Kdyz ladim casovani pameti, tak pouzivam ramsmp benchmark (ramspeed/smp
http://alasir.com/software/ramspeed). Nu i koukam, ze vysledky maji daleko vetsi rozptyl a jsou o neco nizsi, nez drive. Tak jsem vse prozkoumal a na vine je transparent hugepages, ktery je ted zapnut defaultne asi ve vsech distribucich.
S vypnutym nebo madvise mam v MMX nt vysledky o 7% vyzsi, mam K10 a DDR3. Mohl by nekdo vyzkouset nasledujici skript na nejakem intelu, jestli taky dojde ke zpomaleni? Jenom staci zmenit cpus na pocet jader, u intelu nevim, jestli zapocitat HT nebo ne, jestli se vysledek zlepsi. cycles 10 asi staci, pro presnejsi mereni muzete dat vic. Mezi madvise a never nevidim rozdil, always je pomalejsi, v SSE nt asi o 1.5% v MMX nt asi o 7.3%.
cpus=4
cycles=10
#test=(always madvise never)
test=(always madvise)
for i in /sys/devices/system/cpu/cpu?
do
echo "performance" > ${i}/cpufreq/scaling_governor
done
for i in ${test[@]}
do
echo $i > /sys/kernel/mm/transparent_hugepage/enabled
date >> log
cat /sys/kernel/mm/transparent_hugepage/enabled >> log
ramsmp -b3 -l$cycles -p$cpus | grep BatchRun | grep AVE >> log
ramsmp -b6 -l$cycles -p$cpus | grep BatchRun | grep AVE >> log
ramsmp -b9 -l$cycles -p$cpus | grep BatchRun | grep AVE >> log
ramsmp -b12 -l$cycles -p$cpus | grep BatchRun | grep AVE >> log
ramsmp -b15 -l$cycles -p$cpus -t0 | grep BatchRun | grep AVE >> log
ramsmp -b15 -l$cycles -p$cpus -t1 | grep BatchRun | grep AVE >> log
ramsmp -b15 -l$cycles -p$cpus -t2 | grep BatchRun | grep AVE >> log
ramsmp -b18 -l$cycles -p$cpus -t0 | grep BatchRun | grep AVE >> log
ramsmp -b18 -l$cycles -p$cpus -t1 | grep BatchRun | grep AVE >> log
ramsmp -b18 -l$cycles -p$cpus -t2 | grep BatchRun | grep AVE >> log
done
for i in /sys/devices/system/cpu/cpu?
do
echo "ondemand" > ${i}/cpufreq/scaling_governor
done