Open longjiquan opened 3 weeks ago
https://github.com/milvus-io/milvus/blob/77f4f0c5a82f71c148a10e4719df48033dfd4d6c/internal/core/src/segcore/Reduce.cpp#L222-L311
If only one segment was involved in the reduce phase, in fact we waste the cpu of heap-sort to complete the reduce. We can do it more simply.
No response
the tracing can be found in https://github.com/milvus-io/milvus/pull/32734
seems to be a promising optimization
Is there an existing issue for this?
What would you like to be added?
https://github.com/milvus-io/milvus/blob/77f4f0c5a82f71c148a10e4719df48033dfd4d6c/internal/core/src/segcore/Reduce.cpp#L222-L311
Why is this needed?
If only one segment was involved in the reduce phase, in fact we waste the cpu of heap-sort to complete the reduce. We can do it more simply.
Anything else?
No response