Description of problem: The Documentation[1], in section 4.10.2.4 reads: ~~~ Support for huge pages is a new configuration setting for RHV 4.2. Huge pages are pre-allocated when a virtual machine starts to run (dynamic allocation is disabled by default). ~~~ This is partial, there is no mention on how to enable it or use dynamic allocation. I can make it work by: 1. /etc/vdsm/vdsm.conf [performance] use_dynamic_hugepages = true 2. Administration -> Configure -> Scheduling Policies -> Edit -> Disable "Huge Pages" Filter NOTE: without item 2 the engine refuses scheduling the VM on a host without preallocated Huge Pages. It feels like there is missing implementation in the engine to properly support this. If this is supported downstream then it should be fully documented. Otherwise document it as tech preview or not supported.
Maybe even enable this by default, it was just disabled by default because it was not supported by engine yet (https://github.com/oVirt/vdsm/commit/03b7ad91fb2bfc5927dfc4623582e6c280283114)
Sandro, Please see comment 2. Should this setting be enabled by default?
Comment 1 refers to the Virtual Machine Management Guide: https://access.redhat.com/documentation/en-us/red_hat_virtualization/4.3/html/virtual_machine_management_guide/configuring_high_performance_virtual_machines_templates_and_pools#Configuring_Recommended_Manual_Settings
(In reply to Steve Goodman from comment #4) > Comment 1 refers to the Virtual Machine Management Guide: > > https://access.redhat.com/documentation/en-us/red_hat_virtualization/4.3/ > html/virtual_machine_management_guide/ > configuring_high_performance_virtual_machines_templates_and_pools#Configuring > _Recommended_Manual_Settings Oops, indeed. Forgot to add the link to the [1] reference. Thank you.
(In reply to Steve Goodman from comment #3) > Sandro, > > Please see comment 2. Should this setting be enabled by default? Redirecting this question to virt team. Ryan?
Only for high performance VMs, not in general.
@Ryan do I understand this correctly, you want to: 1) enable huge pages by default for high performance VM type 2) change use_dynamic_hugepages in vdsm default to true
Well, no. It's a very specific use case, and the calculation across NUMA nodes can be finicky. I'd like to leave it as it is, thought we do need to document how to enable it
I am not sure if I understood why we can't enable the dynamic hugepages allocation by default, if the VM doesn't use hugepages the setting won't matter - if it uses hugepages it makes using hugepages easier. The engine now supports hugepages so the original argument for disabling it by default is gone, right? Having dynamic hugepages enabled by default also means I can drop it from the filter because that is only relevant for pre-allocated hugepages.
If we plan to make HP default after the related bugs are fixed, a few things to consider: * Default to 2M HPs to avoid contiguous memory allocation issues * Scheduler could use buddyinfo from each host to calculate where the VM could run, and maybe even warn the user about fragmentation
We don't plan to make HP the default
*** Bug 1526754 has been marked as a duplicate of this bug. ***