Our Experimental Setup

Our Shark was equipped with: $8$ GB cache (per cluster), $2$ GB NVS (per cluster), four $600$ MHz PowerPC/RS64IV CPUs (per cluster), and $16$ RAID-$5$ ($6$ + parity + spare) arrays with $72$ GB, $10$K rpm drives. We use only one cluster since enhancements provided by dual cluster are not necessary to study effectiveness of cache algorithms. In a single cluster mode, the behavior of our experimental software on Shark does not change other than the fact that writes now go to both the NVS memory as well as the SMP RAM in the same cluster. The essence of our results does not change with a larger or a smaller cache size.

As a client, we use a powerful AIX host with the following configuration: 16 GB RAM, 2-way SMP with 1GHz PowerPC/Power4 CPUs. The host is connected to the Shark through two fiber channel cards which can support more than sufficient bandwidth for our all experiments.