The fetch phase time for queries is relatively slow

Cluster mode, 3 nodes, each node can play both master and data roles, and the server configuration for each node is 16 cores and 64G. NumberOfShards=3, numberOfReplicas=1. The data volume for each shard is approximately 3G
The total number of documents queried is approximately 6KW, and the type of query field is text. When from and size are 0 and 10, the query is within ms; When the from and size are 0 and 1000, the query occasionally appears for more than 10 seconds. After analyzing the profile, it was found that the bottleneck mainly occurs on load_stored_fields. Why is there such a big difference in query time? Through load_stored_fileds_cout, it was found that there are only 3000, not particularly many. Why is time so slow

@csuchunpeng Welcome to the Elastic community.

May i know how many fields you have stored per doc?

About 6 fields per doc.
For the same query conditions, slow queries mainly occur during the first time.
If you enter the same query criteria again, the subsequent queries will be within milliseconds, which should be due to caching.
Sorry, I currently do not have permission to upload images。
Query criteria: A boost must combination query of two fields, with the first field being a match query and the second field being a terms query
Profile result information:
Node1: load_stored_fields: 98ms, load_stored_fields count: 3084;
Node2: load_stored_fields: 98ms, load_stored_fields count: 3479;
Node3: load_stored_fields: 8.1s, load_stored_fields count: 3437

The following are images of the query criteria:

the result of profile in node1:

the result of profile in node2:

the result of profile in node3: