Hi hackers,
I am using pg_embedding extension for Postgres which implements HNSW
index (some kind of ANN search).
Search query looks something like this:
SELECT _id FROM documents ORDER BY openai <=> ARRAY[0.024466066,
-0.011110042, -0.0012917554,... , -0.008700027] LIMIT 1;
I do not put all query here because vector contains 1536 elements.
The query plans is straightforward:
Limit (cost=4007500.00..4007502.06 rows=1 width=33)
-> Index Scan using pgembedding_hnsw_idx on documents
(cost=4007500.00..6064544.00 rows=1000000 width=33)
Order By: (openai <=>
'{0.024466066,-0.011110042,-0.0012917554,-0.025349319,-0.017500998,0.015570462,-0.015835438,-0.0067442562,0.0015977388,-0.034...
Planning Time: 1.276 ms
JIT:
Functions: 5
Options: Inlining true, Optimization true, Expressions true,
Deforming true
Timing: Generation 0.289 ms, Inlining 3.752 ms, Optimization 34.502
ms, Emission 19.078 ms, Total 57.621 ms
Execution Time: 70.850 ms
without JIT:
Planning Time: 0.472 ms
Execution Time: 13.300 ms
The problem is that JIT slowdown execution of this query 4x times: with
default setting it is 70 msec, with jit=off - 13 msec.
I wonder why do we ever try to use JIT for index scan, especially for
user defined index where no function can be inlined?
Why not to restrict JIT just to seqscan plan? But the way in this
particular case enable_seqscan=off...
May be it is naive suggestions, but 5x slowdown doesn't seems like right
behavior in any case.
Profile of query execution is the following:
- 78.52% 0.00% postgres postgres [.]
standard_ExecutorRun
- 78.52% standard_ExecutorRun
- 55.38% 0000561dc27c89dc
- 55.38% 0000561dc27c89dc
0x7fc3a58dd731
- llvm_get_function
+ 37.66% LLVMRunPassManager
+ 14.99% LLVMOrcLLJITLookup
+ 0.91% llvm_inline
0.90% llvm::PassManagerBuilder::populateModulePassManager
+ 0.79% LLVMRunFunctionPassManager
- 22.79% 0000561dc27c89dc
- ExecScan
- 22.76% 0000561dc27c89dc
index_getnext_tid
- hnsw_gettuple
- 22.76% hnsw_search
- 22.72% searchKnn
+ 22.29% searchBaseLayer
Version of Postgres: Ubuntu 15.3-1.pgdg22.04+1)
