I have been running some tests using 80 workers and I have found a contradiction !!!
Does it make sense that the compiled (even optimised [ … .C+O ]) mode is a factor of 10 times slower than the interpreted mode ?
Also does anybody know why Proof crashes when I setup the parallel mode using compiled mode for less than 16 workers, while exactly the same code set in CINT(simply removing the ++), with the same 16 workers or less runs perfectly ?
By other hand I have realised that the inclusion of vectors of structures (vector muon), or vectors of vectors (vector muon ) runs without having to add the dictionary definitions, and the posterior loading of the PAR, just switching to the compiled mode… is it the rigth way ?
Only if you include the time to build and load the code in compiled mode; building and loading can be considerably longer for small bits of code. If instead you see 1 hour vs 10 hours then I would claim that you are not measuring what you think you are Please send a running, standalone example if you want me to check.
I’ll leave the other questions for our PROOF masters.
After some discussions and tests with Montoya we found that, in the compiled mode, on the workers, there were compilation errors which have to be understood/fixed before any performance tests.
[quote]
Also does anybody know why Proof crashes when I setup the parallel mode using compiled mode for less than 16 workers, while exactly the same code set in CINT(simply removing the ++), with the same 16 workers or less runs perfectly ? [/quote]
Can you post the error messages from the session logs.
The problem was solved. Compiled mode is much faster, some machines had more than one gcc compiler, and in fact the session was not running in all workers
Please send a running, standalone example if you want me to check.