Re: [Bug 217022] New: Extremely Slow Hugepage Allocation

Re: [Bug 217022] New: Extremely Slow Hugepage Allocation

[Andrew Morton]

(switched to email.  Please respond via emailed reply-to-all, not via the
bugzilla web interface).

On Sat, 11 Feb 2023 06:54:45 +0000 bugzilla-daemon@kernel.org wrote:

> https://bugzilla.kernel.org/show_bug.cgi?id=217022
> 
>             Bug ID: 217022
>            Summary: Extremely Slow Hugepage Allocation
>            Product: Memory Management
>            Version: 2.5
>     Kernel Version: 5.15
>           Hardware: All
>                 OS: Linux
>               Tree: Mainline
>             Status: NEW
>           Severity: normal
>           Priority: P1
>          Component: Page Allocator
>           Assignee: akpm@linux-foundation.org
>           Reporter: y.liu@naruida.com
>         Regression: No
> 
> Created attachment 303713
>   --> https://bugzilla.kernel.org/attachment.cgi?id=303713&action=edit
> perf data and kernel config
> 
> We have some ICE lake server with 1TB memory installed. They were all running
> at 5.10.x branch LTS kernel and run fine. After we upgraded kernel to 5.15.x
> LTS kernel, the booting process were extremely slow. After some analysis, we
> realized that was caused by hugepage allocation. Our system used sysctl.conf to
> allocate hugepage at the boot time. In fact, "echo 960 > nr_hugepages"  had the
> same effect. The only way to do a fast allocation is to use boot cmd option:
> "default_hugepagesz=1G hugepages=960".
> 
> Our System is Xeon W3375 with 1TB memory installed. But this bug also occured
> with Xeon 8180 with 1.5TB memory too. Our OS is Gentoo Linux. With 5.10.x, the
> allocation speed is around 300GB/s, and 5.15.x only had 30GB/s. We also tried
> 6.1.1, it is the same as 5.15.x .
> 
> We compiled 5.10.163 and 5.15.88 with debug option and used "perf -a -g sleep
> 2" to catch kernel functions. Here are two perf outputs. I hope this can help.
> 
> -- 
> You may reply to this email to add a comment.
> 
> You are receiving this mail because:
> You are the assignee for the bug.