Open Stack

Hi,

 

I'm research methods to get around high storage latency for some
applications where redundancy does not matter, so using local NVMe
drives in compute nodes seems to be the practical choice.

 

However, there does not appear to be a good solution from what I have
read.  For example, BlockDeviceDriver has been deprecated/removed, LVM
is only supported via iSCSI (which is slow) and localization of LVM
volumes onto the same compute node as VMs is impossible, and other
methods (PCI pass-through, etc.) would require direct access to the
local drives, where device cleansing would need to occur after a device
was removed from a VM, and I don't believe there is a hook for this.

 

Ephemeral storage appears to be an option, but I believe it has the same
issue as PCI pass-through, in that there is no abiilty to automatically
cleanse a device after it has been used.  In our default configuration,
ephemeral storage is redirected to use Ceph, which solves the cleansing
issue, but isn't suitable due to its high latency.  Also, ephemeral
storage appears as a second device, not the root disk, so that
complicates a few configurations we have.

 

Is there any other way to write an operating system image onto a local
drive and boot from it?  Or preferably assign an LVM /dev/mapper path as
a device in libvirt (no iSCSI) after configuring a logical volume?  or
am I missing something?

 

Thanks!


Eric

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack-discuss/attachments/20200805/2c152f3c/attachment-0001.html>