[
https://issues.apache.org/jira/browse/HDDS-15092?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18089535#comment-18089535
]
Wei-Chiu Chuang commented on HDDS-15092:
----------------------------------------
### Key Findings Summary:
1. Datanodes with heterogeneous disk sizes (Challenge 1) are well-supported
via the default CapacityVolumeChoosingPolicy.java and the active disk
balancer on the
datanodes.
2. Varying datanode disk counts and densities (Challenges 2 & 3) are only
partially supported. While the capacity-based placement policy handles this
using utilization ratios, the rack-aware and rack-scatter placement policies
use uniform/random selection. This leads to unbalanced initial placements
that require remediation by the ContainerBalancer later.
3. Heterogeneous rack densities (Challenge 4) are currently unaddressed in
SCM rack selection logic.
4. Varying network bandwidth or processing power (Challenges 5 & 6) are not
modeled or tracked in the scheduling logic.
5. Selective SSD/HDD placement (Challenge 7) is currently unsupported but is
actively in progress under HDDS-11233 https://issues.apache.
org/jira/browse/HDDS-11233 (Storage Policy Support).
> Heterogeneous deployment
> ------------------------
>
> Key: HDDS-15092
> URL: https://issues.apache.org/jira/browse/HDDS-15092
> Project: Apache Ozone
> Issue Type: Epic
> Reporter: Wei-Chiu Chuang
> Priority: Major
>
> Current Ozone implementation assumes homogenous deployments: all datanodes
> are composed of the same number of disks of the same size.
>
> Heterogeneous deployments bring challenges to Ozone:
> # what if a datanode has disks of different size?
> # what if datanodes have different number of disks?
> # what if a rack has datanodes of different density?
> # what if racks are of different storage density?
> # what if datanodes/racks have different network bandwidth?
> # what if datanodes have different processing power?
> # what if some datanodes have SSDs and some have HDD? (storage policy)
>
> The correctness, edge case handling and performance consideration are not
> thought carefully.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
