8 x P3 700Mhz
tbench 192
2.6.0-test5: Throughput 224.949 MB/sec 192 procs
2.6.0-test5-mm1: Throughput 329.827 MB/sec 192 procs
It seems the tbench likes the interactivity tweaks.
Con
> 8 x P3 700Mhz
> tbench 192
> 2.6.0-test5: Throughput 224.949 MB/sec 192 procs
> 2.6.0-test5-mm1: Throughput 329.827 MB/sec 192 procs
> It seems the tbench likes the interactivity tweaks.
4 x P3 700Mhz likes the scheduler tweaks too.
The 8 way box has about 2x throughput of 4 way.
It appears between test2-mm2 and test3-mm2 a big throughput jump
took place. Average below is MB/second for 5 runs. +/- gives
an idea of max variation between runs. (+/- = (high - low) / 2)
The variation between runs is smaller in recent -mm, which may
imply more consistency.
tbench-1.3 192 processes Average +/- MB/sec
2.6.0-test1-bk2 114.8 6.9
2.6.0-test2-bk5 115.3 4.7
2.6.0-test2-mm1 113.2 10.9
2.6.0-test2-mm2 123.2 1.9
2.6.0-test3 114.5 8.8
2.6.0-test3-mm2 151.1 0.7
2.6.0-test3-mm2-gcc-3.3.1 152.8 1.2
2.6.0-test3-mm2-gcc-3.3.1-Os 147.0 1.6
2.6.0-test3-mm2-gcc-3.3.1-Os-falign=2 147.7 1.9
2.6.0-test4 115.0 5.4
2.6.0-test4-mm4 148.0 2.6
2.6.0-test5 113.2 6.2
For fun, test3-mm2 was tested with gcc-2.96, and gcc-3.3.1
with a couple different optimization settings.
This is the readprofile difference between 2.6.0-test4-mm4
and 2.6.0-test5. schedule, __wake_up, and mod_timer and
__mod_timer are where some of the biggest differences occured.
tbench profile
test4-mm4 test5
total 788882 1174599
schedule 200531 249842
default_idle 137447 117369
__wake_up 110705 253144
__mod_timer 77932 211737
local_bh_enable 62582 57927
__might_sleep 59081 0
do_gettimeofday 36704 43132
do_softirq 27731 30611
mod_timer 20784 71059
get_offset_tsc 15790 18543
prepare_to_wait 14032 25463
sched_clock 10398 0
schedule_timeout 5051 6756
finish_wait 3831 17514
del_timer 1587 1793
sys_gettid 1269 1811
add_timer_on 948 0
kmap 723 329
kmap_atomic 464 188
kunmap 421 495
cpu_idle 152 37
do_page_fault 138 154
pte_alloc_one 98 59
copy_mm 60 38
copy_process 58 41
eligible_child 48 102
sys_wait4 41 23
pgd_ctor 37 19
dup_task_struct 32 9
release_task 24 51
exit_notify 22 33
copy_files 16 8
kunmap_atomic 15 430
get_signal_to_deliver 14 10
current_kernel_time 13 14
flush_tlb_page 8 24
do_exit 8 11
group_send_sig_info 8 5
do_sigaction 6 5
schedule_tail 5 4
wake_up_forked_process 4 6
_stext 4 4
sys_mmap2 4 4
mmput 4 3
del_timer_sync 4 0
wait_task_zombie 3 3
count_open_files 3 0
init_new_context 3 0
put_files_struct 3 0
mmgrab 2 8
get_wchan 2 6
handle_IRQ_event 2 6
flush_tlb_mm 2 5
release_x86_irqs 2 4
next_thread 2 0
sys_setitimer 2 0
sys_sigreturn 1 0
.text.lock.fault 1 0
sys_rt_sigprocmask 1 0
system_call 0 58012
syscall_exit 0 6713
syscall_call 0 677
work_resched 0 292
__cond_resched 0 15
mm_release 0 4
handle_signal 0 4
mm_init 0 3
free_task 0 3
ksoftirqd 0 3
restore_all 0 3
ret_from_intr 0 3
Zeros above are when one kernel did not have that
function in readprofile's top 60.
--
Randy Hron
http://home.earthlink.net/~rwhron/kernel/bigbox.html