From: chariya@verizon.not   
      
   baylor wrote:   
   > i have what i consider to be a pretty simple process to debug   
   > performance problems in our IT systems, something our company deals   
   > with a lot. First you divide your system into a couple of high level   
   > pieces that you can measure (time) (normally Web server, app server,   
   > database server and network). You specify the order they're in, what   
   > you consider normal timings to be and the actual times.  The goal is to   
   > isolate the problem. So let's say your app has 4 pieces and one is   
   > significantly over threshhold. We've isolated the problem to that   
   > component (let's say it's the app server). You then dive down into it   
   > by repeating the process - break that piece into a handful of   
   > components, time them, compare times against benchmarks, pick the one   
   > bad piece (so far it's always been one piece) and continue until the   
   > user gives up or has found the problem. i'm not worried about solving   
   > the problem yet, just diagnosing it (walking them through the data   
   > gathering steps)   
      
   What you described can be considered as an interactive   
   version of "closed loop reinforcement learning scheme".   
   Where uers specify parameters in each loop, and the   
   parameters were optimized based on some test results.  Users   
   then partition the optimal parameter further.....   
      
   In an automated closed loop learning scheme, in each cycle,   
   a set of parameters is optimized based on a "reward   
   function"  or "belief function".  The parameter(s) with the   
   optimal value(s) represent(s) "potential problem".   In the   
   next cycle, that parameter is partitioned down further, and   
   the learning continue....   
      
   This is a difficult problem.  The search through the   
   parameter space could end up diverging.  Some pf the key   
   issues are:   
   -  the learning scheme that generates belief values   
   dynamically in each cycle must have a minima, or a finite   
   set of minimum points to guarantee a solution can be chosen   
   for hte next cycle.  Also to guarantee solutions that make   
   sense, the learning scheme will have to contain all key   
   information about causal relationship among components   
   -  how to subparametrize.   The program will have to be   
   supported by a sufficiently rich hierachical knowledgebase   
   of physical components.   
      
      
   However, your program may be a good starting point of future   
   program wiht higher level of automation.  The knowledgebase   
   of components are compiled, and reward function is refined   
   each time your program is run.   
      
   [ comp.ai is moderated.  To submit, just post and be patient, or if ]   
   [ that fails mail your article to , and ]   
   [ ask your news administrator to fix the problems with your system. ]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)