Cassandra scaling on AWS

After I successfully setup and run a Cassandra (v2.1) cluster I'm running some load tests to evaluate its perfomance using the YCSB library (http://ift.tt/1mNLvHH fork).

For my evaluation I use the same configuration on different sizes and number of EC2 nodes, however the performance is not really predictable, in contrast to the claims of the official datastax documentation where it supposedly scales linearly.

To be more precise, 2 x m3.large nodes and 4 x m3.large nodes have almost identical performance, when running the workloada with 100 or 200 threads and 100k or 1m operations. I use the EC2Snitch and 2 seeds nodes on a single region cluster. commitlog_sync: batch and commitlog_sync_batch_window_in_ms: 50 are uncommented.

What am I doing wrong and what I might me missing?

Thanks