[Mochi-devel] An overview of application usage scenarios

[Srinivasan Ramesh]

Hi Phil,

I have been going through your breadcrumb implementation, and I believe 
I understand it well enough to ask a couple of questions:

1. As I understand it, a breadcrumb represents a "call path" in 
traditional HPC profiling terminology, right? A->B is a different 
breadcrumb from C->B. It also appears
that the provider ID is somehow "baked in" to the breadcrumb. Meaning, 
A(provider 1) -> B is a different breadcrumb than A(provider 2)-> B. 
Could you kindly confirm if my understanding is right?
2. If my understanding is accurate, this means that breadcrumb profiling 
provides the COMPLETE profile w.r.t RPC calls (high-level information); 
albeit with TIMESTAMP information missing because you are aggregrating 
information. An idea to enrich this information may be to add some rich 
metadata that will add dimensions to the performance analysis.

As we had discussed in the meeting, combining this information with a 
trace of the same measurement and analyzing (visualizing) them 
side-by-side may be a good first step. This would be in accordance with 
how HPC application performance is typically analyzed.

Going further, I think it may be interesting to ask you (the group) the 
following questions that will help me with the design of the performance 
analysis setup:

====> What exactly are you looking to gain from performance analysis?
       In other words, is there any specific piece of information or 
anomaly that you hope to detect? Or is this more of a general 
performance analysis requirement? I think the answer to that question
is not clear to me entirely.
       ====> The more varied the responses, the better! It will help me 
gather an overall picture of the requirements.


I would love to hear everyone's thoughts on this.

Regards,
-- 
Srinivasan Ramesh