Andrew,
I pretty much agree with you on all these points. More specific answers
below.
On May 8, 2005, at 12:15 AM, Andrew Sapronov wrote:
On Sun, 2005-02-20 at 07:54 -0800, Federico Sacerdoti wrote:
On Feb 19, 2005, at 10:18 PM, Andrew Sapronov wrote:
$telnet localhost 8652
/
will return a last dump of information.
Can I recieve all data for a certain period of time?
For example for the last hour or for december, 2004.
Where can I get information about all possible parameters.
You are going straight for the jugular! We have only a lossy method of
tracking metrics, so retrieving high-fidelity data for an hour long
ago
is not possible.
I see, such restriction is applied because of RRD usage. Then what do
you think for example about the filter
/my_cluster/some_host?filter=primary
which will return high-fidelity data for an hour step-by-step.
In addition i can change primary data period (rrd_helpers.c).
The filters only work on LIVE data (the latest that gmetad has), not on
history data. As you can see from the code, the filters do not query
the RRD databases at all. They are just a filter so you get some
portion of the tree. They could also do some transformation on the tree
before sending it - compute the mean, median, stddev of a metric type
for example.
However in the current design filters are ill-equipped to give
time-series data. This is due to the lossy RRD databases available in
Ganglia at the time.
I tried to make this type of filter (vua using rrd_fetch), but you see
the gmetad design doesn't allow to add your own filters simply.
Could you advise me something?
I believe there is some work on using a real SQL
database to store history, however.
I suppose RRD is quite enough, but still it would be a good feature to
change RDD parameters run-time not compile-time
and perhaps different parameters for different metrics.
I believe Matt has already done some work on this.
The parameters to gmetad's port 8652 (interactive) are very simple:
Possibilities
1. Give a path name:
/Alpha
will return the ganglia tree of the cluster named "Alpha" if there is
one.
/Alpha/compute-0-0
Will return the ganglia tree of a host 'compute-0-0' in cluster Alpha
if there is one.
You mentioned nothing about metrics.
Is this feature planned to be removed?
/Alpha/compute-0-0/some_metric
It works.
This works because of the recursive nature of the code: any child hash
table maps to another /element/of/the/path. Glad to hear you have a use
for it in mind.
I was discouraged when recieved the metrics and saw what part of the
message "<METRIC..." takes in contrast to (comparing with) DTD.
Why DTD is not used in the form of: <!DOCTYPE configuration SYSTEM
"URL">. Besides this way is recommended.
So large DTD invalidates the expediency of separate metric request.
You are right. We should think about moving the DTD retrieval to
another URL rather than tacking it on to each response.
# nc localhost 8652
/?filter=DTD
or similar.
Andrew Sapronov J.
Federico
Rocks Cluster Group, San Diego Supercomputer Center, CA
