Messages in this thread | | | Date | Fri, 16 Aug 2013 09:36:39 -0700 | From | Dave Hansen <> | Subject | Re: [RFC v3 0/5] Transparent on-demand struct page initialization embedded in the buddy allocator |
| |
Hey Nathan,
Could you post your boot timing patches? My machines are much smaller than yours, but I'm curious how things behave here as well.
I did some very imprecise timings (strace -t on a telnet attached to the serial console). The 'struct page' initializations take about a minute of boot time for me to do 1TB across 8 NUMA nodes (this is a glueless QPI system[1]). My _quick_ calculations look like it's 2x as fast to initialize node0's memory vs. the other nodes, and boot time is increased by a second for about every 30G of memory we add.
So even with nothing else fancy, we could get some serious improvements from just doing the initialization locally.
[1] We call anything using pure QPI without any other circuitry for the NUMA interconnects to be "glueless"
| |