Stick your Fusions in the right slot!
Finally got some breathing space for my first post of the year!
In order to get a major client through Christmas which is traditionally their busiest period of the year (where load normally goes up five-fold), we embarked on a major platform refresh where both the back end SQL Server DBMS was upgraded from SQL 2008R2 to SQL Server 2012 and the hardware was upgraded to 2 x Dell PowerEdge R910s with 4 x 10 (physical) cores and 512GB of RAM in each server.
In order to take the disk I/O bottleneck out of the equation we decided to take traditional spinning disk technology out of the equation so the SAN was kicked out sob and a shiny pair of 2.4TB Fusion-io Duos were added to the server spec sheet.
This would allow us to use AlwaysOn Availability Groups in SQL Server 2012 which have the major upside of allowing the secondaries to be readable instead of being completely offline, which, of course is why having Fusion-io drives in both severs really comes into its own.
But I digress, this post isn’t about boasting about new kit or Fusion-io’s hardware but rather it’s a warning about choosing the right PCIe slot to place the Fusion-io cards because although the R910s have plenty of PCIe slots they don’t all run at the same speed and our main concerns when adding an older 785 GB ioDrive2 Fusion-io card we were recycling from the old kit was that the slot had enough power and there was room for airflow between the shiny new 2.4TBs that were already installed and these cards we were adding in from the older kit.
When the servers were powered up there was nothing untoward showing up in event logs etc, no bells or sirens and the card was initialised and formatted, and could have been used without issue. It was only when the fio-status command was run as a final check (with the -a option) that any mention was made that all was not as it should be, as the the screenshot below shows:
And there you have it, a discrete little “The bandwidth of the PCIe slot is not optimal for the ioMemory.” warning for the newly inserted ioDrive2 card.
We quickly realised that the warning was due to the fact that the newly inserted card was in an x4 slot and as soon as we stuck it in an x8 slot the warning disappeared and we were good to go. The Fusion-io docs do mention in passing an x8 slot to be fair, but it’s easily missed as more focus is placed on the height of the slots, and if you do have a server with multiple speed slots make sure you pick the right slot as there’s no ‘loud’ indication that you’re not getting the best out of it, and the x4 and x8 slots look quite similar so check your server motherboard docs in advance.
I’m not sure what the performance penalty would have been, but with one of the cards using only half the potential bandwidth I didn’t want to find out.
So, if you’re migrating to meaty new servers be aware of the proliferation of multiple speed PCIe slots and make sure you pick the right speed slot, especially when dealing with multiple Fusion-io cards, and if you’re considering new hardware you might want to add a check to ensure there are enough correct speed slots for your needs.