Messages in this thread | ![/](/images/icornerl.gif) | | Date | Wed, 06 May 2015 13:58:45 -0400 | From | Waiman Long <> | Subject | Re: [PATCH 0/13] Parallel struct page initialisation v4 |
| |
On 05/06/2015 06:22 AM, Mel Gorman wrote: > On Wed, May 06, 2015 at 08:12:46AM +0100, Mel Gorman wrote: >> On Tue, May 05, 2015 at 03:25:49PM -0700, Andrew Morton wrote: >>> On Tue, 5 May 2015 23:13:29 +0100 Mel Gorman<mgorman@suse.de> wrote: >>> >>>>> Alternatively, the page allocator can go off and synchronously >>>>> initialize some pageframes itself. Keep doing that until the >>>>> allocation attempt succeeds. >>>>> >>>> That was rejected during review of earlier attempts at this feature on >>>> the grounds that it impacted allocator fast paths. >>> eh? Changes are only needed on the allocation-attempt-failed path, >>> which is slow-path. >> We'd have to distinguish between falling back to other zones because the >> high zone is artifically exhausted and normal ALLOC_BATCH exhaustion. We'd >> also have to avoid falling back to remote nodes prematurely. While I have >> not tried an implementation, I expected they would need to be in the fast >> paths unless I used jump labels to get around it. I'm going to try altering >> when we initialise instead so that it happens earlier. >> > Which looks as follows. Waiman, a test on the 24TB machine would be > appreciated again. This patch should be applied instead of "mm: meminit: > Take into account that large system caches scale linearly with memory" > > ---8<--- > mm: meminit: Finish initialisation of memory before basic setup > > Waiman Long reported that 24TB machines hit OOM during basic setup when > struct page initialisation was deferred. One approach is to initialise memory > on demand but it interferes with page allocator paths. This patch creates > dedicated threads to initialise memory before basic setup. It then blocks > on a rw_semaphore until completion as a wait_queue and counter is overkill. > This may be slower to boot but it's simplier overall and also gets rid of a > lot of section mangling which existed so kswapd could do the initialisation. > > Signed-off-by: Mel Gorman<mgorman@suse.de> >
This patch moves the deferred meminit from kswapd to its own kernel threads started after smp_init(). However, the hash table allocation was done earlier than that. It seems like it will still run out of memory in the 24TB machine that I tested on.
I will certainly try it out, but I doubt it will solve the problem on its own.
Cheers, Longman
| ![\](/images/icornerr.gif) |