In SERVER-83233, we added a sys-perf variant to simulate an invisible sharding cluster (a single shard cluster with an embedded config server and colocated mongos processes to simulate mongos embedded in each mongod). For CPU bound workloads, like YCSB 100 read and 95read/5update with less data than the WT cache, max throughput is significantly lower than for the single replica set sys-perf variant. This seems to be because CPU is maxed on the shard primary faster because of the mongos on it. This patch saw ~45% higher max throughput for those YCSB workloads by taking the mongos on the primary out of the connection string, supporting this idea.

Users with a dedicated mongos tier can scale them up enough such that the primary mongod is still the bottleneck for CPU bound workloads. We should add a sys-perf variant with enough dedicated mongoses to test that scenario.