Phil Dibowitz wrote:
Ian Cunningham wrote:
Mr. Dibowitz,
I am unclear as to why you have "all of the boxes...unicast their data
to the ganglia server". Unless you want data being duplicated, this is a
bad idea. I am not sure what you mean by classes or "ganglia setup is a
cluster". I think you should probably set it up with what you call a
class be a cluster instead. You can have multiple clusters under one
gmetad. The gmetad.conf on the 'ganglia server' should look like this IMHO:
data_source "foo" foo1 foo2
data_source "bar" bar1 bar2
The reason they also unicast to the server is to have a 'catch all' - in the
event that a data_source isn't setup for a given class, *every* machine gets
kicked with a default to send to the ganglia server, so we have that data.
I realize the data is then in two data sources, and that's duplication, but
I don't see that as such a big deal. I've tried taking it out but it doesn't
change the problem.
Again, this concept you have of classes is confusing the issue for me.
Each data_source is a cluster, one cluster, only that cluster. Putting
in two data_source lines for the same cluster shouldn't work as far as I
know.
It sounds like you want to have clusters of clusters, and to do that you
need to use grids which is a slightly more complicated ball of wax.
We have 10 data sources within each cluster - that's it. We also use grids
to combine all our clusters, but I've disabled all that since a given
cluster doesn't even work.
The point is, as I understand it when gmetad is configured as cluster FOO,
and gets data for host BAR under data_source BAZ it should write that data
to two places:
FOO/BAR/
BAZ/BAR/
But it's not, it's only writing it to one. Any ideas? As far as I can tell,
I've hit some obscure bug - gmetad has the data, but isn't writing it out...
Each data source is a separate cluster. You should not have 10
data_sources for a cluster.
The gmetad does not represent a single cluster, it is more akin to a
grid and can be used represent one. To make a gmetad represent a grid,
use the gridname tag in gmetad.conf. You can have 'grids of grids' or
'grids of clusters' or even 'grids of clusters and grids' (i have such a
setup). To accomplish grids of grids, the higher (closer to the root)
gmetad polls the lower level gmetad as a data_source.
top_level_gmetad.conf:
data_source "LOWER GRID" lower_gmetad_host:8651
I am not sure if your goal is to duplicate data from one "class" in two
different clusters, or if you want to have multiple grids of 10 clusters
each. I'd guess that whatever it is you want to do is possible, but
would I need to know more about what you want to see at the end to
suggest a solution.
Cheers,
Ian
