I know this is very late, but I'm just responding to the philosophical
issues, rather than the technical ones ... 8^)>
On Thu, 13 Apr 2000 Cameron_C_Caffee@xxxxxxxxxxxxxxxxxx wrote:
> ...
> (3) Axis units - the user could provide a units label for each axis.
The PCP metadata already provides all the information a tool needs to
construct the units label automagically.
> Archive analysis :
>
> ... There's
> nothing wrong with having one tool for real time monitoring and another for
> archive analysis since the goals for each are a bit different. The former has
> the character of an "alert" or "alarm" while the latter is normally oriented
> toward historical trending (e.g. CPU Utilization trend for last 6 months).
I think this line of argument takes you down a rat hole.
In the PCP APIs and in particular the way the archive library support
is implemented in libpcp, there is a very determined effort to make
retrospective and real-time sources of performance metrics semantically
equivalent. The rationale is that many tools, and more importantly the
users of the tools, are best served by operating on the abstraction of
a series of observations over time.
The alert or alarm function is just as useful in historical data as it
is in real-time data, but for different purposes. We use alarming in
real-time for operational management, we use alarming against archive
data for analysis, exception reporting and alarm tuning.
One of the strong points of the successful PCP monitoring tools,
including pmie in the open source release, is that they _do_ operate on
both real-time and historical data. And in fact one of the most severe
criticisms of the pmgagets tool (a 2-D visible alarm constructor) is
that it cannot replay from PCP archives.
|