We simulate benchmarks from the SPECCPU2006 suite [22]. We use SimPoint [18] to identify a single characteristic interval (i.e., simpoint) of each benchmark. Each benchmark is run with the first reference input. We first run 50 million instructions to warm up the internal structures and then run the simulation for 100 million instructions. For multi-core simulations, each benchmark runs simultaneously with the others, restarting after 100 million instructions, until all of the benchmarks have executed at least 100 million instructions after the warmup. We use instructions-per-cycle (IPC) to measure single-core performance. For multi-core workloads, we use the weighted speedup normalized to the baseline. That is, we compute IPCi for each application i sharing the L3 cache, then compute IsolatedIPCi for application i running in isolation with the same L3 cache size. We compute the weighted IPC of a workload as sum of its components’ (IPCi/IsolatedIPCi).