Post-doctoral fellow in the Research Group Parallel Computing at TU Wien
Twitter
LinkedIn
GitHub
GitLab
I work on improving performance of HPC runtime systems, especially distributed runtime systems.
I did my PhD thesis in the TADaaM team in
Inria Bordeaux Sud-Ouest,
under the supervision of Alexandre Denis
and Emmanuel Jeannot.
I worked on interactions between task-based runtime systems and communication libraries for
High Performance Computing, especially StarPU
and NewMadeleine.
I developed two main directions in my thesis:
To be able to optimize broadcasts appearing in task graphs of
StarPU applications, we developed what we called dynamic
broadcasts. Functions such as MPI_Bcast
cannot be used within StarPU, mainly because recipient processes do
not know wheter data comes from a broadcast or a regular
point-to-point request, and in the same fashion they do not know
other nodes involved in the broadcast. Thus, only the original
sender node has all informations to be able to call
MPI_Bcast
. Accurately detecting broadcasts in the task
graph is not straighforward neither.
Our dynamic broadcasts overcome these constraints. The
broadcast communication pattern required by the task-based
algorithm is detected automatically, then the broadcasting
algorithm relies on active messages and source routing, so that
participating nodes do not need to know each other and do not
need to synchronize. Receiver receives data the same way as it
receives point-to-point communication, without having to know
it arrives through a broadcast.
To amortize the cost of MPI communications, distributed
parallel runtime systems can usually overlap network
communications with computations in the hope that it improves
global application performance. When using this technique, both
computations and communications are running at the same time.
We studied the possible interferences between computations and
communications when they are executed in parallel. The main
interference that can occur is memory contention between data
used by computations and data used by network communications.
In some cases, this contention can cause severe slowdown of
both computations and communications.
To predict memory bandwidth for computations and for
communications when they are executed side by side, we proposed
a model taking data locality and contention into account.
Elaboration of the model allowed to better understand locations
of bottlenecks in the memory system and what are the strategies
of the memory system in case of contention.
I defended my thesis On
the Interactions between HPC Task-based Runtime Systems and Communication Libraries
on November 29, 2022. The
manuscript
is available, the slides as well.
Authors are sorted by alphabetical order.
A complete list of my publications is available on HAL.
swartvagher [at] par [dot] tuwien [dot] ac [dot] at
I have a PGP key: 0x6EC3C10693C090C3 (942A 2C17 4547 99B5 62E3 4127 6EC3 C106 93C0 90C3).