Hardware Reference
In-Depth Information
FIGURE 18.1: Relationships between GLEAN and principal components of
an HPC application.
nodes. Using GLEAN, one can apply custom analyses to the data on the com-
pute resource or on the staging nodes. This can help reduce the amount of
data written out to storage. On the staging nodes, GLEAN uses MPI-IO or
higher-level I/O libraries to write the data out asynchronously to storage.
GLEAN is implemented in C++ leveraging MPI and pthreads, and pro-
vides interfaces for Fortran and C-based parallel applications. It offers a flex-
ible and extensible API that can be customized to meet the needs of the
application. The following subsections describe GLEAN's design in terms of
four principal features: network topology and reduced synchronization, data
semantics, asynchronous data staging, compression and subfiling.
18.2.1 Exploiting Network Topology and Reduced
Synchronization for I/O
As system designs continue to evolve toward heterogeneous and complex
network topologies, effective ways to fully exploit their heterogeneity is becom-
ing more critical. For example, in BG/P, the system has different networks
with varying throughputs and topologies. The 3D torus interconnects a com-
pute node with its six neighbors at 425 MB/s over each link. In contrast, the
collective network is a shared network with a maximum throughput of 850
MB/s to the I/O network. The collective network is the only way to get to
the I/O network in order to perform I/O. Recent supercomputing systems
have a higher radix interconnect|BG/Q has a 5D torus and the K-machine
