Re: 2.6.7rc1 vs 2.6.0

[FabF <fabian.frederick@skynet.be>]

Martin,

	When profiling latest mm against some fgrep I've got the following:

Extracting ext3 relevant fx:

Occ / Function
71 ext3_get_block_handle
68 ext3_dx_find_entry
57 ext3_get_branch
48 ext3_getblk
40 ext3_do_update_inode
35 ext3_find_entry
35 ext3_read_inode
31 ext3_block_to_path
30 ext3_get_inode_block
20 ext3_get_block
19 ext3_htree_store_dirent
15 ext3_lookup
12 ext3_release_file
12 ext3fs_dirhash
11 ext3_get_inode_loc
10 ext3_check_dir_entry

All those ext3_fxs consume almost 0.1% CPU and are not called as much as I thought

(e.g.
9573      3.2735  vmlinux                  __copy_to_user_ll
6914      2.3642  vmlinux                  mark_offset_tsc
4399      1.5042  vmlinux                  mask_and_ack_8259A
2703      0.9243  vmlinux                  fast_clear_page
1750      0.5984  vmlinux                  enable_8259A_irq
933       0.3190  vmlinux                  default_idle
826       0.2824  vmlinux                  irq_entries_start
822       0.2811  vmlinux                  ide_outb
614       0.2100  vmlinux                  apic_timer_interrupt
513       0.1754  vmlinux                  ide_inb
355       0.1214  vmlinux                  page_fault
331       0.1132  vmlinux                  find_vma
243       0.0831  vmlinux                  get_exec_dcookie
189       0.0646  vmlinux                  __d_lookup
170       0.0581  vmlinux                  unix_poll
160       0.0547  vmlinux                  do_generic_mapping_read
159       0.0544  vmlinux                  radix_tree_delete
155       0.0530  vmlinux                  __rmqueue
155       0.0530  vmlinux                  schedule
153       0.0523  vmlinux                  buffered_rmqueue
143       0.0489  vmlinux                  handle_mm_fault
133       0.0455  vmlinux                  disable_8259A_irq
130       0.0445  vmlinux                  delay_tsc
127       0.0434  vmlinux                  __page_cache_release
125       0.0427  vmlinux                  __find_get_block
124       0.0424  vmlinux                  __alloc_pages
124       0.0424  vmlinux                  kmem_cache_alloc
112       0.0383  vmlinux                  page_add_anon_rmap
102       0.0349  vmlinux                  link_path_walk
...)

So I guess we can consider the 2.6 / 2.6.7 I/O regression not fs relevant (ie ext3) in this case ....

PS: I'll try to bring same profile for 2.6.0.

Regards,
FabF

On Tue, 2004-05-25 at 16:27, Martin J. Bligh wrote:
> > 	Here's trivial fgrep vs report (using ffb1) :
> > 
> > 2.6.7rc1 : 
> > Grepping  /usr/bin  :
> > 45% cpu - 38.06 RT Sec - 1.10 Sec in KM
> > Entries scanned : 1527
> > 85112 Kb analyzed this time
> > 
> > 2.6.0 : 
> > Grepping  /usr/bin  :
> > 51% cpu - 33.12 RT Sec - 1.10 Sec in KM
> > Entries scanned :1527
> > 85112 Kb analyzed this time
> > 
> > 	This is done against ext3 fs. Same .config, same box, same box state.
> > What relevance could explain this 5s delta ? IOW, what big ext3, mm new functionnalities have been plugged in-between ?
> 
> Take kernel profiles of both (see Documentation/basic_profiling.txt)
> 
> M.
> 
>