Now 38 years
this is basically counting back to my diploma thesis where I developed my
first performance tools in 1987.
I just realized by discussions this morning I should have slightly made a slight modification
to the title.
It should be Memories of a Parallel Performance Engineer.
I always was interested in how to optimize communication
synchronization
parallelism
and not so much the more node level single core performance things where the experts
sitting here in Erlangen.
So that we are kind of nicely match each other
you experience and our experience on the other
side.
Okay
I basically made this talk historically so people at least in the room should recognize
the picture.
So it's the Meanwhile Handler Tower
I think it's called in English.
So it's the Institute for Informatics in Erlangen and I've been there like
you know
basically
I've been there 82, starting my computer science there but working on parallel tools was with
my diploma series in 87.
And the context was at that time there was a Sonderforschungsbereich 182
multiprocessor
and network configurations
like multiprocessor network configurations
went over 12 years
in four times three phases and I was part of it in the first two phases.
So with people who are non-chairmen
so Sonderforschungsbereich is like a large grant given to a university.
I think in our case it was over 30 PhD people working on different aspects from like operating
system all the way up and as I said it's early 80s and it was like really very early time
for a university to think about parallelism
parallel computers and so on.
And I personally was part of Work Package C1
Messung
Modellierung
Bewertung von multiprocessor
Rechnernetzen
like measurement
modelling and evaluation of processes and networks.
And there's a lot of parallel system development so I encourage you to come here and there's
actually still
these machines are partially visible here and you can see them.
So ECBA
which was the first parallel system already in the 60s or 70s and then DIRMU and
MEMSY in the 80s, early 90s or something.
Another thing you will see is
Presenters
Zugänglich über
Offener Zugang
Dauer
01:13:00 Min
Aufnahmedatum
2025-10-21
Hochgeladen am
2025-11-07 13:20:05
Sprache
en-US
Speaker: Dr. Bernd Mohr, Jülich Supercomputing Center
Slides
Abstract:
I created my first performance tool for parallel systems as part of my diploma thesis at the Friedrich-Alexander-Universität Erlangen-Nürnberg, Bavaria, in 1987. Since then, I never changed my research topic and was involved in the development of many (well-known?) HPC performance tools: TDL/POET, ZM4, SIMPLE, TAU, Vampir, EPILOG, KOJAK, Score-P, Scalasca, and Cube. The talk will highlight successes and failures of the different tools and will provide a personal assessment of the state of HPC performance tools.