Hi Tim, here is what I found in my experience:
> 1. Employ a tickless kernel, probably a kernel close to 2.6.24. This has very little effect, because C-states have little effect on this processor. > 2. Turn on the ondemand governor (set > /sys/devices/system/devices/system/cpu/cpu*/cpufreq/scaling_governor to > ondemand). > > 3. Trun on the menu idle governor (set > /sys/devices/system/cpu/cpuidle_governor_ro to menu) This was both activated automatically on my Fedora/Rawhide system. > 4. Turning on the power aware scheduler (set > /sys/devices/system/cpu/sched_mc_power_savings to 1) No big effect observed so far. This may be due to the fact that we need to bind our Benchmark JVMs to CPUs (each JVM is bound to 2 cores sharing a cache). If we don't, we loose about 3% max throughput which ruins the benchmark result (no way to compensate that by saving power on lower loads). > 5. Use a larger sampling (like 1000000) versus the default value of 2000 > in /sys/devices/system/cpu/cpu*/cpufreq/ondemand/sampling_rate. For > SPECpower workload, the load level is very steady, so it does not pay to > shift gear to a high cpu frequency too quickly. This was the key for better savings in the intermediate load range. I don't have the numbers here right now, but it's around 10W at loadlevels 20%-60%. Thanks a lot! It even helped the benchmark result by improving the maximum throughput by ~1%, which is very important for the final benchmark result. I got the same benchmark result as under Windows now, despite a slightly lower max throughput. Martin -- Martin Wilck PRIMERGY System Software Engineer FSC IP ESP DE6 Fujitsu Siemens Computers GmbH Heinz-Nixdorf-Ring 1 33106 Paderborn Germany Tel: ++49 5251 8 15113 Fax: ++49 5251 8 20209 Email: mailto:[EMAIL PROTECTED] Internet: http://www.fujitsu-siemens.com Company Details: http://www.fujitsu-siemens.com/imprint.html _______________________________________________ Discuss mailing list [email protected] http://mail.lesswatts.org/mailman/listinfo/discuss
