mongoinserts

The Java MongoDB driver (here at ver: 3.2.0) exposes several flavours by which documents can be inserted. mongoinserts.Inserts exercises some of them while recording TPS information.

The data

The data used here is the primer-dataset (copy checked in) . It contains 25359 restaurant listings in JSON format

The flavours

• insertOne maps to MongoCollection.insertOne
• insertOnePar a concurrent version of insertOne using Java 8 parallel stream
• insertMany maps to MongoCollection.insertMany
• insertManyPar a concurrent version of insertMany using Java 8 parallel stream with custom batching of the input documents
• bulkWriteOrdered maps to MongoCollection.bulkWrite with BulkWriteOptions().ordered(true)
• bulkWriteOrderedPar a concurrent version of bulkWriteOrderedusing Java 8 parallel stream with custom batching of the input documents
• bulkWriteUnOrdered maps to MongoCollection.bulkWrite with BulkWriteOptions().ordered(false)
• bulkWriteUnOrderedPar a concurrent version of bulkWriteUnOrderedusing Java 8 parallel stream with custom batching of the input documents

Methodology

Local instance of MongoDB, default concerns. Each flavour is run multiple times, before each run the collection is dropped. The duration of each run is recorded and the last N runs are averaged to produce a TPS number.

Numbers

The absolute values are not very useful as they depend on hardware to a very large degree (if you are really curious I get between 10k and 70k doc/sec depending on the flavour). Consequently all results here are normalized by the slowest result which is insertOne:

insertOne              1.0
insertOnePar           3.4

insertMany             2.6
insertManyPar          6.2

bulkWriteUnordered     2.5
bulkWriteUnorderedPar  5.7

bulkWriteOrdered       2.7
bulkWriteOrderedPar    6.0

Observations

1. In every case the concurrent version produces better throughput then the serial.
2. bulkWriteOrdered performs (slightly) better then bulkWriteUnOrdered. This remains true for the parallel versions.
3. insertMany is similar to the bulk operations and is better when done in parallel.