Re: [PATCH] GFP_THISNODE for the slab allocator

[David Rientjes <rientjes@google.com>]

On Fri, 15 Sep 2006, David Rientjes wrote:

> On Fri, 15 Sep 2006, Andrew Morton wrote:
> 
> > Yes.  Speeding up get_page_from_freelist() is less than totally trivial. 
> > I've been putting off thinking about it until we're pretty sure that there
> > aren't any other showstoppers.
> > 
> > I'm (very) impressed at how well the infrastructre which you and Christoph
> > have put together has held up under this.
> > 
> 
> I used numa=fake=64 for 64 nodes of 48M each (with my numa=fake fix).  I 
> created a 2G cpuset with 43 nodes (43*48M = ~2G) and attached 'usemem -m 
> 1500 -s 10000000 &' to it for 1.5G of anonymous memory.  I then used 
> readprofile to time and profile a kernel build of 2.6.18-rc5 with x86_64 
> defconfig in the remaining 21 nodes.
> 
> Kernel build within the 2G container:
> 	real	5m23.057s
> 	user	9m13.395s
> 	sys	1m15.417s
> 

Kernel build within 2G cpuset with Christoph's patch:
	real	5m15.758s
	user	9m10.398s
	sys	1m9.072s

> Unrestricted kernel build (no NUMA emulation or usemem):
> 	real	5m3.213s
> 	user	9m19.483s
> 	sys	0m32.014s	<-- over twice faster
> 
> In 2G container:
> 	10599 __cpuset_zone_allowed			50.4714
> 	 3521 mwait_idle				45.1410
> 	 1149 clear_page				20.1579
> 	   24 clear_page_end				 3.4286
> 	  215 find_get_page				 3.0282
> 	  110 pfn_to_page				 2.3913
> 	  130 __down_read_trylock			 1.9697
> 	   86 page_remove_rmap				 1.9545
> 	  150 find_vma					 1.7241
> 	   46 __strnlen_user				 1.1795
> 	   32 nr_free_pages				 1.1034
> 	   55 page_to_pfn				 1.0784
> 	   22 page_add_file_rmap			 1.0000
> 	  829 get_page_from_freelist			 0.8904
> 	 1548 do_page_fault				 0.8586
> 	   17 file_ra_state_init			 0.8500
> 	   63 _atomic_dec_and_lock			 0.7500
> 	   85 ia32_sysenter_target			 0.7083
> 	   47 cond_resched				 0.6912
> 	  198 copy_user_generic				 0.6644
> 

In 2G cpuset with Christoph's patch:
	  9232 __cpuset_zone_allowed                     43.9619
	  2083 mwait_idle                                26.7051
	   973 clear_page                                17.0702
	   131 pfn_to_page                                2.8478
	   201 find_get_page                              2.8310
	    18 clear_page_end                             2.5714
	   142 __down_read_trylock                        2.1515
	    87 page_remove_rmap                           1.9773
	   149 find_vma                                   1.7126
	    85 page_to_pfn                                1.6667
	    51 __strnlen_user                             1.3077
	    34 nr_free_pages                              1.1724
	    25 page_add_file_rmap                         1.1364
	    18 fput                                       1.0588
	    21 file_ra_state_init                         1.0500
	   108 ia32_sysenter_target                       0.9000
	    74 _atomic_dec_and_lock                       0.8810
	   818 get_page_from_freelist                     0.8786
	   133 zone_watermark_ok                          0.8526
	  1524 do_page_fault                              0.8453
	...

> Unrestricted:
> 	 3719 mwait_idle				47.6795
> 	 1083 clear_page				19.0000
> 	   20 clear_page_end				 2.8571
> 	  175 find_get_page				 2.4648
> 	   77 page_remove_rmap				 1.7500
> 	  114 __down_read_trylock			 1.7273
> 	   77 pfn_to_page				 1.6739
> 	  144 find_vma					 1.6552
> 	   60 __strnlen_user				 1.5385
> 	   71 page_to_pfn				 1.3922
> 	   24 page_add_file_rmap			 1.0909
> 	   17 fput					 1.0000
> 	   80 _atomic_dec_and_lock			 0.9524
> 	    4 up_write					 0.8000
> 	 1439 do_page_fault				 0.7981
> 	   13 compat_sys_open				 0.7647
> 	  227 copy_user_generic				 0.7617
> 	   89 ia32_sysenter_target			 0.7417
> 	   21 memcmp					 0.6562
> 	   13 file_ra_state_init			 0.6500
> 	...
> 	  389 get_page_from_freelist			 0.4178
> 

		David

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>