kernbench-16 on 2.5.59 vs 2.5.59-mm6

This test does a make -j X vmlinux on a 2.4.17 kernel compile with a very large config set. X is 16 times the number of cpus. This is on a 16-way NUMA-Q so we end up with a make -j256 (it's fastest with about 1.5 * num_cpus), but this test puts more stress on the kernel.

None of the other tests I ran showed anything very interesting. (the new NUMA sched stuff from Ingo seems to give mild degredations in -mjb ... probably needs some more tuning).

Going from 59 to 59-mm6, I get:

Kernbench-16:

                                   Elapsed        User      System         CPU
                        2.5.59       47.45      568.02      143.17     1498.17
                    2.5.59-mm6       47.18      567.15      138.62     1495.50

Summary: Scheduler stuff seems like a wash (schedule -> do_schedule). Seems to be some sort of rearrangement of the dcache stuff which appears to be mildly beneficial (what's going in there?). current_kernel_time seems to be less than half the cost, I'm assuming the new frlock kernel time stuff is doing that. This workload doesn't stress that very much, so I'll find a better test for that one ...

2.5.59: 1657 current_kernel_time
2.5.59-mm6: 747 current_kernel_time

diffprofile (+ gets worse, - gets better).

2023 do_schedule
485 dentry_open
289 .text.lock.file_table
132 clear_page_tables
131 pgd_ctor
113 vma_merge
75 kmap_atomic
62 get_empty_filp
51 can_vma_merge_after
-52 dget_locked
-54 vfs_follow_link
-55 kmem_cache_free
-66 buffered_rmqueue
-74 __copy_to_user_ll
-94 page_add_rmap
-102 fd_install
-110 __copy_from_user_ll
-117 __d_lookup
-157 do_generic_mapping_read
-188 path_lookup
-273 .text.lock.dec_and_lock
-275 file_ra_state_init
-283 do_anonymous_page
-331 pfn_to_nid
-405 page_remove_rmap
-413 pgd_alloc
-427 vm_enough_memory
-910 current_kernel_time
-1222 .text.lock.namei
-2076 total
-2133 schedule

The pagetable preconstruction cache hit is spread across clear_page_tables() and pgd_ctor() with the pgd_ctor patches. This is the equivalent of the explicit zeroing in pgd_alloc().

Your result appears to imply the overhead has been reduced by 36%, which is useful evidence for the PAE case. Before this the pgd_alloc() overhead had only been observed on non-PAE systems.

Now, YTF hadn't I seen this before if all it took to bring it out was a kernel compile? Perhaps diffprof (I prefer the multiplicative flavor but nm that) of some flavor was lacking.

-- wli