[Bug]: It takes long time to build index for Float16Vector

[yhmo]

Is there an existing issue for this?

• [X] I have searched the existing issues

Environment

- Milvus version: v2.4.4
- Deployment mode(standalone or cluster):
- MQ type(rocksmq, pulsar or kafka):    
- SDK version(e.g. pymilvus v2.0.0rc2):
- OS(Ubuntu or CentOS): 
- CPU/Memory: 
- GPU: 
- Others:

Current Behavior

Index process for Float16Vector is so poor. One million vector 768 dim, HNSW{"M": 8, "efConstruction": 200}, it takes 1.5 hours to build index. Much slower than FloatVector.

Expected Behavior

The time cost to build index for Float16Vector should be less or equal to FloatVector.

Steps To Reproduce

1. create a collection with Float16Vector field
2. insert 1000000 vectors 768 dim into the collection
3. create index HNSW{"M": 8, "efConstruction": 200}
4. wait index done

Milvus Log

No response

Anything else?

No response

[yhmo]

Test script:

import random
import time

import numpy as np

from pymilvus import (
    connections,
    FieldSchema, CollectionSchema, DataType,
    Collection,
    utility,
)

HOST = 'localhost'
PORT = '19530'

connections.connect(host=HOST, port=PORT)

F16_COLLECTION = "f16_col"

DIM = 768
METRIC_TYPE = "L2"

ID_FIELD = "id"
VECTOR_FIELD = "vector"

def gen_fp16_vectors(num):
    raw_vectors = []
    fp16_vectors = []
    for _ in range(num):
        raw_vector = [random.random() for _ in range(DIM)]
        raw_vectors.append(raw_vector)
        fp16_vector = np.array(raw_vector, dtype=np.float16)
        fp16_vectors.append(fp16_vector)
    return raw_vectors, fp16_vectors

def create_collection():
    if utility.has_collection(F16_COLLECTION):
        utility.drop_collection(F16_COLLECTION)

    fields = [
        FieldSchema(name="id", dtype=DataType.INT64, is_primary=True, auto_id=True),
        FieldSchema(name="vector", dtype=DataType.FLOAT16_VECTOR, dim=DIM),
    ]
    schema = CollectionSchema(fields=fields)
    collection = Collection(name=F16_COLLECTION, schema=schema)
    print(f"Collection '{F16_COLLECTION}' created")

def prepare_data():
    collection_16 = Collection(name=F16_COLLECTION)
    count = 10000
    for i in range(100):
        raw_vectors, fp16_vectors = gen_fp16_vectors(count)
        collection_16.insert(data=[
            fp16_vectors,
        ])
        print(f"insert batch {i}")
    print("insert done")

    time.sleep(5)
    collection_16.flush()
    print("flush done")

    start = time.time()
    index_params = {
        'metric_type': METRIC_TYPE,
        'index_type': "HNSW",
        'params': {"M": 8, "efConstruction": 200},
    }
    collection_16.create_index(field_name=VECTOR_FIELD, index_params=index_params)
    utility.wait_for_index_building_complete(collection_name=F16_COLLECTION)
    print("index done")

    end = time.time()
    print(f"fp16 index time cost: {end-start} seconds")

if __name__ == '__main__':
    create_collection()
    prepare_data()

[yanliang567]

/assign @cqy123456 /unassign

[cqy123456]

/assign @yhmo use latest 2.4 to try, fp16 and bf16 add simd support in latest 2.4.

[yhmo]

Tested in milvus v2.4.9 with the same script, the time cost of fp16 index reduces to 446 seconds, ten times faster than v2.4.4

fp16 index time cost: 446.4446921348572 seconds