As the regular author of this blog I wanted to provide a short introduction for a guest contributor Over the past few months I have been transitioning to new challenges so my day to day work on VMware and IBM storage has become limited. A new technical expert has taken on the role of author VMware papers and best practices and he has graciously agreed to write about some of his recent work. Thank you Jeremy Canady for the great contributions. – Rawley Burbridge
At this point, vSphere Storage APIs for Array Integration (VAAI) is old hat for most VMware administrators. In fact this blog has already written about VAAI in a previous post and white paper back in 2011. Sense that time there has been multiple releases of vSphere and support for various VAAI primitives continues to increase among storage systems for both the Block and NAS primitives. With the 184.108.40.206 code release for the Storwize family we decided to revisit the topic using current vSphere versions and the new 220.127.116.11 release.
I have no intention of diving into all the details of VAAI as there are many very good resources online as well as the full white paper. What I would like to highlight in this post is the testing and results of the Atomic Test and Test (ATS) primitive.
Atomic Test and Set
As you are likely aware, VMFS is a clustered file system that allows multiple hosts to simultaneously access the data located on it. The key to a clustered file system is handling conflicting requests from multiple hosts. VMFS utilizes file level locking to prevent conflicting access to the same file. The locking information is stored in the VMFS metadata which also must be protected from conflicting access. To provide locking for the metadata, VMFS originally only relied upon a SCSI reservation for the full VMFS volume. The SCSI reservation locks the complete LUN and prevents access from other hosts. This results in actions such as snapshot creation or virtual machine power on temporarily locking the complete VMFS datastore. This locking mechanism poses performance and scalability issues.
Atomic Test and Set (ATS) provides an alternative to the SCSI reservation method of locking. Instead of locking the complete LUN with a SCSI reservation, ATS allows the host to lock a single sector that contains the metadata it would like to update. This results in only conflicting access being prevented.
With the update to the VAAI paper we wanted to show the actual benefits of ATS with the Storwize family. To do so, two tests were devised, one with artificially generated worst case locking and one a with more of a real world configuration. The setup consisted of two IBM Flex System x240 compute nodes running ESXi 5.1 sharing a single VMFS datastore provided via FC from a Storwize V7000. Each test was run with the locking load and ATS disabled, without the locking load, and with the locking load and ATS enabled.
To generate worst case locking we needed a way to quickly generate metadata updates on the VMFS file system. To do this a simple bash script was created that continually executed the touch command upon a file located on the VMFS datastore. The touch command updates the time stamp of the file resulting in a temporary metadata lock each time it is ran.
To measure the impact, a VMDK located on the shared VMFS datastore was attached to a VM that had IOMeter loaded. A simple 4 KB sequential read workload was placed on the VMDK and measured for each run. The results can be seen below.
As you can see without ATS there is a severe impact upon the disk performance. When enabling ATS the performance increased by over 400% and equaled the performance when no locking workload was applied. This is to be expected as the artificially generated locking should never conflict with accessing the VMDK when ATS is enabled.
To measure the severity of the locking esxtop was utilized to monitor and measure the number of conflicts per second. As you can see, ATS all but eliminated the conflicts.
Real World Backup Snapshot Workload
Backup processes need to be quick and have the smallest impact possible. In a VMware environment most backup processes require a snapshot of the virtual machine to be created and deleted. Snapshot creation and deletion require metadata updates and when scaled out can cause issues on heavily populated datastores.
To test the effects of ATS upon large scale snapshot creation and deletion, a VMDK was located on a VMFS volume that contained ten additional virtual machines. The VMDK was attached to a VM running Iometer with a simple 4KB sequentail workload applied to the VMDK. A script was created to continually create and remove snapshots from the ten running VMs. The resulting Iometer performance was monitored while the snapshot workload was applied. The results can be seen below.
As you can see there was a sizable impact on performance with ATS disabled. Additionally notice that the performance with ATS enabled is nearly identical to the performance when no snapshot workload was running.
To measure the severity of the locking esxtop was utilized to measure the conflicts per second. As you can see below, the number of conflicts was significantly reduced.
I hope these simple tests show the huge improvement that ATS provides to the scaling of a single VMFS datastore.
A complete white paper on this topic as well as the other supported VAAI primitives is available here.