After reading a recent blog post by Duncan Epping, SIOC, tying up some loose ends, I decided to explore whether or not VMware's Storage I/O Control feature actually requires an Enterprise Plus license and vCenter Server. To be completely honest, Duncan's article got me thinking but it was also my recent experience with VMware's vsish and the blog post I wrote What is VMware vsish? that made me think this might be a possibility. vsish is only available on ESXi 4.1 within the Tech Support Mode, but if you have access to debugging rpm from VMware, you can also obtain vsish for classic ESX.
Within vsish, there is a storage section and within that section there is a devices sub-section which provides information regarding your storage devices that includes paths, partitions, IO statistics, queue depth and new SIOC state information.
Here is an example of the various devices that I can view on an ESXi 4.1 host:
~ # vsish
/> ls /storage/scsifw/devices/
Here is an example of various properties accessible to a given storage device:
/> ls /storage/scsifw/devices/t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4/
In particular, we are interested in iormState and you can see the value by just using the cat command:
/> cat /storage/scsifw/devices/t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4/iormState
This value may not mean a whole lot and I have seen this as the default value when SIOC is disabled as well as 2000 from my limited set of tests. Now, since we can access these particular SIOC parameter, I wanted to see how this value was affected when SIOC is enabled and disabled. To test this, I used the following VMware KB1022091 to enabling additional SIOC logging which goes directly to /var/log/messages with the logger tag "storageRM", this allows you to easily filter out SIOC logs via simple grep.
For testing purposes, you can just enable the logging to level 2 which is more than sufficient to get the necessary output. You will perform the following command to change the default SIOC logging level from 0 to 2 using Tech Support Mode:
~ # esxcfg-advcfg -s 2 /Misc/SIOControlLogLevel
Value of SIOControlLoglevel is 2
Now you will want to open a separate SSH session to your ESXi host and tail /var/log/messages to monitor the SIOC logs:
~ # tail -f /var/log/messages | grep storageRM
Oct 10 18:39:05 storageRM: Number of devices on host = 3
Oct 10 18:39:05 storageRM: Checked device t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4 iormEnabled= 0 LatThreshold =30
Oct 10 18:39:05 storageRM: Checked device mpx.vmhba1:C0:T0:L0 iormEnabled= 0 LatThreshold =232
Oct 10 18:39:05 storageRM: Checked device mpx.vmhba32:C0:T0:L0 iormEnabled= 0 LatThreshold =30
Oct 10 18:39:05 storageRM: rateControl: Current log level: 2, new: 2
Oct 10 18:39:05 storageRM: rateControl: Alas - No device with IORM enabled!
You should see something similar to the above. In my lab, it is seeing an iSCSI volume, local storage and CD-ROM and you will notice there is an iormEnabled flag and all three has SIOC disabled and the default latency threshold specified by LatThreshold, which is 30ms by default.
Now we know what these values are when SIOC is disabled, let's see what happens when we enable SIOC from vCenter on this ESXi 4.1 host. I am using evaluation license for the host which supports the Storage I/O Control from a licensing perspective. After enabling SIOC and using the default 30ms on my iSCSI volume, I took a look at the SIOC logs and saw that there were some changes in the logs:
~ # tail -f /var/log/messages | grep storageRM
Oct 10 18:48:56 storageRM: Number of devices on host = 3
Oct 10 18:48:56 storageRM: Checked device t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4 iormEnabled= 1 LatThreshold =30
Oct 10 18:48:56 storageRM: Found device t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4 with datastore openfiler-iSCSI-1
Oct 10 18:48:56 storageRM: Adding device t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4 with datastore openfiler-iSCSI-1
As you can see now, SIOC is enabled and the iormEnabled flag has changed from 0 to 1. This should not be a surprise, now let's take a look at vsish storage property and see if that has changed:
~ # vsish -e get /storage/scsifw/devices/t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4/iormState
If you recall from the previous command above, the default value was 1596 and after enabling SIOC, the value has incremented by one. I found this to be an interesting observation and I tried a few other configurations including enabling SIOC on local storage and found that this value was always incremented by 1 if SIOC was enabled and decrement or kept the same if SIOC is disabled.
As you may or may not know, SIOC does not use vCenter, it is only required when enabling the feature and from this simple test, this looks to be the case. It is also important to note, as pointed out by Duncan in his blog post that the latency statistics is stored in .iormstats.sf file which is stored within each of the VMFS datastores that has SIOC enabled. Putting all this together, I hypothesize that Storage I/O Control could actually be enabled without an Enterprise Plus license and without vCenter Server.
The test utilized the following configuration:
- 2 x virtual ESXi 4.1 hosts licensed as free ESXi (vSphere 4.1 Hypervisor)
- 1 x 50GB iSCSI volume exported from Openfiler VM
- 2 x CentOS VM installed with iozone to generate some IO
Here is a screenshot of the two ESXi 4.1 free licensed hosts displaying the licensed version and each running CentOS VM residing on this shared VMFS iSCSI volume:
I configured vm1 which resides on esxi4-1 and set the disk shares to Low with default value of 500 and configured vm2 which resides on esxi4-4 and set the disk shares to High with default value of 2000:
I then ran iozone a filesystem benchmark tool on both CentOSes VM which is generating some amount of IO on the single VMFS volume shared by both ESXi 4.1 hosts:
I then view the SIOC logs on both ESXi 4.1 in /var/log/vmkernel and tail the IO statistics for the VMFS iSCSI datastore using vsish:
Note: The gray bar tells you which host you are viewing the data for which is being displayed by using GNU Screen. The first two screen displays the current status of SIOC which is currently disabled and the second two screen displays the current device queue depth which in my lab environment defaulted to 128, by default it should be 32 as I remember.
Now I enable SIOC on both ESXi 4.1 hosts using vsish and perform the command:
~ # vsish -e set /storage/scsifw/devices/t10.F405E46494C45400A50555567414D2D443E6A7D276D6F6E4/iormState 1597
Note: Remember to run a "get" operation to check the default value and you will just need to increment by one to enable SIOC. From my testing, the default value will either be 1596 or 2000 and you will change it to 1597 or 2001 respectively.
You can now validate that SIOC is enabled by going back to your SSH session and verify the logs:
As you can see, the iormEnabled flag has now changed from 0 to 1, which means SIOC is now enabled.
If you have running virtual machines on the VMFS volume and SIOC is enabled, you should now see a iormstats.sf latency file stored in the VMFS volume:
After awhile, you can view IO statistics via vsish to see what the device queue is currently configured to and slowly see the throttle based on latency. For this particular snapshot, you can see the vm1 was configured with "high" disk share and vm2 was configured with "low" disk share and there is a large queue depth for the very bottom ESXi host versus the other host which only has a smaller queue depth.
Note: During my test I did notice that the queue depth was dramatically decreased from 128 and even from 32 to single digits. I am pretty sure it was due to the limited amount of resources in my lab that some of these numbers were a little odd.
To summarize, it seems that you can actually enable Storage I/O Control without the use of vCenter Server and an Enterprise Plus license, however, this requires the use of vsish which is only found on ESXi 4.1 and not on classic ESX 4.1. I also found that if you enabled SIOC via this method and joined your host to vCenter Server, it is not aware of these changes and marks SIOC as disabled even though the host actually has SIOC enabled. If you want vCenter to see the update, you will need to enabled SIOC via vCenter.
I would also like to thank Raphael Schitz for helping me validate some of my initial findings.
Update: I also found hidden Storage I/O Control API method called ConfigureDatastoreIORMOnHost which allows you to enable SIOC directly on the host, which validates the claim from above that this can be done directly on an ESX or ESXi host.