Re: Benchmarking ROOT 2.21/07 + PGCC (2)

Rene Brun (Rene.Brun@cern.ch)
Fri, 12 Mar 1999 13:28:22 +0000

Hi Jacek,
To answer your question, one has to see more benchmarks.
The overal result 2:1 is a very good result in favour of the
interpreter.
However, note that when running with the interpreter, we call most
of the time compiled functions. I believe this reflects quite well
what could be a realistic data analysis scenario.
Note also that stress 16 gives a very bad result for CINT. I added
this test on purpose. This is an example with several nested loops
for which the current CINT cannot generate byte code. Probably, one day,
you will see suddenly a tremendous speed-up in this last step.
But, I repeat, send us the benchmark results only once we have
released 2.21 next week. I still want to change a few things
in these benchmark programs.

Rene Brun

Jacek M. Holeczek wrote:
>
> > Could you tell me also which I/O system your PII 450 has (SCSI, ultra-wide
> > SCSI, UATA, disk rpms, etc). I've a dual PII 450 with UATA 7200rpm disk at
> The "data" disk is a local HD 10GB UDMA ( I have forgotten the manufacturer ).
> ROOT executables/libraries are on a NFS mounted disk ( via 10Mb ethernet ).
> > home so I will make the comparison this evening.
> Fine.
> I' interested not only in the I/O performance, but ( mainly ) in Real/Cpu
> Times of "stress" ( and of ".x benchmarks.C" ).
> I mean, for example, is the 2:1 ratio between interpreted and compiled
> "stress" very good, good, ok, bad, or very bad ?
> If it's worse then yours I'd like to try different optimization flags.
> ( BTW. maybe the one who reported a significant performance increase when
> using "loop unrolling" did it on the rootdev list, not roottalk ? But I
> can't find the archive of rootdev to search it. Any ideas ? )
> Jacek.