On 10/6/05, Peter Bodik <[EMAIL PROTECTED]> wrote: > > > Does anybody have experience with such setup? Basically, > > the number of > > > hosts would be pretty small, but the number of metrics > > could be large > > > (100s or > > > 1000s) and for some of them we'll need very fast updates. > > > > While it would be nice to have per-process or per-application > > metrics, I believe it was recently stated in an e-mail just > > sent in the last 24 hours about a similar task that it's not > > currently available in Ganglia, although it might be a nice > > feature to put into the libmetrics codes. > > We're planning to add all the application-specific metrics through gmetric. > So the J2EE app would call gmetric every time it wants to update. For > example, I would create a metric 'EJB_componentX_errorrate' for each > componentX and report the value from J2EE. > > So it certainly is possible, but I'm just worried (for example) that some > data might get lost because of such frequent updates or something. > > The user-defined metrics are less-efficient, right? So having 1000s (from > each host) could cause some problems ... I'd like to know what kind of > problems we can expect and what do to about them.
Someone else, here or on ganglia-developer might know about this -- I'm too new to know about such things... I don't believe personally that all the data will get transmitted; it depends on the polling rate of the gmonds and gmetads and web interface in the cluster/system. I have no reason to believe that a good portion won't get through, however. -- ~Mike - Just my two cents - No man is an island, and no man is unable.
