Scaling physics and material science applications on a massively parallel Blue Gene/L system

Abstract

Blue Gene/L represents a new way to build supercomputers, using a large number of low power processors, together with multiple integrated interconnection networks. Whether real applications can scale to tens of thousands of processors (on a machine like Blue Gene/L) has been an open question. In this paper, we describe early experience with several physics and material science applications on a 32,768 node Blue Gene/L system, which was installed recently at the Lawrence Livermore National Laboratory. Our study shows some problems in the applications and in the current software implementation, but overall, excellent scaling of these applications to 32K nodes on the current Blue Gene/L system. While there is clearly room for improvement, these results represent the first proof point that MPI applications can effectively scale to over ten thousand processors. They also validate the scalability of the hardware and software architecture of Blue Gene/L.

Task mapping on supercomputers with cellular networks

“…Without an appropriate assignment, even an ultra-scalable supercomputer system such as Blue Gene/L with huge computing power will only realize a small portion of its potential [33]. Our models are great effective tools to avoid this kind of power waste.…”

“…We select the BG/L system as the testing platform because it meets all our major requirements: It is used worldwide and large partitions are readily available. It supports MPI [32,33] and offers an MPI re-map interface [35].…”

“…A good mapping can reduces the average running time by efficient utilization of the network resource [33].…”

High Performance File I/O for The Blue Gene/L Supercomputer

“…I/O related system calls trapped in CNK are shipped to the corresponding I/O node and processed by a console daemon CIOD, thus a POSIX-like I/O interface is supported for userlevel compute processes. The separation and cooperation accomplished via the simple function shipping off-loads non-computation related services to the I/O nodes to keep a noise-free state for the compute nodes while achieving superior scalability [11]. In addition, the separation of I/O nodes and compute nodes, together with organizing them into balanced psets and providing almost dedicated communication channels for I/O operations, essentially provide the capability for scalable I/O.…”

“…BG/L is currently ranked as the world's fastest supercomputer in Top500 list of supercomputers, and represents five of the top ten entities in the list. While impressive scaling results have been obtained up to 32768 nodes [11], there has been relatively few results published on I/O performance for data-intensive applications on BG/L or any other massively parallel system.…”