The TestSmallDocBulkInsert task of the driver benchmarks reveals an inefficiency: as you call mongoc_bulk_operation_insert all docs are appended to the same buffer. In mongoc_bulk_operation_execute, those catenated documents are then split into batches of 1000, the batches have to be copied into "insert" command documents before they're sent.

The simplest fix is to pre-split batches: when the current batch has 1000 documents, mongoc_bulk_operation_insert should start a new one.

Same for updates and deletes.

A bulk op can be constructed before the driver has selected a server to send it to, or before the driver has connected at all, so the server's maxWriteBatchSize isn't known. Assume 1000 for now.