From design doc:

We will add a genny test which will do multiple parallel sharded transactions.  Parameters of the test should include

• Number of transactions to run in parallel
• Number of operations within a given transaction
• Size of operations within a transaction

This test will run on an unsharded replica set (to set a baseline) and a sharded cluster, on a collection which has shard keys well-distributed between shards (e.g. hashed by _id), with majority write concern.  The relevant metrics are

• Total transaction throughput.
• Transaction latencies
• If we can get it, secondary lag during the test.

This could be either a sys-perf suite or c++ benchmark test.