Experimental Results

In this section, we highlight the key features from the design of SSM. We present benchmarks illustrating each of the recovery-friendly, self-tuning, and self-protecting features. We also present numerous benchmarks demonstrating the self-healing nature of SSM when integrated with Pinpoint.

Each benchmark is conducted on the UC Berkeley Millennium Cluster. Our load generator models hyperactive users who continually make requests to read and write session state; each hyperactive user is modeled using a thread which does a sequence of alternating write and read requests, which is representative of the workload for session state, as described earlier in Section 2. As soon as a request returns, the thread immediately makes a subsequent request. In the following benchmarks, we vary the number of sending threads as well as the number of bricks. All bricks are run on separate, dedicated machines.