Table of Contents
Statistical profiling
Minix 3 supports statistical profiling. It is available in two versions. The basic one is based on BIOS RTC interrupts. This is available on any x86 machine and in virtual machines. As interrupts are not enabled in kernel this method does not make kernel profiling possible. The second method is using non-maskable interrupts (NMI) which also interrupt the kernel. This requires APIC mode. The NMIs are generated using performance counters. Each platform provides different counters, therefore Minix has to support each platform separately.
Prerequisites
Before statistical profiling can be used, some configuration changes must be made in MINIX. In /usr/src/include/minix/config.h, set the SPROFILE constant to 1. This causes the kernel to allocate a huge buffer to store profiling samples. After these changes, a make world command (in /usr/src) is needed to compile and install the new image.
If you want to use NMI rather than RTC, additional configuration changes are needed. In /usr/src/kernel/config.h, add:
#define USE_APIC 1 #define USE_ACPI 1 #define USE_WATCHDOG 1
After this, the boot image must be recompiled and installed (make install in /usr/src/tools or simply make world in /usr/src to do everything in one go). You must also pass some kernel parameters from the bootloader to make this work. Edit /boot.cfg to add a menu item that includes the following arguments: watchdog=1 no_apic=0 acpi=yes. Note that /boot.cfg is overwritten when the system is compiled, so it may be more convenient to change the local additions file /etc/boot.cfg.local, that is never overwritten, instead. After booting this new option, NMI statistical profiling should work.
Usage
profiling is started and stopped using the profile utility, collected result are written in a file. Use sprofalyze script to process the data and visualize the results.
Statistical Profiling:
profile start [--rtc | --nmi] [-m memsize] [-o outfile] [-f frequency]
profile stop
--rtc is default, --nmi allows kernel profiling
Call Profiling:
profile get [-m memsize] [-o outfile]
profile reset
- memsize in MB, default: 64
- default output file: profile.{stat|call}.out
- sample frequencies for --rtc (default: 6):
3 8192 Hz 10 64 Hz
4 4096 Hz 11 32 Hz
5 2048 Hz 12 16 Hz
6 1024 Hz 13 8 Hz
7 512 Hz 14 4 Hz
8 256 Hz 15 2 Hz
9 128 Hz
NMI mode
When using NMI more any frequncy may be specified in Hz, e.g. -f 10000 for 10kHz.
Beware, if no frequency is specified 6 is taken as default which in case of NMI means 6Hz!
Suported platforms:
- Intel architecture performance counters - virtually any reasonably Intel cpu
Sample sprofalyze output
================================================================================
Data file: profile.stat.out
================================================================================
System process ticks: 19935 ( 51%)
User process ticks: 15339 ( 39%) Details of system process
Idle time ticks: 3690 ( 10%) samples, aggregated and
---------- ---- per process, are below.
Total ticks: 38964 (100%)
--------------------------------------------------------------------------------
Total system process time 19935 samples
--------------------------------------------------------------------------------
kernel _ipc_entry ***** 5.6%
kernel _read_tsc **** 4.6%
kernel ___switch_address_space *** 3.9%
kernel _phys_copy *** 3.9%
kernel _kernel_call_entry *** 3.8%
kernel _isokendpt_f *** 3.6%
kernel _context_stop *** 3.5%
kernel _read_tsc_64 ** 2.4%
kernel _switch_to_user ** 2.4%
kernel _do_ipc ** 2.3%
vm __do_kernel_call ** 2.2%
kernel _restore_user_context ** 1.8%
kernel _sub64 ** 1.8%
kernel fill_start ** 1.7%
kernel _copy_msg_to_user ** 1.6%
kernel _lapic_restart_timer * 1.4%
vfs _release * 1.4%
mfs _search_dir * 1.3%
mfs __receive * 1.3%
vfs __sendrec * 1.1%
kernel _virtual_copy_f * 1.1%
kernel _mini_send * 1.1%
mfs __send * 1.1%
mfs _get_block * 1.1%
kernel _lin_lin_copy * 1.0%
<1% ******************************* 43.0%
--------------------------------------------------------------------------------
total 100.0%
--------------------------------------------------------------------------------
kernel 59.3% of system process samples
--------------------------------------------------------------------------------
_ipc_entry ************************** 9.4%
_read_tsc ********************** 7.8%
___switch_address_space ******************* 6.6%
_phys_copy ****************** 6.5%
_kernel_call_entry ****************** 6.4%
_isokendpt_f ***************** 6.0%
_context_stop ***************** 5.9%
_read_tsc_64 ************ 4.1%
_switch_to_user ************ 4.0%
_do_ipc *********** 3.9%
_restore_user_context ********* 3.1%
_sub64 ********* 3.0%
fill_start ******** 2.9%
_copy_msg_to_user ******** 2.7%
_lapic_restart_timer ******* 2.3%
_virtual_copy_f ****** 1.9%
_mini_send ****** 1.8%
_lin_lin_copy ***** 1.7%
_enqueue **** 1.4%
_dequeue **** 1.4%
_copy_msg_from_user **** 1.3%
_createpde **** 1.2%
<1% **************************************** 14.7%
--------------------------------------------------------------------------------
kernel 100.0%
--------------------------------------------------------------------------------
vm 15.0% of system process samples
--------------------------------------------------------------------------------
__do_kernel_call **************************************** 14.6%
_pt_ptalloc_in_range ****************** 6.5%
__kernel_call *************** 5.4%
_pt_writemap ************** 5.2%
_slaballoc ************** 5.1%
_physr_incr_iter ************* 4.6%
_physr_remove ********** 3.7%
_physr_insert ********** 3.5%
__receive ********** 3.4%
_map_ph_writept ******* 2.4%
_pt_new ****** 2.2%
_slabfree ****** 2.0%
_physr_search ****** 2.0%
_pt_free ***** 1.8%
_physr_start_iter_least ***** 1.7%
_addr_insert ***** 1.7%
_addr_remove ***** 1.6%
_do_fork ***** 1.6%
_main **** 1.5%
_objstats **** 1.4%
_map_lookup **** 1.4%
_vm_isokendpt **** 1.4%
_addr_search **** 1.3%
_sys_vmctl **** 1.3%
__send **** 1.2%
_addr_start_iter **** 1.2%
_map_writept **** 1.2%
_map_pf *** 1.1%
_map_new_physblock *** 1.1%
_map_subfree *** 1.0%
_physr_get_iter *** 1.0%
<1% **************************************** 14.9%
--------------------------------------------------------------------------------
vm 100.0%
