TOPIC:

Hi,

We have a two node HA setup with 2 Dell R430 Servers. The servers have 2 disks in a RAID 1 array which is where Xen Centre 6.5, with SP1 and the two hot fixes, is installed. There are three other disks on the each of the servers in RAID 5 which have been used as a storage repository for the DRBD partition setup by the halizard scripts.

Once the system was setup a Windows 2012 R2 Standard 64-bit machine was setup with SQL 2012 SP2 Standard 32-bit. After this HA was tested and it worked very nicely with the machine bouncing from one server to the other as it should.

Next we used a little test application which inserted data into a SQL 2012 database across the network a row at a time and then read the row back, When this application was running and we simulated a power failure on the active/master server the application stopped and we were able to determine how many rows were entered by determining the id of the last row read. However when the VM started up on the other server and we looked at the DB not all the rows were present and we were missing data. Our test results were as follows:

Time between Inserts	Number of rows inserted by test application	Number of rows present in the DB after a power failure
1 ms	2011	1061
10 ms	1261	1194
5 ms	1422	0
5 ms	5997	5621

The missing data is bit of a concern especially with our third test where everything disappeared. Does anyone know why this could be happening?

We also tried this test on another Dell R430 Server with the similar setup where Xen 6.5 was installed on the first two disks and the storage repository was on the three other disks. However this was a single server and was not part of the pool. With this test when we simulated a power failure we did not lose any data and when the server rebooted and the VM started all the data was there. So the only difference with this test and the previous test was:

1. This test was on a single server not a HA pool
2. This server had a local storage repository and not a DRDB iSCSI repository

The VM used on this setup was exactly the same as the VM on the other setup apart from a few security patches for Windows.

This test was also done on yet another Dell R430 server which had windows directly install on the hardware, so NO Xen. The tests passed on this setup as well.

So at this point it appears that the issue is either with Xen or with how the halizard scripts setup the storage repository. Does anyone have any suggestions on how we could resolve this or what we should try next?

Thanks

Udyant

As a further test we turned off the slave on the HA Pool and carried out the test on the master and it still failed; i.e. the all the rows that were written were not in the DB after power up.

This was done because we thought that DRBD might be an issue. Any thoughts?

Try disabling write cache in targets.conf. Per the TGT man page it appears that there is a small write cache ON by default.

write-cache <val>

"on" or "off", default on.

Hi SC,

Thanks for that suggestion.

Changing the write-cache to off in the /etc/tgt/targets.conf file did the trick and we are no longer experiencing data loss during power fail tests.

Thanks again