During file related activities on a VMFS datastore, the ESX/ESXi host acquires a lock on a series of on-disk resource metadata structures to either allocate file blocks to a file or free file blocks while deleting the file from VMFS datastore. After it completes the resource metadata structure related operations, the ESX/ESXi host continues to retain the lock for a short period of time in case the host gets further requests for allocations. In an ESX/ESXi cluster where multiple hosts are actively working with the same VMFS datastore, this lock from one host can delay operations from another.
Under most conditions the delay is unnoticed operationally. However, the chance of this delay becoming apparent to user level operations increases under these conditions:
-
Large number of hosts (>10) in a cluster attempting to allocate files on VMFS datastores with low % free space (< 10%).
-
Power-off or power-on operations on many virtual machines in a short period of time.
-
vMotion of virtual machines that involves VMFS datastores with low % free space (< 10%).
-
High activity on disk trees that contain a lot of linked clones.
-
High rate of growth with thin disk provisioning.