Menu
▾
▴
[fuse-devel] Kernel 2.6.26 very slow.
|
From: Paul S. <pa...@up...> - 2008-08-09 15:12:43
|
Hi Miklos and other filesystem guru's.
I upgraded the kernel on my pc from 2.6.20.11 to 2.6.26 and fuse-2.8.0
from the cvs (co yesterday).
I see very slow performance compared to 2.6.20.11.
I like to use bonnie++ to test certain aspects of my filesystem.
I ran "time bonnie++ -uroot -s0 -n100" under 2.6.20.11 with the
following results:
Version 1.03 ------Sequential Create------ --------Random
Create--------
debian -Create-- --Read--- -Delete-- -Create-- --Read---
-Delete--
files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
/sec %CP
100 24592 29 +++++ +++ 37742 23 22396 26 +++++ +++
36432 23
debian,,,,,,,,,,,,,,100,24592,29,+++++,+++,37742,23,22396,26,+++++,+++,36432,23
User:0.312 Kernel:3.876 Total:0:14.93 CPU: 27.9%
The same command under 2.6.26 results in:
Version 1.03 ------Sequential Create------ --------Random
Create--------
debian -Create-- --Read--- -Delete-- -Create-- --Read---
-Delete--
files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
/sec %CP
100 8054 37 146627 41 13037 31 7790 34 +++++ +++
13432 32
debian,,,,,,,,,,,,,,100,8054,37,146627,41,13037,31,7790,34,+++++,+++,13432,32
User:0.368 Kernel:14.608 Total:0:42.47 CPU: 35.2%
The newer kernel is slower and uses a lot more CPU. (14.93 sec vs 42.47
sec. total time) (27.9% vs 35.2% CPU usage).
I did an "oprofile -l" on both but could not isolate the problem.
2.6.20.11:
CPU: Core 2, speed 2397.7 MHz (estimated)
Counted CPU_CLK_UNHALTED events (Clock cycles when not halted) with a
unit mask
of 0x00 (Unhalted core cycles) count 100000
samples % app name symbol name
19619 4.7623 vmlinux-2.6.20.11 mwait_idle_with_hints
11306 2.7444 vmlinux-2.6.20.11 fuse_dev_read
10661 2.5878 vmlinux-2.6.20.11 system_call
8945 2.1713 libc-2.3.6.so strlen
8081 1.9616 vmlinux-2.6.20.11 task_rq_lock
8049 1.9538 libc-2.3.6.so strncpy
7901 1.9179 vmlinux-2.6.20.11 kunmap_atomic
7019 1.7038 libJudy.so.1.0.3 j__udyLGet
6931 1.6824 vmlinux-2.6.20.11 try_to_wake_up
6909 1.6771 libpthread-2.3.6.so pthread_mutex_lock
6582 1.5977 libc-2.3.6.so _int_malloc
6108 1.4826 libc-2.3.6.so memset
5642 1.3695 libJudy.so.1.0.3 j__udyInsWalk
5345 1.2974 libpthread-2.3.6.so pthread_mutex_unlock
4944 1.2001 libJudy.so.1.0.3 JudyLNext
4834 1.1734 vmlinux-2.6.20.11 find_busiest_group
4665 1.1324 libJudy.so.1.0.3 j__udyDelWalk
4661 1.1314 vmlinux-2.6.20.11 prepare_to_wait
4613 1.1197 libJudy.so.1.0.3 JudySLGet
-----snip-----
2.6.26:
samples % app name symbol name
209608 20.9357 vmlinux-2.6.26 mach_get_cmos_time
56031 5.5964 vmlinux-2.6.26 lock_acquired
42243 4.2192 vmlinux-2.6.26 __lock_acquire
30939 3.0902 vmlinux-2.6.26 clear_lock_stats
27708 2.7675 vmlinux-2.6.26 lock_stats
22123 2.2096 vmlinux-2.6.26 __rwlock_init
20888 2.0863 vmlinux-2.6.26 lockdep_reset
20378 2.0354 vmlinux-2.6.26 in_gate_area
18412 1.8390 vmlinux-2.6.26 lockdep_free_key_range
17508 1.7487 vmlinux-2.6.26 lockdep_init_map
17351 1.7330 vmlinux-2.6.26 irq_entries_start
12552 1.2537 vmlinux-2.6.26 count_matching_names
10445 1.0432 vmlinux-2.6.26 fuse_copy_args
9208 0.9197 libpthread-2.3.6.so pthread_mutex_lock
9022 0.9011 libc-2.3.6.so _int_malloc
9020 0.9009 libJudy.so.1.0.3 j__udyLGet
8721 0.8711 libc-2.3.6.so strlen
8564 0.8554 vmlinux-2.6.26 unmap_vmas
8558 0.8548 libc-2.3.6.so strncpy
7609 0.7600 vmlinux-2.6.26 __spin_lock_init
7275 0.7266 libpthread-2.3.6.so pthread_mutex_unlock
6647 0.6639 libc-2.3.6.so memset
One can see from this that there is a lot more kernel overhead in 2.6.26.
"mach_get_cmos_time" is by far the function that is called the most.
It would probably help the performance a lot if I can eliminate it.
I therefore tried to include/exclude various RTC modules during the
kernel build.
None made a difference.
Did I mess up the kernel config ? If so please tell me what to include
and exclude.
Or maybe this is the price for "mmap and big_writes" ?
If anyone has any insight into this, please let me know.
Regards
Paul
|