Speaker
Description
The use of accelerators such as GPUs has become mainstream to
achieve high performance on modern computing systems. GPUs come with
their own (limited) memory and are connected to the main memory of
the machine through a bus (with limited bandwidth). When a
computation is started on a GPU, the corresponding data needs to
be transferred to the GPU before the computation starts. Such data movements
may become a bottleneck for performance, especially when several
GPUs have to share the communication bus.
Task-based runtime schedulers have emerged as a convenient and
efficient way to use such heterogeneous platforms. When processing an application,
the scheduler has the knowledge of all tasks available for
processing on a GPU, as well as their input data dependencies.
Hence, it is possible to produce a tasks processing order
aiming at reducing the total processing time through three objectives:
minimizing data transfers, overlapping transfers and computation and
optimizing the eviction of previously-loaded data. We
focus on this problem of partitioning and ordering
tasks that share some of their input data on multiple GPUs. We present a novel
dynamic strategy based on data selection
to efficiently allocate tasks to GPUs and a custom eviction policy, and compare
them to existing strategies using either a well-known graph
partitioner or standard scheduling techniques in runtime
systems.
We present their performance on tasks from tiled 2D, 3D matrix products and Cholesky factorization,
as well as a sparse matrix product.
All strategies have been implemented on top of the StarPU
runtime, and we show that our dynamic strategy achieves better performance
when scheduling tasks on multiple GPUs with limited memory.
