Mongo write taking too long with Pyspark (sharded cluster)

Question

I'm trying to read parquet files and dump it onto mongodb collection (sharded). When i do it without sharding, the write throughput is really good. But after sharding it has gone down drastically.

A single task is taking 30 mins plus, which is only processing 16 mb data

I'm using below Spark config

(
     SparkConf()
    .setMaster("yarn")
    .set("spark.executor.memory", "30g")
    .set("spark.executor.instances", "10")
    .set("spark.executor.cores", "5")
    .set("spark.sql.shuffle.partitions", "2000")
    .set("spark.network.timeout", "800")
    .set("spark.sql.broadcastTimeout", "1200")
    .set("spark.default.parallelism", "2000") 
    .set('spark.jars', './mongo*.jar')
    .set("spark.mongodb.input.uri", mongo_uri)
    .set("spark.mongodb.input.database", db)
    .set("spark.mongodb.input.collection", db_collection)
    .set("spark.mongodb.output.uri", mongo_uri)
    .set("spark.mongodb.output.database", db)
    .set("spark.mongodb.output.collection", db_collection)
    .set("spark.mongodb.input.partitionerOptions.partitionKey", shard_key)
    .set("spark.mongodb.input.partitioner", "MongoShardedPartitioner")
    .set("spark.mongodb.input.partitionerOptions.shardkey", shard_key) 
)

I'm looking to dump 20 Billion plus records, 8 hours up and it's only inserted around 800 million documents.

The documents are same size, each document is of 250 KB

No additional indexes are being used.