[collectd] Graphing NFS statistics
cell at ices.utexas.edu
Fri Nov 18 16:35:03 CET 2005
If recently added a "stack" graph of all the nfs stuff to our graphs.
I would probably reduce it to 6 or 8 items, with the last one being
"other" - a summation of all remaining DSes.
probably also nice to have:
readdir + readdirplus
the rest can probably be lumped into "other".
or, you could lump the DSes by category:
meta ops = access + getattr + setattr ...
file ops = read + write + create + remove + rename ...
dir ops = readdir + readdirplus + mkdir + rmdir ...
misc ops = fsstat + pathconf ...
On Fri, Nov 18, 2005 at 01:41:39PM +0100, Florian Forster wrote:
> Hello everyone,
> I'm trying to add a graph for the NFS statistics to `collection.cgi' and
> need some opinions. I'll talk about NFSv3 here but the problem basically
> applies to versions 2 and 4 as well.
> The NFS-RRD-files have many (22 in this case) DS'es. A graph with 22
> lines is not very readable/informative. So I want to reduce the number
> of lines/DSes actually graphed. I see three possible solutions:
> - Group DSes together (a first try was `read', `write', `create' and
> - Only graph some (i.e. the interesting ones) DSes
> - A combination of both (graph the interesting seperate and one `Other'
> Here's some sample data of my fileserver:
> read: 14242804 63.88% ( 63.88%)
> getattr: 4075875 18.28% ( 82.16%)
> access: 1589911 7.13% ( 89.29%)
> lookup: 980753 4.40% ( 93.69%)
> write: 629690 2.82% ( 96.51%)
> commit: 340567 1.53% ( 98.04%)
> setattr: 143552 0.64% ( 98.68%)
> create: 136865 0.61% ( 99.30%)
> remove: 120388 0.54% ( 99.84%)
> readdirplus: 18813 0.08% ( 99.92%)
> rename: 12000 0.05% ( 99.97%)
> mkdir: 2106 0.01% ( 99.98%)
> symlink: 2069 0.01% ( 99.99%)
> rmdir: 1026 0.00% (100.00%)
> link: 290 0.00% (100.00%)
> readlink: 203 0.00% (100.00%)
> fsinfo: 90 0.00% (100.00%)
> mknod: 32 0.00% (100.00%)
> fsstat: 2 0.00% (100.00%)
> null: 1 0.00% (100.00%)
> readdir: 0 0.00% (100.00%)
> pathconf: 0 0.00% (100.00%)
> As you can see the first three procedures sum up to almost 90% of all
> procedures, the first eight sum up to over 99%. Drawing the other 14 is
> definitely not useful. That's what I think at least..
> Looking at Jason's graphs I see that it's similar with his NFS server: a
> whole bunch of `access' and `getattr' requests, a reasonable amount of
> `read's and `writes' (though nowhere as many reads as my fileserver has
> for whatever reason) and pretty much nothing else..
> Oh, and one more limitation: For the sake of simplicity and because my
> time is very limited at the moment I want to have exactly one graph per
> RRD-file. This is not a hard requirement for the (far) future but right
> now I don't want to spend that much time on this..
> Any suggestions?
> Florian octo Forster
> Hacker in training
> GnuPG: 0x91523C3D
> Collectd mailing list
> Collectd at verplant.org
More information about the Collectd