>-----Original Message----- >From: [EMAIL PROTECTED] >[mailto:[EMAIL PROTECTED] On Behalf Of >Vaidyanathan Srinivasan >Sent: Monday, February 25, 2008 9:24 AM >To: Martin Wilck >Cc: [email protected] >Subject: Re: [Discuss] SPECpower Linux vs. Windows > >* Martin Wilck <[EMAIL PROTECTED]> [2008-02-25 >15:42:13]: > >> Hello, >> >> we recently did SPECpower benchmark measurements on a 2xQuad-Core >> Harpertown (Intel E5440) Server and found that the results were >> generally slightly better under Windows than under Linux, >despite Linux >> technology like the tickless kernel. The reason was twofold: >> >> 1 The raw performance value under Windows is 1-2% better. >This seems to >> be a matter of optimization of the JVM used (BEA), not a lot >we can do >> about it. >> >> 2 The power consumption under Linux at 100% and 0% load is >approx. the >> same as under Windows. But in the intermediate levels >50%-10%, Windows >> is roughly 3-5W better (increasing towards lower load levels). This >> suggests that Windows power management is more aggressive >than Linux. Is >> there a description somewhere how to modify cpuidle governor >tunables? >> >> The kernel used was Fedora Rawhide 2.6.24-fc9 (x86_64). We >are about to >> upgrade to an even more recent kernel. The "menu" cpuidle >governor was used. >> >> For those who don't know the SPECpower benchmark: It uses a >Java-based >> database-type load which runs entirely in system memory (no disk IO). >> First a maximum load is calibrated. After that, the load is >reduced in >> steps of 10%. A final "active idle" measurement has the JVM >loaded, but >> no requests. >> >> We'd be thankful for any suggestions wrt choice of the >kernel, tunables, >> and other hints, which hel us to improve the Linux results. > >Hi Martin, > >Some comparison points that come to my mind are: > >* Did Linux exploit all the P-States available >* Whats the P-State and C-State usage at various loads >* PowerTop may provide some clue > >Tunable: > >echo 1 > /sys/devices/system/cpu/sched_mc_power_savings > >Please check if CONFIG_SCHED_MC=y in the kernel among other power >management related options like > >CONFIG_TICK_ONESHOT=y >CONFIG_NO_HZ=y >CONFIG_HIGH_RES_TIMERS=y > >I am trying to play with these tunables and bring down the idle power. >
We have seen sched_mc_power_savings mentioned above help SPECpower. Another _hack_ that seems to help the benchmark is to increase the ondemand sampling rate at /sys/devices/system/cpu/cpu0/cpufreq/ondemand/sampling_rate. You may have to try different values upto max to get best number. Thanks, Venki _______________________________________________ Discuss mailing list [email protected] http://mail.lesswatts.org/mailman/listinfo/discuss
