- there is not much overhead from the transaction API
- throughput improves as the work done per transaction increases
- throughput improves with the concurrent memtable
The test server has 2 sockets, 24 CPU cores and 48 HW threads. The database used fast storage (tmpfs). The benchmark is db_bench --benchmarks=randomtransaction via this test script. It was run was run for 1 to 48 concurrent clients to understand the impact of concurrency. It was run without transactions and with pessimistic transactions to understand the impact of the transaction API. It was run with --batch_size=1 and --batch_size=4 to understand the impact of doing more work per transaction. The test database was cached by RocksDB.
The first two graphs show the transaction throughput with (concurMT=1) and without (concurMT=0) the concurrent memtable enabled. Throughput is much larger with the concurrent memtable enabled. The benefit is larger with batch_size=4 than batch_size=1 because there is more work done with a larger batch_size and more mutex contention to avoid. Throughput is larger with batch_size=1 because there is 4X more work done per transaction with batch_size=4.
The next two graphs show the efficiency of the transaction API. It compares the ratio of the throughput with pessimistic transactions versus the throughput without transactions. When the value is 1.0 then the throughput with transactions matches the throughput without transactions. From the graphs below the efficiency is better with batch_size=1 than with batch_size=4 and the efficiency improves with concurrency.