2-node HA DRBD Concurrent local writes

[thheo]

Hello,

Once in a while I was getting this warning in dmesg ( kern.log ):
block drbd0: kvm[433746] Concurrent local write detected! [DISCARD L] new: 623356039s +8192; pending: 623356039s +8192

So I decided to study it more and I was able to trigger message whenever I was running sqlio benchmark tool in Windows 2003 guests.
I run a 2-node HA cluster with drbd as "shared" storage, however this message is about subsequent local (not from the other node) writes in the same location ( or overlapping locations, although in my case they were always identical: new and pending )
I run 6 Windows 2003r2 64-bit guests with virtio driver for disk with version 0-1.49 ( I am getting BSOD if I try to run any newer version ). So I am not sure if
this is a driver problem or something else, but I was able to get rid of these warnings after I activated caching "writethrough" for the virtual disks of the guests (in the proxmox gui).
The guest disks are raw-LVM on top of DRBD ( not using CLVM ). I have played a bit with all the settings ( https://pve.proxmox.com/wiki/Performance_Tweaks#Disk_Cache )
but only writethrough and writeback where the only ones that prevented this behaviour. Maybe somehow with this caching the I/O requests get properly queued. As far as I understood this message means that a subsequent request for the same location arrives before the first one gets completed.

I'll leave it like this for now, but if anyone can clarify what's going on it would be nice to know

 

[whitewater]

Hello thheo,
i have 2 nodes with LSI Raid Card, with BBU.
I haven't got the answser about your question.
I have similar message : Concurrent local write detected!

So, i will test it and give my results.
Do you know if a stop / restart of the VM is required ?

EDIT :
have you got seen update at wiki DRBD page ?
http://pve.proxmox.com/wiki/DRBD

go to "warning" section and the link.
There is interesting informations.

 

[thheo]

Yup.. good finding. As far as I tested with a guest VM only writethrough and writeback avoids those messages. With directsync for example I still have them..
You have to stop the VM so that the new caching kicks in..
Let me know what are your test results..

 

[whitewater]

i will test with shutdown the VM.
It's in production, so... i must find the good moment

I had looking for it in forum, but not find if it's better to disable caching in windows VM with DRBD storage, for data security and performance inpact.
Do you know ?