From: Peter P. <pet...@in...> - 2010-10-26 15:13:24
|
Bernard, Thanks for committing the patch. Hopefully more people will try the sFlow functionality and we can get additional feedback. On Oct 26, 2010, at 12:47 AM, Bernard Li wrote: > If I could make just one comment: right now if I run gmond in debug > mode, it shows that the sFlow metrics are actually spoofed gmond > metrics. Is there a way to make it so they are labelled sFlow > metrics? The metrics really are exactly equivalent to gmond native metrics, the sFlow standard was based in part on gmond's libmetrics and the Host sFlow agent leverages code from libmetrics. By spoofing them as gmond metrics, they seamlessly integrate in the UI and are combined in charts and rolled up correctly. There is a broader question relating enhancing the metadata attributes to be more expressive, for instance to be able to express the containment relationship between virtual machine metrics and the host physical machine. The sFlow module is currently discarding virtual machine statistics (based on libvirt) because it isn't clear what metadata to associate with them to ensure that they would appear in the UI in a useful way. http://www.mail-archive.com/gan...@li.../msg06009.html This would seem to tie in to the effort to re-work the Ganglia UI. Additional metadata associated with the metrics would allow the UI to be more intelligent about combining and grouping metrics. > > Perhaps we should start a separate thread to discuss Kostas' idea of > having a pluggable interface for metric sources. With sFlow support, > gmond can now handle both gmond sources and sFlow sources. With a > pluggable interface, we could in the future be able to aggregate data > collected by other tools, such as collectd, collectl, etc. The > possibilities are endless. The gmond_internal.h file contains some of the gmond functions that would need to be exposed to a module in order to support the sFlow functionality. I am not very familiar with collectd and collectl, but it appears that gmond would be polling remote statistics repositories in these cases. The sFlow case is different, the metrics are asynchronously pushed by the sFlow agents to to gmond, so the code has to site in the UDP packet processing loop. Regards, Peter |