High Performance Computing

Since we began using the high performance computing (HPC) cluster in 2008, its needs have grown rapidly each year. The CPU core count has increased from 1024 to nearly 3000 and the storage has increased from 100TB to over 1PB. After so many years, the cluster has significant amounts of old equipment which need to be replaced.

In Q1 2022, we made a major upgrade to our cluster. We purchased a number of compute nodes, consisting of 4096 CPU cores and 16TB of memory (4GB/core). We also added a new 1PB storage server for usage with the cluster. This MKI-owned cluster is part of a much larger MIT cluster, which is contributed to and shared among many departments and PI groups within the MIT research community. We are currently in the process of adding a GPU server with two A100 processors, as well as six 4-way NV-linked A100 GPU servers to cover different research needs. 

Our user base are the PI groups in many research areas. For example:

Mark Vogelsberger’s Group

uses the cluster to perform large hydrodynamical cosmological simulations to understand galaxy and structure formation. Research topics include, for example, studying the growth and evolution of galaxies, the detailed astrophysical processes in galaxies and the nature of dark matter. Furthermore, the group also hosts large simulation datasets on the cluster to make them available to collaborators.

Jacqueline Hewitt’s Group

uses HPC to simulate the formation of stars in galaxies at high redshift, making predictions about the properties of the 21cm line of neutral hydrogen.  They are building radio arrays to detect this line, so having the simulations helps us design arrays and interpret our data. We also use HPC to analyze the radio data.  Our calibration and mapping algorithms are very compute-intensive.

Scott Hughes’ group

uses large-scale computing to model binary sources of gravitational waves, and in the past has used computing to examine how well waveform properties can be determined by realistic measurements.  Modeling sources is presently the main focus, using large, parallel, high-precision and high-performance computations in order to solve the equations of general relativity over a wide-span of astrophysically relevant parameters.  We use large-scale computing both to compute very accurate waveforms, and to develop high-fidelity approximations which describe aspects of waveform physics with reduced computational cost.

Lina Necib’s Group

runs simulations and analyzes data from the Gaia mission. More specifically, the Necib group runs isolated simulations of merger events in which the Milky Way collides with a second galaxy. This is important in understanding the formation of our Galaxy as well as the current distribution of Dark Matter under different models. These simulated results are used to compare with observational data from the largest stellar kinematic catalog to data: Gaia. The group uses the cluster to isolate kinematic structures, calculate properties of the Galaxy, and extract properties of Dark Matter to make experimental predictions. 

Andrew Vanderburg’s Group

uses high performance computing for a wide range of tasks involving the detection and characterization of exoplanets. These can range from the Bayesian analysis of telescope observations to determine physical sizes and masses of exoplanets, to the study of the orbits of binary stars that host exoplanets, to machine learning tools to study the elemental compositions of extrasolar asteroids.