On Mon, May 4, 2020 at 2:13 AM Nikolay Borisov <nborisov@xxxxxxxx> wrote:
>
>
>
> On 4.05.20 г. 11:03 ч., Chris Murphy wrote:
> > receive (rw,noatime,seclabel,compress-force=zstd:5,space_cache=v2,subvolid=5,subvol=/)
> > send (rw,noatime,seclabel,compress=zstd:1,nossd,notreelog,space_cache=v2,subvolid=5,subvol=/)
> >
> > Both are on dm-crypt.
> >
> > perf top -g -U consumes about 85% CPU according to top, and every time
> > I run it, the btrfs send performance *increases*. When I cancel this
> > perf top command, it returns to the slower performance. Curious.
> >
>
> Well this still doesn't show the stack traces, after all the + sign
> means you can expand that (with the 'e' key). But looking at this I
> don't see any particular lock contention - just compression-related
> functions.
I'm not sure which ones...
Samples
Children Self Shared Object Symbol
- 62.58% 0.10% [kernel] [k]
entry_SYSCALL_64_after_hwframe
◆
- 62.47% entry_SYSCALL_64_after_hwframe
▒
- 62.17% do_syscall_64
▒
- 23.84% ksys_read
▒
- 23.62% vfs_read
▒
- 14.79% proc_reg_read
▒
- seq_read
▒
- 7.07% s_show
▒
- seq_printf
▒
- vsnprintf
▒
1.87% format_decode
▒
1.49% number
▒
0.84% string
▒
0.68% memcpy_erms
▒
- 6.23% s_next
▒
- update_iter
▒
4.49% module_get_kallsym
▒
1.41% kallsyms_expand_symbol.constprop.0
▒
- 0.79% s_start
▒
- update_iter
▒
0.57% module_get_kallsym
▒
- 8.38% new_sync_read
▒
- 8.35% pipe_read
▒
- 6.46% __mutex_lock.constprop.0
▒
6.33% mutex_spin_on_owner
▒
- 0.86% copy_page_to_iter
▒
- 0.78% copyout
▒
0.77% copy_user_enhanced_fast_string
▒
- 17.96% __x64_sys_splice
▒
- 17.92% do_splice
▒
7.80% mutex_unlock
▒
- 4.55% pipe_double_lock
▒
- 2.88% mutex_lock
▒
0.95% _cond_resched
▒
- 2.61% mutex_lock
▒
0.82% _cond_resched
▒
0.52% pipe_unlock
▒
- 9.80% __x64_sys_ioctl
▒
- ksys_ioctl
▒
- 9.79% rpc_populate.constprop.0
▒
For a higher level overview, try: perf top --sort comm,dso
▒
--
Chris Murphy
