 |
In this section, we compare three basic memory power-management techniques: the default built-in power-management policy implemented in current RDRAM memory controllers (Base), a simple On/Off technique, and our initial PAVM implementation. Recall from Section 2 that, under the Base policy, the controller keeps devices in Standby mode, and quickly switches them to Attention mode when accessed. The On/Off technique simply involves putting all nodes into Nap mode upon detecting system is idle, and restoring all nodes to Standby when any process is ready to run. It requires minimal kernel modifications to implement, and is worth considering here for its simplicity. PAVM, as described in Section 4.1, is compared with these two methods.
As the Base policy always keeps nodes in Standby, while the
other two put all nodes into Nap mode when the system is idle, we use 
for Base and 
for the other policies when
computing energy with Eq. 1. As neither Base nor On/Off uses Nap mode while processes are running,
the second term in Eq. 1 simplifies to
for these two policies.
To show how these power-management policies perform in real systems, we run the three workloads described earlier. The results are shown in Figures 6(a-c) for Light, Poweruser, and Multimedia workloads, respectively. Each graph shows cumulative energy consumed over time, normalized with respect to the Base policy. As one can see, for Light and Poweruser workloads, the simple On/Off policy performs well since it can exploit the large amount of idle time in the system to put nodes into Nap mode. With the Multimedia workload, idle time in the system is minuscule, and therefore, the On/Off policy approaches the Base policy at the end of the workload. PAVM, on the other hand, not only exploits idle time, but also reduces memory power dissipation when processes are actively running. Compared to the On/Off policy, it can save an additional 48-66%, 51-63%, 30-62% of energy for Light, Poweruser, and Multimedia workloads, respectively.