[RFC] doc/spec: document host requirements

doc/spec/host-requirements.md

@@ -0,0 +1,126 @@
+# Host Requirement Assessment
+
+This document describes the basic hardware requirements for a host to
+join / deploy a cluster.
+
+## Problem Description
+
+Aquarium's Host Management should assess whether a node meets the
+requirements for joining an existing cluster, or to deploy an initial
+seed cluster, providing immediate feedback to the user on its
+feasibility.
+
+## Proposed Changes
+
+Considering that the requirements depened on a cluster purpose
+(production, testing, educational) it seems we have two provide at
+least two profiles: "minimal" requirements (enough for Ceph to work),
+and "optimal" requirements (to meet user expectations).
+
+In future we might need to have several "optimal" profiles for
+different use cases.
+
+## Minimal profile
+
+It is based on Ceph upstream minimum hardware recommendations [^1].
+
+### ceph-osd
+
+- Processor: 1+ core per daemon
+- RAM: 4GB+ per daemon (2GB will probably work, but will require
+  changing osd_memory_target)
+- Volume Storage: 1x 20GB+ storage drive per daemon
+- DB/WAL: optional, 1x 10Gb+ SSD partition per daemon
+- Network: 1x 1GbE+ NICs
+
+### ceph-mon, ceph-mgr
+
+- Processor: 2+ cores per daemon
+- RAM: 2GB+ per daemon
+- Disk Space: may be the OS disk, 60GB+ per daemon
+- Network: 1x 1GbE+ NICs
+
+### ceph-mds
+
+- Processor: 2+ cores per daemon
+- RAM: 2GB+ per daemon
+- Disk Space: may be the OS disk, no special requirements
+- Network: 1x 1GbE+ NICs
+
+## Optimal profile
+
+It is based both on Ceph upstream hardware recommendations [^2] and SES
+hardware requirements and recommendations [^3].
+
+### ceph-osd
+
+- Processor: 4+ core per daemon (1x 2GHz CPU Thread per spinner, 2x
+  2GHz CPU Thread per SSD, 4x 2GHz CPU Thread per NVMe)
+- RAM: 5GB+ (osd_memory_target + 1GB) per daemon + 16GB+ for OS
+- Volume Storage: 1x 1TB+ storage drive per daemon
+- DB/WAL: 1x SSD/NVMe 200G+ partition per daemon, not more than six
+  WAL/DB partitions on the same SSD disk and 12 on NVMe disks. Not
+  needed if the volume storage is SSD.
+- Network: two bonded 25 GbE (or faster) network interfaces bonded
+  using 802.3ad (LACP)
+
+### ceph-mon, ceph-mgr
+
+- Processor: 4+ cores per daemon
+- RAM: 4GB+ per daemon + 16GB+ for OS
+- Disk Space: 500GB+ SSD (RAID?) disk/partition for mon db,
+  no requirements for ceph-mgr
+- Network: two bonded 25 GbE (or faster) network interfaces bonded
+  using 802.3ad (LACP)
+
+### ceph-mds
+
+- Processor: 4+ cores per daemon
+- RAM: 16GB+ per daemon (for caching) + 16GB+ for OS
+- Disk Space: may be the OS disk, no special requirements
+- Network: two bonded 25 GbE (or faster) network interfaces bonded
+  using 802.3ad (LACP)
+
+## Additional considerations
+
+### Solid State Drives
+
+Some drives are known to behave badly with Ceph. Should we have a
+blacklist for optimal profile?
+
+### Controllers
+
+A controller may be a bottleneck. Should we check number of disks per
+controller? Topology?
+
+### Write Caches
+
+Usual recommendation is to disable write cache. Should we do this
+automatically?
+
+### Networks
+
+Make separate public and cluster network mandatory for optimal
+profile? Check for jumbo frames enabled and test network for "large"
+package (jumbo frames misconfiguration is a common problem)?
+
+### "Asymmetric" configurations
+
+Mixing hosts with different number of osd disks or disks of different
+size is usually a bad idea. Should we allow only "symmetric"
+configurations for optimal profile?
+
+### Failure Domains
+
+Should we check the minimum "host" failure domain for the optimal
+profile? Number of nodes (at least 4 physical OSD nodes, with 8? OSD
+disks each)?
+
+### Log storage
+
+Should we check storage for logs?
+
+
+[^1]: https://docs.ceph.com/en/latest/start/hardware-recommendations/#minimum-hardware-recommendations
+[^2]: https://docs.ceph.com/en/latest/start/hardware-recommendations/
+[^3]: https://documentation.suse.com/ses/7/html/ses-all/storage-bp-hwreq.html