Using preference _local in elasticsearch

I am currently tuning elasticsearch for search api. The specification are:

• 2 Nodes single cluster with single index on a VM (2 cores 2GB of RAM)
• 5 shards
• 1 replication
• load balanced using nginx

When I test it using Jmeter using Nginx it get throughput about ~220 req/s but when I specify the ?preference=_local it can get to ~320 req/s. That's very good performance improvement. What I want to ask is :

1. What ?preference=_local actually do and how it can improve the performance of the query?
2. What's the tradeoff of using ?preference=_local?

Query :

{
"query": {
    "multi_match": {
        "query": "trump",
        "type": "most_fields",
        "operator": "and",
        "fields": ["title", "content"]
    }
},
"sort": {
    "published_at": {
        "order": "desc"
    },
    "_score": {
        "order": "desc"
    }
},
"from": 0,
"size": 20,
"min_score": 1
}