Hi.
Unfortunately, there isn't a good guide for sizing Ganesha. It's pretty
light weight, and so the machines it needs are generally smaller than
what Ceph needs, so you probably won't have much of a problem.
The scaling of Ganesha is in 2 factors, based on the workload involved:
the CPU usage scales with the number of clients, and the memory scales
with the size of the working set of files.
The number of clients is controlled by the RPC_Max_Connections parameter
(default 1024), and the max number of parallel operations is controlled
by the RPC_Ioq_ThrdMax parameter, both in the NFS_CORE_PARAM block. You
probably won't need to change these, unless you have a *lot* of clients.
With these settings, any decent server-class multi-core CPU can keep
up with demand.
Memory usage is controlled by several parameters, controlling 2 separate
caches: the handle cache, and the dirent cache.
The handle cache is global to the machine, and there is one handle per
file/directory/symlink/etc. When this cache is full, entries in it will
start to be re-used, and performance will degrade. This is controlled
by the Entries_HWMark parameter in the MDCACHE block, and defaults to
100,000. This is deliberately sized low, so that small deployments can
be made in containers or on small systems. If you have a large data
set, you will definitely need to raise this. Memory per handle is
fairly small, in the mid 10s of k, so this can be raised a lot. The
handle cache is the largest user of memory on a Ganesha system.
The dirent cache is per-directory, and it makes a very large difference
in directory listing performance. Dirents are stored in chunks of 1000,
and the number of chunks saved per directory is controlled by the
Dir_Chunk parameter in the MDCACHE block. It defaults to 128, which is
again low. If you have large directories that are listed commonly (or
are listed by multiple clients at once) you probably want to raise it.
Dirent memory is generally small, dominated by the size of the filename
in the dirent, but keep in mind the dirent cache is per-directory, so if
you have lots of directories in your working set, this could take a
significant amount of memory.
Ganesha does not do any data caching, only metadata caching, so it's
memory doesn't need to scale with the amount of data being used.
In general, Ganesha is a lightweight daemon, since it's primarily a
translator, and it will use much less resources than the equivalent
CephFS MDS or RGW serving the same workload.
Daniel
On 3/20/21 5:01 AM, Quang Lê wrote:
Hi guys,
I'm using manila-openstack to provide a filesystem service using
backend CEPHFS. My design use nfs-ganesha as the gateway for the VM in
openstack mount to CephFS. I am having problems with sizing the
ganesha-servers.
Can anyone suggest me *what are the hardware requirements of the ganesha
server are* or *what parameters are needed to consider when sizing a
ganesha server* ?
My simple topology in link: *https://i.imgur.com/xrYqxAh.png
<https://i.imgur.com/xrYqxAh.png>*
Thank you guys.
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
_______________________________________________
ceph-users mailing list -- ceph-users@xxxxxxx
To unsubscribe send an email to ceph-users-leave@xxxxxxx
