Concurrency vs. Throughput: why more parallelism can make databases slower