Method and apparatus for supporting multiple outstanding vector requests in vector database

1 . A method of processing multiple outstanding vector requests in a vector search system, the method comprising:

transmitting, by a vector index host apparatus for executing a vector index host, information on a plurality of vector queries to a vector search hardware apparatus for executing a vector search upon receiving the plurality of vector queries;

creating, by the vector index host apparatus, a candidate node list for each of the plurality of vector queries;

confirming, by the vector index host apparatus, a multiple-search counting number indicating a total number of search tasks being processed or waiting in the vector search hardware apparatus to identify a number of requests that can be additionally processed by the vector search hardware apparatus;

selecting by the vector index host apparatus, search target nodes from the candidate node list;

buffering, by the vector index host apparatus, a number of search requests including search requests for neighbor nodes of the search target nodes;

transmitting, by the vector index host apparatus, the search requests for the neighbor nodes of the search target nodes to the vector search hardware apparatus by referring to the multiple-search counting number without waiting for a search response to a previous search request from the vector search hardware apparatus to reduce an idle time of the vector search hardware apparatus; and

transmitting, by the vector search hardware apparatus, search results for the search requests to the vector index host apparatus.

2 . The method of claim 1 , wherein the transmitting of the search requests includes transmitting, by the vector index host apparatus, a first search request for searching for similarity to an arbitrary query to neighbor nodes of a first node of a candidate node list for searching for the arbitrary query, and a second search request for searching for similarity to the arbitrary query to neighbor nodes of a second node of the candidate node list.

3 . The method of claim 1 , further comprising scheduling, by the vector search hardware apparatus, a search task according to each of the search requests.

4 . The method of claim 1 , wherein the transmitting of the search requests includes transmitting, by the vector index host apparatus, a third search request for searching for similarity to a first query to neighbor nodes of an arbitrary node of a candidate node list for searching for the first query, and a fourth search request for searching for similarity to a second query to neighbor nodes of an arbitrary node of a candidate node list for searching for the second query.

5 . The method of claim 4 , further comprising scheduling, by the vector search hardware apparatus, a search task according to the third search request or the fourth search request.

6 . A vector search system for supporting multiple outstanding vector requests, the system comprising:

a vector index host apparatus for executing a vector index host, wherein the vector index host apparatus comprises a memory storing instructions and a processor coupled to the memory and configured to execute the instructions to:

transmit information on a plurality of vector queries to a vector search hardware apparatus for executing a vector search upon receiving the plurality of vector queries,

create a candidate node list for each of the plurality of vector queries,

confirm a multiple-search counting number indicating a total number of search tasks being processed or waiting in the vector search hardware apparatus to identify a number of search requests that can be additionally processed by the vector search hardware apparatus,

select search target nodes from the candidate node list,

buffer a number of search requests including search requests for neighbor nodes of the search target nodes, and

transmit the search requests for the neighbor nodes of the search target nodes to the vector search hardware apparatus by referring to the multiple-search counting number without waiting for a search response to a previous search request from the vector search hardware apparatus to reduce an idle time of the vector search hardware apparatus; and

the vector search hardware apparatus that transmits search results for the search requests to the vector index host apparatus.

7 . An apparatus for executing a vector index host supporting multiple outstanding vector requests, the apparatus comprising:

a communication unit;

a memory storing instructions; and

processor coupled to the memory and configured to execute the instructions to:

upon receiving a plurality of queries, transmit information on a plurality of vector queries to a vector search hardware apparatus for executing a vector search through the communication unit,

create a candidate node list for each of the plurality of vector queries,

confirm a multiple-search counting number indicating a total number of search tasks being processed or waiting in the vector search hardware apparatus to identify a number of requests that can be additionally processed by the vector search hardware apparatus,

select search target nodes from the candidate node list,

buffer a number of search requests including search requests for neighbor nodes of the search target nodes,

transmit the search requests for the neighbor nodes of the search target nodes to the vector search hardware apparatus though the communication unit by referring to the multiple-search counting number without waiting for a search response to a previous search request from the vector search hardware apparatus to reduce an idle time of the vector search hardware apparatus, and

receive search results for the search requests from the vector search hardware apparatus through the communication unit.

8 . A computer-readable recording non-transitory medium in which a computer program executed by a computing device is recorded, the computer program comprising:

upon receiving a plurality of queries, transmitting information on a plurality of vector queries to a vector search hardware apparatus for executing a vector search;

creating a candidate node list for searching for each of the plurality of queries;

confirming a multiple-search counting number indicating a total number of search tasks being processed or waiting in the vector search hardware apparatus to identify a number of requests that can be additionally processed by the vector search hardware apparatus;

selecting search target nodes from the candidate node list;

buffering a number of search requests including search requests for neighbor nodes of the search target nodes;

transmitting the search requests for the neighbor nodes of the search target nodes to the vector search hardware apparatus by referring to the multiple-search counting number without waiting for a search response to a previous search request from the vector search hardware apparatus to reduce an idle time of the vector search hardware apparatus; and

receiving search results for the search requests from the vector search hardware apparatus.