Messages in this thread | | | Date | Mon, 04 May 2015 23:32:32 -0400 | From | Waiman Long <> | Subject | Re: [PATCH 0/13] Parallel struct page initialisation v4 |
| |
On 05/04/2015 05:30 PM, Andrew Morton wrote: > On Fri, 01 May 2015 20:09:21 -0400 Waiman Long<waiman.long@hp.com> wrote: > >> On 05/01/2015 06:02 PM, Waiman Long wrote: >>> Bad news! >>> >>> I tried your patch on a 24-TB DragonHawk and got an out of memory >>> panic. The kernel log messages were: >> ... >> >>> [ 81.360287] [<ffffffff8151b0c9>] dump_stack+0x68/0x77 >>> [ 81.365942] [<ffffffff8151ae1e>] panic+0xb9/0x219 >>> [ 81.371213] [<ffffffff810785c3>] ? >>> __blocking_notifier_call_chain+0x63/0x80 >>> [ 81.378971] [<ffffffff811384ce>] __out_of_memory+0x34e/0x350 >>> [ 81.385292] [<ffffffff811385ee>] out_of_memory+0x5e/0x90 >>> [ 81.391230] [<ffffffff8113ce9e>] __alloc_pages_slowpath+0x6be/0x740 >>> [ 81.398219] [<ffffffff8113d15c>] __alloc_pages_nodemask+0x23c/0x250 >>> [ 81.405212] [<ffffffff81186346>] kmem_getpages+0x56/0x110 >>> [ 81.411246] [<ffffffff81187f44>] fallback_alloc+0x164/0x200 >>> [ 81.417474] [<ffffffff81187cfd>] ____cache_alloc_node+0x8d/0x170 >>> [ 81.424179] [<ffffffff811887bb>] kmem_cache_alloc_trace+0x17b/0x240 >>> [ 81.431169] [<ffffffff813d5f3a>] init_memory_block+0x3a/0x110 >>> [ 81.437586] [<ffffffff81b5f687>] memory_dev_init+0xd7/0x13d >>> [ 81.443810] [<ffffffff81b5f2af>] driver_init+0x2f/0x37 >>> [ 81.449556] [<ffffffff81b1599b>] do_basic_setup+0x29/0xd5 >>> [ 81.455597] [<ffffffff81b372c4>] ? sched_init_smp+0x140/0x147 >>> [ 81.462015] [<ffffffff81b15c55>] kernel_init_freeable+0x20e/0x297 >>> [ 81.468815] [<ffffffff81512ea0>] ? rest_init+0x80/0x80 >>> [ 81.474565] [<ffffffff81512ea9>] kernel_init+0x9/0xf0 >>> [ 81.480216] [<ffffffff8151f788>] ret_from_fork+0x58/0x90 >>> [ 81.486156] [<ffffffff81512ea0>] ? rest_init+0x80/0x80 >>> [ 81.492350] ---[ end Kernel panic - not syncing: Out of memory and >>> no killable processes... >>> [ 81.492350] >>> >>> -Longman >> I increased the pre-initialized memory per node in update_defer_init() >> of mm/page_alloc.c from 2G to 4G. Now I am able to boot the 24-TB >> machine without error. The 12-TB has 0.75TB/node, while the 24-TB >> machine has 1.5TB/node. I would suggest something like pre-initializing >> 1G per 0.25TB/node. In this way, it will scale properly with the memory >> size. > We're using more than 2G before we've even completed do_basic_setup()? > Where did it all go?
I think they may be used in the allocation of the hash tables like:
[ 2.367440] Dentry cache hash table entries: 2147483648 (order: 22, 17179869184 bytes) [ 11.522768] Inode-cache hash table entries: 2147483648 (order: 22, 17179869184 bytes) [ 18.598513] Mount-cache hash table entries: 67108864 (order: 17, 536870912 bytes) [ 18.667485] Mountpoint-cache hash table entries: 67108864 (order: 17, 536870912 bytes)
The size of those hash tables do scale somewhat linearly with the amount of total memory available.
>> Before the patch, the boot time from elilo prompt to ssh login was 694s. >> After the patch, the boot up time was 346s, a saving of 348s (about 50%). > Having to guesstimate the amount of memory which is needed for a > successful boot will be painful. Any number we choose will be wrong > 99% of the time. > > If the kswapd threads have started, all we need to do is to wait: take > a little nap in the allocator's page==NULL slowpath. > > I'm not seeing any reason why we can't start kswapd much earlier - > right at the start of do_basic_setup()?
I think we can, we just have to change the hash table allocator to do that.
Cheers, Longman
| |