Netflix is deploying Apache Cassandra as a data store for production use. While setting it up, Netflix has created benchmarks to stress-test Cassandra running on AWS, and they've shared a lot of results.
The Netflix team used M1 Extra Large (M1.XL) instances to run Cassandra, though they also use M2 Quadruple Extra Large (M2.4XL) units to run Cassandra in production. The clients were running the M2.4XL instance types. The clients ran in a single availability zone, but the Cassandra instances were running in three availability zones. Netflix used EC2's auto-scaling feature to create instances.
Netflix ran tests with 48, 96, 144 and 288 nodes. The results were impressive, as Netflix basically achieved linear scaling for client writes each time they doubled the number of nodes. I've included the chart from Netflix here.
An interesting observation buried in the results is that Netflix found network latency and response times are lower in AWS Europe than in the US East.
If you're wondering what kind of costs were incurred, Netflix shows us that too. For a 48-node test that ran about 15 minutes, the cost was $60.97. The 96-node test ran 22 minutes, and had a bill of $121.94. The 288-node test took 66 minutes and cost Netflix $561.86.
Overall, a really good look at how Netflix is using and benchmarking Cassandra in AWS.