[IMPALA-2726] Templatize RawValue::GetHashValue to avoid per row PrimitiveType check for 6% Perf improvement

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 2.2
• Fix Version/s: Impala 2.5.0
• Component/s: Backend
• Labels: performance ramp-up
• Target Version: Impala 2.5.0
• Assignee: Mostafa Mokhtar
• Reporter: Mostafa Mokhtar
• Created: 30/Nov/15 21:52
• Updated: 05/Feb/16 18:56
• Resolved: 05/Feb/16 18:56

Description

When looking at performance of ProbeFilters I realized that the type check for every row is consuming 7% of overall CPU.
We should use templates for

RawValue::GetHashValue()

to avoid all the expensive branches.

Function-Stack	Effective-Time-by-Utilization	Effective-Time-by-Utilization
----impala::Thread::SuperviseThread	97.20%	0s
------boost::function0<void>::operator()	97.20%	0s
--------impala::HdfsScanNode::ScannerThread	87.60%	0s
----------impala::HdfsParquetScanner::ProcessSplit	87.60%	0s
-----------impala::HdfsParquetScanner::AssembleRows<(bool)0,(bool)0>	87.60%	2.144s
--------------ReadRow<false>	75.10%	1.433s
---------------impala::HdfsParquetScanner::ScalarColumnReader<long,(bool)1>::ReadNonRepeatedValue	69.90%	1.188s
------------------ReadValue<false>	66.30%	0.210s
--------------------ReadSlot<false>	55.10%	1.930s
----------------------Get<true>	25.70%	8.539s
----------------------impala::DictDecoder<long>::GetValue	11.60%	1.510s
----------------------impala::HashUtil::Hash	6.20%	1.010s
----------------------impala::RawValue::GetHashValue	5.80%	1.918s

Perf data for

RawValue::GetHashValue

, sampling is based on cycles

impala::RawValue::GetHashValue(void const*, impala::ColumnType const&, unsigned int)
       │     Disassembly of section .text:                                                                                               
       │                                                                                                                                 
       │     00000000009ba270 <impala::RawValue::GetHashValue(void const*, impala::ColumnType const&, unsigned int)>:                    
  8.15 │       test   %rdi,%rdi                                                                                                          
       │     ↓ je     e0                                                                                                                 
  7.98 │       mov    (%rsi),%eax                                                                                                        
  1.78 │       cmp    $0x10,%eax                                                                                                         
       │     ↓ jbe    18                                                                                                                 
       │       xor    %eax,%eax                                                                                                          
       │     ← retq                                                                                                                      
       │       nop                                                                                                                       
  7.47 │ 18:   lea    typeinfo name for boost::detail::sp_counted_impl_p<impala::ClientCacheHelper::PerHostCache,%rcx                    
       │       mov    %eax,%r8d                                                                                                          
  7.08 │       movslq (%rcx,%r8,4),%r8                                                                                                   
 38.06 │       lea    (%r8,%rcx,1),%rcx                                                                                                  
 15.39 │     ↓ jmpq   ffffffffffffffff                                                                                                   
       │       nop                                                                                                                       
 13.81 │       mov    $0x8,%esi                                                                                                          
  0.29 │     ↓ jmpq   ffffffffffffffff                                                                                                   
       │       nop                                                                                                                       ◆
       │       mov    $0x4,%esi                                                                                                          
       │     ↓ jmpq   ffffffffffffffff                                                                                                   
       │       nop                                                                                                                       
       │       mov    0x8(%rdi),%esi                                                                                                     
       │       test   %esi,%esi                                                                                                          
       │     ↓ jne    130                                                                                                                
       │       mov    %edx,%ecx                                                                                                          
       │       mov    %edx,%eax                                                                                                          
       │       shl    $0x6,%ecx                                                                                                          
       │       shr    $0x2,%eax                                                                                                          
       │       lea    0x1c01f8a7(%rcx,%rax,1),%eax                                                                                       
       │       xor    %edx,%eax                                                                                                          
       │     ← retq                                                                                                                      
       │       nop                                                                                                                       
       │       mov    $0xc,%esi                                                                                                          
       │     ↓ jmpq   ffffffffffffffff                                                                                                   
       │       nop                                                                                                                       
       │       cmp    $0x10,%eax                                                                                                         
       │       mov    0x4(%rsi),%esi                                                                                                     
       │     ↓ je     100                                                                                                                
       │       cmp    $0xa,%eax                                                                                                          

Issue Links

• is related to: IMPALA-2715 Probe filter are not pushed to scan for joins on partitioned column
• is related to: IMPALA-2744 Codegen tuple copy for Hash Join , Agg and possibly Exchange
• is related to: IMPALA-2419 Extend Probe filter implementation