Our goal with phase-guided auto-tuning is to improve the utilization of performance asymmetric multi-core processors. Phase-guided auto-tuning can be divided into two parts. First, group together code segments that are likely to exhibit similar runtime behavior. Second, (the key idea) use actual runtime characteristics of a small number of code segments to give insight into the behavior of all similar code segments.

Experimental results are from an actual quad core machine with two cores running at 2.4GHz and two cores running at 1.6GHz. Workloads are made up of 18 to 84 SPEC CPU 2000 and 2006 benchmarks running simultaneously. Upon completion of a benchmark, another is immediately started to maintain constant workload size. Here, our technique is compared to the stock Linux scheduler.

We measure three things: overhead, throughput, and fairness. Details for each of these are given separately. Here, we describe the overall performance for variations of our technique.

Speedup vs Fairness

Figure 1: Application speedup vs fairness for variations of our technique.

Speedup vs Throughput

Figure 2: Application speedup vs throughput for variations of our technique.

Throughput vs Fairness

Figure 3: Throughput vs fairness for variations of our technique.

---

1. These results are shown in our latest technical report.