No abstract

We investigate operating system noise, which we identify as one of the main reasons for a lack of synchronicity in parallel applications. Using a microbenchmark, we measure the noise on several contemporary platforms and find that, even with a general-purpose operating system, noise can be limited if certain precautions are taken. We then inject artificially generated noise into a massively parallel system and measure its influence on the performance of collective operations. Our experiments indicate that on extreme-scale platforms, the performance is correlated with the largest interruption to the application, even if the probability of such an interruption on a single process is extremely small. We demonstrate that synchronizing the noise can significantly reduce its negative influence.

No abstract

Current leadership-class machines suffer from a significant imbalance between their computational power and their I/O bandwidth. I/O forwarding is a paradigm that attempts to bridge the increasing performance and scalability gap between the compute and I/O components of leadership-class machines to meet the requirements of data-intensive applications by shipping I/O calls from compute nodes to dedicated I/O nodes. I/O forwarding is a critical component of the I/O subsystem of the IBM Blue Gene/P supercomputer currently deployed at several leadership computing facilities. In this paper, we evaluate the performance of the existing I/O forwarding mechanisms for BG/P and identify the performance bottlenecks in the current design. We augment the I/O forwarding with two approaches: I/O scheduling using a work-queue model and asynchronous data staging. We evaluate the efficacy of our approaches using microbenchmarks and application-level benchmarks on leadershipclass systems. I. INTRODUCTIONLeadership-class systems are providing unprecedented opportunities to advance science in numerous fields, such as climate sciences, biosciences, astrophysics, computational chemistry, materials sciences, high-energy physics, and nuclear physics [17]. Current leadership-class machines such as the IBM Blue Gene/P (BG/P) supercomputer at the Argonne National Laboratory and the Cray XT system at the Oak Ridge National Laboratory consist of a few hundred thousand processing elements. BG/P is the second generation of supercomputers in the Blue Gene series and has demonstrated ultrascale performance together with a novel energy-efficient design. As of November 2009, five of the top 20 systems in the Top 500 list [20] and thirteen of the top 20 most powerefficient systems were based on the Blue Gene solution [8].While the computational power of supercomputers keeps increasing with every generation, the I/O systems have not kept pace, resulting in a significant performance bottleneck. In order to achieve higher performance, many HPC systems run a stripped-down operating system kernel on the compute nodes to reduce the operating system "noise." The IBM Blue Gene series of supercomputers takes this a step further, restricting I/O operations from the compute nodes. In order to enable applications to perform I/O, the compute node kernel ships all I/O operations to a dedicated I/O node, which performs I/O on behalf of the compute nodes. This process is known as I/O forwarding [3].

No abstract