Interactive massively parallel computations are critical for machine learning and data analysis. These computations are a staple of the MIT Lincoln Laboratory Supercomputing Center (LLSC) and has required the LLSC to develop unique interactive supercomputing capabilities. Scaling interactive machine learning frameworks, such as TensorFlow, and data analysis environments, such as MATLAB/Octave, to tens of thousands of cores presents many technical challenges -in particular, rapidly dispatching many tasks through a scheduler, such as Slurm, and starting many instances of applications with thousands of dependencies. Careful tuning of launches and prepositioning of applications overcome these challenges and allow the launching of thousands of tasks in seconds on a 40,000-core supercomputer. Specifically, this work demonstrates launching 32,000 TensorFlow processes in 4 seconds and launching 262,000 Octave processes in 40 seconds. These capabilities allow researchers to rapidly explore novel machine learning architecture and data analysis algorithms.

No abstract

In the rapidly expanding field of parallel processing, job schedulers are the "operating systems" of modern big data architectures and supercomputing systems. Job schedulers allocate computing resources and control the execution of processes on those resources. Historically, job schedulers were the domain of supercomputers, and job schedulers were designed to run massive, long-running computations over days and weeks. More recently, big data workloads have created a need for a new class of computations consisting of many short computations taking seconds or minutes that process enormous quantities of data. For both supercomputers and big data systems, the efficiency of the job scheduler represents a fundamental limit on the efficiency of the system. Detailed measurement and modeling of the performance of schedulers are critical for maximizing the performance of a large-scale computing system. This paper presents a detailed feature analysis of 15 supercomputing and big data schedulers. For big data workloads, the scheduler latency is the most important performance characteristic of the scheduler. A theoretical model of the latency of these schedulers is developed and used to design experiments targeted at measuring scheduler latency. Detailed benchmarking of four of the most popular schedulers (Slurm, Son of Grid Engine, Mesos, and Hadoop YARN) are conducted. The theoretical model is compared with data and demonstrates that scheduler performance can be characterized by two key parameters: the marginal latency of the scheduler t s and a nonlinear exponent α s . For all four schedulers, the utilization of the computing system decreases to <10% for computations lasting only a few seconds. Multi-level schedulers (such as LLMapReduce) that transparently aggregate short computations can improve utilization for these short computations to >90% for all four of the schedulers that were tested.

Abstract-The Apache Accumulo database is an open source relaxed consistency database that is widely used for government applications. Accumulo is designed to deliver high performance on unstructured data such as graphs of network data. This paper tests the performance of Accumulo using data from the Graph500 benchmark. The Dynamic Distributed Dimensional Data Model (D4M) software is used to implement the benchmark on a 216-node cluster running the MIT SuperCloud software stack. A peak performance of over 100,000,000 database inserts per second was achieved which is 100x larger than the highest previously published value for any other database.The performance scales linearly with the number of ingest clients, number of database servers, and data size. The performance was achieved by adapting several supercomputing techniques to this application: distributed arrays, domain decomposition, adaptive load balancing, and single-program-multiple-data programming.