[SimGrid-user] SimGrid 3.16 is expected to be released by June 21.

Arnaud Legrand arnaud.legrand at imag.fr
Fri Jun 9 12:46:07 CEST 2017


Le 08/06/2017 22:42, Martin Quinson a écrit :
> SMPI performance were further improved, using various OS-level magic
> to speed up our virtualization of the MPI code to be run. This allowed
> Tom to run a simulation of HPL involving 10^6 processes... Wow. 

Not quite. To be perfectly clear, we (Tom Cornebize actually) have
been able to emulate HPL with 4096 MPI process, which may seem few but
the hardest part is actually not the number of MPI processes but the
rank of the Matrix used in HPL as both simulated and simulation time
inherently grow quadratically with the rank. At the moment, systems
that have the structure of stampede (thousands of nodes with
multi-cores/accelerators) seem thus reachable with SMPI (if MPI is not
run in a flat way at the core level) even with non trivial workloads
(such as the ones made when applying for the top500), which is a great
news. This was much more complicated than I anticipated and the dlopen
OS-level magic was a key ingredient to achieve this.

Regarding executions with a large number of processes, we can
obviously run a { MPI_init(); printf("Hello world!\n"); MPI_Finalize() }
with 1E6 processes but it seems quite pointless to me. More
information about scalability can be found in our latest article on
this topic: https://hal.inria.fr/hal-01415484v2

> And the work for v3.17 started already. Here are some changesets that
> are not ready to be included yet but may well be part of 3.17:
>  - Energy modeling for the network too
>  - New reduction algorithm for the model-checker, based on event
>    folding structures
>  - multi-model simulations, to specify a differing networking model
>    for each netzone.

I can't wait for it! :) Thanks for all the great work Martin.



If your computer speaks English, it was probably made in Japan.
   -- "Epigrams in Programming", by Alan J. Perlis of Yale University.

More information about the Simgrid-user mailing list