- Data ONTAP 8 7-Mode
- Clustered ONTAP 8.x
- ONTAP 9
|Warning: This article applies to hard drive disk (HDD) aggregates, but reallocate must not be done to a SSD aggregate, FabricPool, or in Cloud Volumes ONTAP (CVO) aggregates. For CVO please create a new aggregate, volume move the existing volumes and destroy the existing aggregate once vacated.|
For best performance, it is advisable to add a new RAID group of equal size to existing RAID groups. If a new RAID group cannot be added, then at minimum, three or more disks should be added at the same time to an existing RAID group. This allows the storage system to write new data across multiple disks.
Also, a reallocate must be done to even out the RAID stripe layout across the aggregate, otherwise all new writes will go to the new disk.
How should a reallocate be done?Forced reallocation ignores the optimization thresholds and completely rewrites the data to disk, unlike the normal reallocation process. Although this improves the layout, routine use of
reallocate -fis not a best practice, except when adding disks as we know the layout is uneven. Also, because all of the data is optimized, forced reallocation cannot be run against volumes that have existing Snapshot copies unless the physical reallocation method (-p) is also used.
reallocate -f -o -p /vol/volname
- Clustered ONTAP/ONTAP 9:
reallocate start -vserver svm0 -path /vol/vol1 -once true -space-optimized true -no-check true
Reallocate can be done on an aggregate level, but often times it is costly on disk cycles and takes several days or weeks. Instead, a more measured approach is to do the busiest volume first, working to the quietest in the aggregate. One job at a time may be ran, and if performance overhead exists, a second may be added.
Note: It is best to look at Active IQ Unififed Manager under Aggregates, then performance, and Nodes under performance, to determine the quietest times such as after 5 PM or on weekends. Reallocate will cause additional overhead so this must be accounted for. It's estimated to cause between 10-30% of a performance overhead per job, but this is an estimate and could easily take more or less.