On Wed, Aug 14, 2013 at 1:35 PM, Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> wrote: > On Wed, Aug 14, 2013 at 01:05:33PM -0700, Dave Hansen wrote: >> On 08/14/2013 12:43 PM, Greg Kroah-Hartman wrote: >> > On Wed, Aug 14, 2013 at 02:31:45PM -0500, Seth Jennings wrote: >> >> ppc64 has a normal memory block size of 256M (however sometimes as low >> >> as 16M depending on the system LMB size), and (I think) x86 is 128M. With >> >> 1TB of RAM and a 256M block size, that's 4k memory blocks with 20 sysfs >> >> entries per block that's around 80k items that need be created at boot >> >> time in sysfs. Some systems go up to 16TB where the issue is even more >> >> severe. >> > >> > The x86 developers are working with larger memory sizes and they haven't >> > seen the problem in this area, for them it's in other places, as I >> > referred to in my other email. >> >> The SGI guys don't run normal distro kernels and don't turn on memory >> hotplug, so they don't see this. I do the same in my testing of >> large-memory x86 systems to speed up my boots. I'll go stick it back in >> there and see if I can generate some numbers for a 1TB machine. >> >> But, the problem on x86 is at _worst_ 1/8 of the problem on ppc64 since >> the SECTION_SIZE is so 8x bigger by default. >> >> Also, the cost of creating sections on ppc is *MUCH* higher than x86 >> when amortized across the number of pages that you're initializing. A >> section on ppc64 has to be created for each (2^24/2^16)=256 pages while >> one on x86 is created for each (2^27/2^12)=32768 pages. >> >> Thus, x86 folks with our small pages and large sections tend to be >> focused on per-page costs. The ppc folks with their small sections and >> larger pages tend to be focused on the per-section costs. > > Ah, thanks for the explaination, now it makes more sense why they are > both optimizing in different places. I had one local patch that sent before, it will probe block size for generic x86_64. set it to 2G looks more reasonable for system with 1T+ ram. Also can we add block_size in that /sys directly so could generate less entries ? Thanks Yinghai
Attachment:
block_size_x86_64.patch
Description: Binary data