在 2021/10/29 下午9:38, Michal Hocko 写道:
On Thu 28-10-21 19:56:49, Ning Zhang wrote:
As we know, thp may lead to memory bloat which may cause OOM.
Through testing with some apps, we found that the reason of
memory bloat is a huge page may contain some zero subpages
(may accessed or not). And we found that most zero subpages
are centralized in a few huge pages.
Following is a text_classification_rnn case for tensorflow:
zero_subpages huge_pages waste
[ 0, 1) 186 0.00%
[ 1, 2) 23 0.01%
[ 2, 4) 36 0.02%
[ 4, 8) 67 0.08%
[ 8, 16) 80 0.23%
[ 16, 32) 109 0.61%
[ 32, 64) 44 0.49%
[ 64, 128) 12 0.30%
[ 128, 256) 28 1.54%
[ 256, 513) 159 18.03%
In the case, there are 187 huge pages (25% of the total huge pages)
which contain more then 128 zero subpages. And these huge pages
lead to 19.57% waste of the total rss. It means we can reclaim
19.57% memory by splitting the 187 huge pages and reclaiming the
zero subpages.
What is the THP policy configuration in your testing? I assume you are
using defaults right? That would be always for THP and madvise for
defrag. Would it make more sense to use madvise mode for THP for your
workload? The THP code is rather complex and just by looking at the
diffstat this add quite a lot on top. Is this really worth it?
The THP configuration is always.
Madvise needs users to set MADV_HUGEPAGE by themselves if they want use
huge page, while many users don't do set this, and they can't control
this well.
Such as java, users can set heap and metaspace to use huge pages with
madvise, but there is also memory bloat. Users still need to test
whether their app can accept the waste.
For the case above, if we set THP configuration to be madvise, all the
pages it uses will be 4K-page.
Memory bloat is one of the most important reasons that users disable
THP. We do this to popularize THP to be default enabled.
