[BENCHMARK] 12% improvement in fork load from mm3 to mm4

[BENCHMARK] 12% improvement in fork load from mm3 to mm4

[rwhron]

autoconf build test creates about 1.2 million processes in
35 minutes.  On quad xeon, there was an improvement of 
~ 12% between 2.5.70-mm3 and 2.5.70-mm4.

              Avg of 3 runs
2.5.69-mm5    830.3 seconds
2.5.70        732.8
2.5.70-mjb1   719.5
2.5.70-mm3    829.0
2.5.70-mm4    737.9
-- 
Randy Hron
http://home.earthlink.net/~rwhron/kernel/bigbox.html

Re: [BENCHMARK] 12% improvement in fork load from mm3 to mm4

[rwhron]

> autoconf build test creates about 1.2 million processes in
> 35 minutes.  On quad xeon, there was an improvement of
> ~ 12% between 2.5.70-mm3 and 2.5.70-mm4.
> 
>               Avg of 3 runs
> 2.5.70-mm3    829.0
> 2.5.70-mm4    737.9

profile shows the biggest change is in do_page_fault.


                    2.5.70-mm4 2.5.70-mm3
                 total  478507  521921
          default_idle  229230  231410
         do_page_fault   53727  103744
         pte_alloc_one   22227   21114
           system_call   21589   21750
         kunmap_atomic   17231   17091
               copy_mm   14913   13690
          copy_process   14514   12789
        flush_tlb_page   10946   11092
           sigprocmask    9478    9038
          release_task    7553    7243
           exit_notify    5923    5257
            copy_files    5473    5143
          do_sigaction    5249    5200
    sys_rt_sigprocmask    4831    4596
   current_kernel_time    4729    4428
           kmap_atomic    4566    4309
              schedule    4430    3903
       dup_task_struct    3741    3281
 get_signal_to_deliver    2255    2177
               do_exit    2096    1985
         schedule_tail    2002    1852
         ret_from_intr    1511    2298
      sys_rt_sigaction    1473    1367
          flush_tlb_mm    1419    1060
              old_mmap    1186    1096
             sys_wait4    1115    1332
           setup_frame     982     550
        del_timer_sync     963     968
      do_flush_tlb_all     950     879
               do_fork     912     768
             sys_mmap2     906     844
               mm_init     889     909
          syscall_exit     888     951
             __wake_up     886     814
 flush_signal_handlers     855     758
wake_up_forked_process     845     826
                 mmput     842     879
      put_files_struct     818     899
      .text.lock.fault     808    1489
      release_x86_irqs     754     505
         handle_signal     677     424
      wait_task_zombie     654     684
      setup_sigcontext     569     285
      count_open_files     525     399
           copy_thread     474     407
              mm_alloc     473     483
          syscall_call     453     524
     remove_wait_queue     449     463
            error_code     446     713
    restore_sigcontext     435     286
         sys_sigreturn     431     373
      init_new_context     424     327
        get_jiffies_64     412     344
        add_wait_queue     388     325
        eligible_child     374     378
            mm_release     369     316

-- 
Randy Hron
http://home.earthlink.net/~rwhron/kernel/bigbox.html

Re: [BENCHMARK] 12% improvement in fork load from mm3 to mm4

[Andrew Morton]

rwhron@earthlink.net wrote:
>
> > autoconf build test creates about 1.2 million processes in
> > 35 minutes.  On quad xeon, there was an improvement of
> > ~ 12% between 2.5.70-mm3 and 2.5.70-mm4.
> > 
> >               Avg of 3 runs
> > 2.5.70-mm3    829.0
> > 2.5.70-mm4    737.9
> 
> profile shows the biggest change is in do_page_fault.

Well that's amusing.  In mm4 I replaced the patch which runs
remap_file_pages() against all prot_exec mappings with a patch which just
prefaults those mappings into pagecache.

The difference _should_ be that in mm3, all pte's are set up at mmap time.
So -mm4 should have more pagefaults, not less.

Something fishy is going on there.  Makes one wonder if remap_file_pages()
is successfully avoiding the minor fault.