q14 | parquet / none / none | 5.93 | 6.00 | -1.07% | 0.15% | 0.69% | 1 | 5 | | TPCH(60) | TPCH-Q20 | parquet / none / none | 5.72 | 5.79 | -1.09% | 0.59% | 0.51% | 1 | 5 | | TPCH(60) | TPCH-Q18 | parquet / none / none | 45.42 | 45.93 | -1.10% | 1.42% | 0.50% | 1 | 5 | | TPCH(60) | TPCH-Q2 | parquet / none / none | 4.81 | 4.89 | -1.52% | 1.68% | 1.01% | 1 | 5 | | TPCH(60) | TPCH-Q16 | parquet / none / none | 5.41 | 5.52 | -1.98% | 0.66% | 0.73% | 1 | 5 | | TPCH(60) | TPCH-Q1 | parquet / none / none | 27.58 | 29.13 | -5.34% | 0.24% | 1.51% | 1 | 5 | | TPCH(60) | TPCH-Q8 | parquet / none / none | 12.61 | 14.30 | -11.78% | 6.20% | * 15.28% * | 1 | 5 | | TPCH(60) | TPCH-Q17 | parquet / none / none | 43.74 | 126.58 | I -65.44% | 1.34% | 9.60% | 1 | 5 | +----------+----------+-----------------------+--------+-------------+------------+-----------+----------------+-------------+-------+ Change-Id: Ife90cf27989f98ffb5ef5c39f1e09ce92e8cb87c Reviewed-on: http://gerrit.cloudera.org:8080/7526 Tested-by: Impala Public Jenkins Reviewed-by: Tim Armstrong <tarmstrong@cloudera.com>	2017-08-17 03:12:48 +00:00
Taras Bobrovytsky	529a5f99b9	IMPALA-4787: Optimize APPX_MEDIAN() memory usage ... Before this change, ReservoirSample functions (such as APPX_MEDIAN()) allocated memory for 20,000 elements up front per grouping key. This caused inefficient memory usage for aggregations with many grouping keys. This patch fixes this by initially allocating memory for 16 elements. Once the buffer becomes full, we reallocate a new buffer with double capacity and copy the original buffer into the new one. We continue doubling the buffer size until the buffer has room for 20,000 elements as before. Testing: Added some EE APPX_MEDIAN() tests on larger datasets that exercise the resize code path. Perf Benchrmark (about 35,000 elements per bucket): SELECT MAX(a) from ( SELECT c1, appx_median(c2) as a FROM benchmark GROUP BY c1) t BEFORE: 11s067ms Operator #Hosts Avg Time Max Time #Rows Est. #Rows Peak Mem Est. Peak Mem Detail ------------------------------------------------------------------------------------------------------------------------- 06:AGGREGATE 1 124.726us 124.726us 1 1 28.00 KB -1.00 B FINALIZE 05:EXCHANGE 1 29.544us 29.544us 3 1 0 -1.00 B UNPARTITIONED 02:AGGREGATE 3 86.406us 120.372us 3 1 44.00 KB 10.00 MB 04:AGGREGATE 3 1s840ms 2s824ms 2.00K -1 1.02 GB 128.00 MB FINALIZE 03:EXCHANGE 3 1s163ms 1s989ms 6.00K -1 0 0 HASH(c1) 01:AGGREGATE 3 3s356ms 3s416ms 6.00K -1 1.95 GB 128.00 MB STREAMING 00:SCAN HDFS 3 64.962ms 65.490ms 65.54M -1 25.97 MB 64.00 MB tpcds_10_parquet.benchmark AFTER: 9s465ms Operator #Hosts Avg Time Max Time #Rows Est. #Rows Peak Mem Est. Peak Mem Detail ------------------------------------------------------------------------------------------------------------------------ 06:AGGREGATE 1 73.961us 73.961us 1 1 28.00 KB -1.00 B FINALIZE 05:EXCHANGE 1 18.101us 18.101us 3 1 0 -1.00 B UNPARTITIONED 02:AGGREGATE 3 75.795us 83.969us 3 1 44.00 KB 10.00 MB 04:AGGREGATE 3 1s608ms 2s683ms 2.00K -1 1.02 GB 128.00 MB FINALIZE 03:EXCHANGE 3 826.683ms 1s322ms 6.00K -1 0 0 HASH(c1) 01:AGGREGATE 3 2s457ms 2s672ms 6.00K -1 3.14 GB 128.00 MB STREAMING 00:SCAN HDFS 3 81.514ms 89.056ms 65.54M -1 25.94 MB 64.00 MB tpcds_10_parquet.benchmark Memory Benchmark (about 12 elements per bucket): SELECT MAX(a) FROM ( SELECT ss_customer_sk, APPX_MEDIAN(ss_sold_date_sk) as a FROM tpcds_parquet.store_sales GROUP BY ss_customer_sk) t BEFORE: 7s477ms Operator #Hosts Avg Time Max Time #Rows Est. #Rows Peak Mem Est. Peak Mem Detail --------------------------------------------------------------------------------------------------------------------- 06:AGGREGATE 1 114.686us 114.686us 1 1 28.00 KB -1.00 B FINALIZE 05:EXCHANGE 1 18.214us 18.214us 3 1 0 -1.00 B UNPARTITIONED 02:AGGREGATE 3 147.055us 165.464us 3 1 28.00 KB 10.00 MB 04:AGGREGATE 3 2s043ms 2s147ms 14.82K -1 4.94 GB 128.00 MB FINALIZE 03:EXCHANGE 3 840.528ms 943.254ms 15.61K -1 0 0 HASH(ss_customer_sk) 01:AGGREGATE 3 1s769ms 1s869ms 15.61K -1 5.32 GB 128.00 MB STREAMING 00:SCAN HDFS 3 17.941ms 37.109ms 183.59K -1 1.94 MB 16.00 MB tpcds_parquet.store_sales AFTER: 434ms Operator #Hosts Avg Time Max Time #Rows Est. #Rows Peak Mem Est. Peak Mem Detail --------------------------------------------------------------------------------------------------------------------- 06:AGGREGATE 1 125.915us 125.915us 1 1 28.00 KB -1.00 B FINALIZE 05:EXCHANGE 1 72.179us 72.179us 3 1 0 -1.00 B UNPARTITIONED 02:AGGREGATE 3 79.054us 83.385us 3 1 28.00 KB 10.00 MB 04:AGGREGATE 3 6.559ms 7.669ms 14.82K -1 17.32 MB 128.00 MB FINALIZE 03:EXCHANGE 3 67.370us 85.068us 15.60K -1 0 0 HASH(ss_customer_sk) 01:AGGREGATE 3 19.245ms 24.472ms 15.60K -1 9.48 MB 128.00 MB STREAMING 00:SCAN HDFS 3 53.173ms 55.844ms 183.59K -1 1.18 MB 16.00 MB tpcds_parquet.store_sales Change-Id: I99adaad574d4fb0a3cf38c6cbad8b2a23df12968 Reviewed-on: http://gerrit.cloudera.org:8080/6025 Reviewed-by: Taras Bobrovytsky <tbobrovytsky@cloudera.com> Tested-by: Impala Public Jenkins	2017-03-16 05:59:40 +00:00
Michael Ho	37a73ae9f7	IMPALA-3350: Add some missing StringVal.is_null checks ... Impala has a hardcoded limit of 1GB in size for StringVal. If the length of the string exceeds 1GB, Impala will simply mark the StringVal as NULL (i.e. is_null = true). It's important that string functions or built-in UDFs check this field before accessing the pointer or Impala may end up doing null pointer access, leading to crashes. Change-Id: I55777487fff15a521818e39b4f93a8a242770ec2 Reviewed-on: http://gerrit.cloudera.org:8080/2786 Reviewed-by: Dan Hecht <dhecht@cloudera.com> Tested-by: Internal Jenkins	2016-05-12 14:17:39 -07:00
Michael Ho	e01ab4f1b2	IMPALA-2620: FunctionContext::Allocate() and friends should check for memory limits. ... FunctionContext::Allocate(), FunctionContextImpl::AllocateLocal() and FunctionContext::Reallocate() allocate memory without taking memory limits into account. The problem is that these functions invoke FreePool::Allocate() which may call MemPool::Allocate() that doesn't check against the memory limits. This patch fixes the problem by making these FunctionContext functions check for memory limits and set an error in the FunctionContext object if memory limits are exceeded. An alternative would be for these functions to call MemPool::TryAllocate() instead and return NULL if memory limits are exceeded. However, this may break some existing external UDAs which don't check for allocation failures, leading to unexpected crashes of Impala. Therefore, we stick with this ad hoc approach until the UDF/UDA interfaces are updated in the future releases. Callers of these FunctionContext functions are also updated to handle potential failed allocations instead of operating on NULL pointers. The query status will be polled at various locations and terminate the query. This patch also fixes MemPool to handle the case in which malloc may return NULL. It propagates the failure to the callers instead of continuing to run with NULL pointers. In addition, errors during aggregate functions' initialization are now properly propagated. Change-Id: Icefda795cd685e5d0d8a518cbadd37f02ea5e733 Reviewed-on: http://gerrit.cloudera.org:8080/1445 Reviewed-by: Michael Ho <kwho@cloudera.com> Tested-by: Internal Jenkins	2015-12-19 04:45:55 +00:00