Micron Confidential Hi Jeff and Jens, About +ACI-FAN+AF8-MODIFY fsnotify watch set on /dev+ACI-. Was using Fedora34 distro (with 6.3.9 kernel), and fio. Without any particular/specific setting. I tried to see what could watch /dev but failed at that. I used the inotify-info tool, but that display watchers using the inotify interface. And nothing was watching /dev via inotify. Need to figure out how to do the same but for the fanotify interface. I'll look at it again and let you know. Regards, Pierre Micron Confidential +AD4- -----Original Message----- +AD4- From: Jens Axboe +ADw-axboe+AEA-kernel.dk+AD4- +AD4- Sent: Tuesday, August 8, 2023 2:41 PM +AD4- To: Jeff Moyer +ADw-jmoyer+AEA-redhat.com+AD4AOw- Pierre Labat +ADw-plabat+AEA-micron.com+AD4- +AD4- Cc: 'io-uring+AEA-vger.kernel.org' +ADw-io-uring+AEA-vger.kernel.org+AD4- +AD4- Subject: +AFs-EXT+AF0- Re: FYI, fsnotify contention with aio and io+AF8-uring. +AD4- +AD4- CAUTION: EXTERNAL EMAIL. Do not click links or open attachments unless you +AD4- recognize the sender and were expecting this message. +AD4- +AD4- +AD4- On 8/7/23 2:11?PM, Jeff Moyer wrote: +AD4- +AD4- Hi, Pierre, +AD4- +AD4- +AD4- +AD4- Pierre Labat +ADw-plabat+AEA-micron.com+AD4- writes: +AD4- +AD4- +AD4- +AD4APg- Hi, +AD4- +AD4APg- +AD4- +AD4APg- This is FYI, may be you already knows about that, but in case you +AD4- don't.... +AD4- +AD4APg- +AD4- +AD4APg- I was pushing the limit of the number of nvme read IOPS, the FIO +- +AD4- +AD4APg- the Linux OS can handle. For that, I have something special under the +AD4- +AD4APg- Linux nvme driver. As a consequence I am not limited by whatever the +AD4- +AD4APg- NVME SSD max IOPS or IO latency would be. +AD4- +AD4APg- +AD4- +AD4APg- As I cranked the number of system cores and FIO jobs doing direct 4k +AD4- +AD4APg- random read on /dev/nvme0n1, I hit a wall. The IOPS scaling slows +AD4- +AD4APg- (less than linear) and around 15 FIO jobs on 15 core threads, the +AD4- +AD4APg- overall IOPS, in fact, goes down as I add more FIO jobs. For example +AD4- +AD4APg- on a system with 24 cores/48 threads, when I goes beyond 15 FIO jobs, +AD4- +AD4APg- the overall IOPS starts to go down. +AD4- +AD4APg- +AD4- +AD4APg- This happens the same for io+AF8-uring and aio. Was using kernel version +AD4- 6.3.9. Using one namespace (/dev/nvme0n1). +AD4- +AD4- +AD4- +AD4- +AFs-snip+AF0- +AD4- +AD4- +AD4- +AD4APg- As you can see 76+ACU- of the cpu on the box is sucked up by +AD4- +AD4APg- lockref+AF8-get+AF8-not+AF8-zero() and lockref+AF8-put+AF8-return(). Looking at the +AD4- +AD4APg- code, there is contention when IO+AF8-uring call fsnotify+AF8-access(). +AD4- +AD4- +AD4- +AD4- Is there a FAN+AF8-MODIFY fsnotify watch set on /dev? If so, it might be +AD4- +AD4- a good idea to find out what set it and why. +AD4- +AD4- This would be my guess too, some distros do seem to do that. The +AD4- notification bits scale horribly, nobody should use it for anything high +AD4- performance... +AD4- +AD4- -- +AD4- Jens Axboe
