arXiv subjects

David Brooks

Publications and source records attributed to David Brooks.

At least 199 records · Page 11Linked to original sources

arXiv2021

Application-driven Design Exploration for Dense Ferroelectric Embedded Non-volatile Memories

The memory wall bottleneck is a key challenge across many data-intensive applications. Multi-level FeFET-based embedded non-volatile memories are a promising solution for denser and more energy-efficient on-chip memory. However, reliable multi-level cell storage requires careful optimizations to minimize the design overhead costs. In this work, we investigate the interplay between FeFET device characteristics, programming schemes, and memory array architecture, and explore different design choices to optimize performance, energy, area, and accuracy metrics for critical data-intensive workloads. From our cross-stack design exploration, we find that we can store DNN weights and social network graphs at a density of over 8MB/mm^2 and sub-2ns read access latency without loss in application accuracy.

cs.DC↗

arXiv2021

Gradient Disaggregation: Breaking Privacy in Federated Learning by Reconstructing the User Participant Matrix

We show that aggregated model updates in federated learning may be insecure. An untrusted central server may disaggregate user updates from sums of updates across participants given repeated observations, enabling the server to recover privileged information about individual users' private training data via traditional gradient inference attacks. Our method revolves around reconstructing participant information (e.g: which rounds of training users participated in) from aggregated model updates by leveraging summary information from device analytics commonly used to monitor, debug, and manage federated learning systems. Our attack is parallelizable and we successfully disaggregate user updates on settings with up to thousands of participants. We quantitatively and qualitatively demonstrate significant improvements in the capability of various inference attacks on the disaggregated updates. Our attack enables the attribution of learned properties to individual users, violating anonymity, and shows that a determined central server may undermine the secure aggregation protocol to break individual users' data privacy in federated learning.

cs.CR↗

arXiv2021

Logic Compatible High-Performance Ferroelectric Transistor Memory

Silicon ferroelectric field-effect transistors (FeFETs) with low-k interfacial layer (IL) between ferroelectric gate stack and silicon channel suffers from high write voltage, limited write endurance and large read-after-write latency due to early IL breakdown and charge trapping and detrapping at the interface. We demonstrate low voltage, high speed memory operation with high write endurance using an IL-free back-end-of-line (BEOL) compatible FeFET. We fabricate IL-free FeFETs with 28nm channel length and 126nm width under a thermal budget <400C by integrating 5nm thick Hf0.5Zr0.5O2 gate stack with amorphous Indium Tungsten Oxide (IWO) semiconductor channel. We report 1.2V memory window and read current window of 10^5 for program and erase, write latency of 20ns with +/-2V write pulses, read-after-write latency <200ns, write endurance cycles exceeding 5x10^10 and 2-bit/cell programming capability. Array-level analysis establishes IL-free BEOL FeFET as a promising candidate for logic-compatible high-performance on-chip buffer memory and multi-bit weight cell for compute-in-memory accelerators.

cond-mat.mes-hall↗

arXiv2021

RecPipe: Co-designing Models and Hardware to Jointly Optimize Recommendation Quality and Performance

Deep learning recommendation systems must provide high quality, personalized content under strict tail-latency targets and high system loads. This paper presents RecPipe, a system to jointly optimize recommendation quality and inference performance. Central to RecPipe is decomposing recommendation models into multi-stage pipelines to maintain quality while reducing compute complexity and exposing distinct parallelism opportunities. RecPipe implements an inference scheduler to map multi-stage recommendation engines onto commodity, heterogeneous platforms (e.g., CPUs, GPUs).While the hardware-aware scheduling improves ranking efficiency, the commodity platforms suffer from many limitations requiring specialized hardware. Thus, we design RecPipeAccel (RPAccel), a custom accelerator that jointly optimizes quality, tail-latency, and system throughput. RPAc-cel is designed specifically to exploit the distinct design space opened via RecPipe. In particular, RPAccel processes queries in sub-batches to pipeline recommendation stages, implements dual static and dynamic embedding caches, a set of top-k filtering units, and a reconfigurable systolic array. Com-pared to prior-art and at iso-quality, we demonstrate that RPAccel improves latency and throughput by 3x and 6x.

cs.AR↗

arXiv2021

Dark Energy Survey Year 1 Results: Constraining Baryonic Physics in the Universe

Measurements of large-scale structure are interpreted using theoretical predictions for the matter distribution, including potential impacts of baryonic physics. We constrain the feedback strength of baryons jointly with cosmology using weak lensing and galaxy clustering observables (3$\times$2pt) of Dark Energy Survey (DES) Year 1 data in combination with external information from baryon acoustic oscillations (BAO) and Planck cosmic microwave background polarization. Our baryon modeling is informed by a set of hydrodynamical simulations that span a variety of baryon scenarios; we span this space via a Principal Component (PC) analysis of the summary statistics extracted from these simulations. We show that at the level of DES Y1 constraining power, one PC is sufficient to describe the variation of baryonic effects in the observables, and the first PC amplitude ($Q_1$) generally reflects the strength of baryon feedback. With the upper limit of $Q_1$ prior being bound by the Illustris feedback scenarios, we reach $\sim 20\%$ improvement in the constraint of $S_8=σ_8(Ω_{\rm m}/0.3)^{0.5}=0.788^{+0.018}_{-0.021}$ compared to the original DES 3$\times$2pt analysis. This gain is driven by the inclusion of small-scale cosmic shear information down to 2.5 arcmin, which was excluded in previous DES analyses that did not model baryonic physics. We obtain $S_8=0.781^{+0.014}_{-0.015}$ for the combined DES Y1+Planck EE+BAO analysis with a non-informative $Q_1$ prior. In terms of the baryon constraints, we measure $Q_1=1.14^{+2.20}_{-2.80}$ for DES Y1 only and $Q_1=1.42^{+1.63}_{-1.48}$ for DESY1+Planck EE+BAO, allowing us to exclude one of the most extreme AGN feedback hydrodynamical scenario at more than $2 σ$.

astro-ph.CO↗

arXiv2021

The Clustering of DESI-like Luminous Red Galaxies Using Photometric Redshifts

We present measurements of the redshift-dependent clustering of a DESI-like luminous red galaxy (LRG) sample selected from the Legacy Survey imaging dataset, and use the halo occupation distribution (HOD) framework to fit the clustering signal. The photometric LRG sample in this study contains 2.7 million objects over the redshift range of $0.4 < z < 0.9$ over 5655 deg$^2$. We have developed new photometric redshift (photo-$z$) estimates using the Legacy Survey DECam and WISE photometry, with $σ_{\mathrm{NMAD}} = 0.02$ precision for LRGs. We compute the projected correlation function using new methods that maximize signal-to-noise ratio while incorporating redshift uncertainties. We present a novel algorithm for dividing irregular survey geometries into equal-area patches for jackknife resampling. For a five-parameter HOD model fit using the MultiDark halo catalog, we find that there is little evolution in HOD parameters except at the highest redshifts. The inferred large-scale structure bias is largely consistent with constant clustering amplitude over time. In an appendix, we explore limitations of Markov chain Monte Carlo fitting using stochastic likelihood estimates resulting from applying HOD methods to N-body catalogs, and present a new technique for finding best-fit parameters in this situation. Accompanying this paper we have released the Photometric Redshifts for the Legacy Surveys (PRLS) catalog of photo-$z$'s obtained by applying the methods used in this work to the full Legacy Survey Data Release 8 dataset. This catalog provides accurate photometric redshifts for objects with $z < 21$ over more than 16,000 deg$^2$ of sky.

astro-ph.CO↗

arXiv2021

RecSSD: Near Data Processing for Solid State Drive Based Recommendation Inference

Neural personalized recommendation models are used across a wide variety of datacenter applications including search, social media, and entertainment. State-of-the-art models comprise large embedding tables that have billions of parameters requiring large memory capacities. Unfortunately, large and fast DRAM-based memories levy high infrastructure costs. Conventional SSD-based storage solutions offer an order of magnitude larger capacity, but have worse read latency and bandwidth, degrading inference performance. RecSSD is a near data processing based SSD memory system customized for neural recommendation inference that reduces end-to-end model inference latency by 2X compared to using COTS SSDs across eight industry-representative models.

cs.AR↗

arXiv2021

Installation of the Dark Energy Spectroscopic Instrument at the Mayall 4-meter telescope

The Dark Energy Spectroscopic Instrument (DESI) is a Stage IV ground-based dark energy experiment that will measure the expansion history of the Universe using the Baryon Acoustic Oscillation technique. The spectra of 35 million galaxies and quasars over 14000 square degrees will be measured during the life of the experiment. We describe the installation of the major elements of the instrument at the Mayall 4m telescope, completed in late 2019. The previous prime focus corrector, spider vanes, and upper rings were removed from the Mayall's Serrurier truss and replaced with the newly-constructed DESI ring, vanes, cage, hexapod, and optical corrector. The new corrector was optically aligned with the primary mirror using a laser tracker system. The DESI focal plane system was integrated to the corrector, with each of its ten 500-fiber-positioner petal segments installed using custom installation hardware and the laser tracker. Ten DESI spectrographs with 30 cryostats were installed in a newly assembled clean room in the Large Coude Room. The ten cables carrying 5000 optical fibers from the positioners in the focal plane were routed down the telescope through cable wraps at the declination and hour angle axes, and their integral slitheads were integrated with the ten spectrographs. The fiber view camera assembly was installed to the Mayall's primary mirror cell. Servers for the instrument control system replaced existing computer equipment. The fully integrated instrument has been commissioned and is ready to start its operations phase.

astro-ph.IM↗

arXiv2021

The DESI Sky Continuum Monitor System

The Dark Energy Spectroscopic Instrument (DESI) is an ongoing spectroscopic survey to measure the dark energy equation of state to unprecedented precision. We describe the DESI Sky Continuum Monitor System, which tracks the night sky brightness as part of a system that dynamically adjusts the spectroscopic exposure time to produce more uniform data quality and to maximize observing efficiency. The DESI dynamic exposure time calculator (ETC) will combine sky brightness measurements from the Sky Monitor with data from the guider system to calculate the exposure time to achieve uniform signal-to-noise ratio (SNR) in the spectra under various observing conditions. The DESI design includes 20 sky fibers, and these are split between two identical Sky Monitor units to provide redundancy. Each Sky Monitor unit uses an SBIG STXL-6303e CCD camera and supports an eight-position filter wheel. Both units have been completed and delivered to the Mayall Telescope at the Kitt Peak National Observatory. Commissioning results show that the Sky Monitor delivers the required performance necessary for the ETC.

astro-ph.IM↗

arXiv2021

Performance of Kitt Peak's Mayall 4-meter Telescope During DESI Commissioning

In preparation for the Dark Energy Spectroscopic Instrument (DESI), a new top end was installed on the Mayall 4-meter telescope at Kitt Peak National Observatory. The refurbished telescope and the DESI instrument were successfully commissioned on sky between 2019 October and 2020 March. Here we describe the pointing, tracking and imaging performance of the Mayall telescope equipped with its new DESI prime focus corrector, as measured by six guider cameras sampling the outer edge of DESI's focal plane. Analyzing ~500,000 guider images acquired during commissioning, we find a median delivered image FWHM of 1.1 arcseconds (in the r-band at 650 nm), with the distribution extending to a best-case value of ~0.6 arcseconds. The point spread function is well characterized by a Moffat profile with a power-law index of $β$ ~ 3.5 and little dependence of $β$ on FWHM. The shape and size of the PSF delivered by the new corrector at a field angle of 1.57 degrees are very similar to those measured with the old Mayall corrector on axis. We also find that the Mayall achieves excellent pointing accuracy (several arcseconds RMS) and minimal open-loop tracking drift (< 1 milliarcsecond per second), improvements on the telecope's pre-DESI performance. In the future, employing DESI's active focus adjustment capabilities will likely further improve the Mayall/DESI delivered image quality.

astro-ph.IM↗

arXiv2020

Chasing Carbon: The Elusive Environmental Footprint of Computing

Given recent algorithm, software, and hardware innovation, computing has enabled a plethora of new applications. As computing becomes increasingly ubiquitous, however, so does its environmental impact. This paper brings the issue to the attention of computer-systems researchers. Our analysis, built on industry-reported characterization, quantifies the environmental effects of computing in terms of carbon emissions. Broadly, carbon emissions have two sources: operational energy consumption, and hardware manufacturing and infrastructure. Although carbon emissions from the former are decreasing thanks to algorithmic, software, and hardware innovations that boost performance and power efficiency, the overall carbon footprint of computer systems continues to grow. This work quantifies the carbon output of computer systems to show that most emissions related to modern mobile and data-center equipment come from hardware manufacturing and infrastructure. We therefore outline future directions for minimizing the environmental impact of computing systems.

cs.AR↗

arXiv2020

Preliminary Target Selection for the DESI Milky Way Survey (MWS)

The DESI Milky Way Survey (MWS) will observe $\ge$8 million stars between $16 < r < 19$ mag, supplemented by observations of brighter targets under poor observing conditions. The survey will permit an accurate determination of stellar kinematics and population gradients; characterize diffuse substructure in the thick disk and stellar halo; enable the discovery of extremely metal-poor stars and other rare stellar types; and improve constraints on the Galaxy's 3D dark matter distribution from halo star kinematics. MWS will also enable a detailed characterization of the stellar populations within 100 pc of the Sun, including a complete census of white dwarfs. The target catalog from the preliminary selection described here is public.

astro-ph.GA↗

arXiv2020

Preliminary Target Selection for the DESI Bright Galaxy Survey (BGS)

The Dark Energy Spectroscopic Instrument (DESI) will execute a nearly magnitude-limited survey of low redshift galaxies ($0.05 \leq z \leq 0.4$, median $z \approx 0.2$). Clustering analyses of this Bright Galaxy Survey (BGS) will yield the most precise measurements to date of baryon acoustic oscillations and redshift-space distortions at low redshift. DESI BGS will comprise two target classes: (i) BRIGHT ($r<19.5$~mag), and (ii) FAINT ($19.5<r<20$~mag). Here we present a summary of the star-galaxy separation, and different photometric and geometrical masks, used in BGS to reduce the number of spurious targets. The selection results in a total density of $\sim 800$ objects/deg$^2$ for the BRIGHT and $\sim 600$ objects/deg$^2$ for the FAINT selections.A full characterization of the BGS selection can be found in Ruiz-Macias et al. (2020).

astro-ph.GA↗

arXiv2020

LyaCoLoRe: Synthetic Datasets for Current and Future Lyman-$α$ Forest BAO Surveys

The statistical power of Lyman-$α$ forest Baryon Acoustic Oscillation (BAO) measurements is set to increase significantly in the coming years as new instruments such as the Dark Energy Spectroscopic Instrument deliver progressively more constraining data. Generating mock datasets for such measurements will be important for validating analysis pipelines and evaluating the effects of systematics. With such studies in mind, we present LyaCoLoRe: a package for producing synthetic Lyman-$α$ forest survey datasets for BAO analyses. LyaCoLoRe transforms initial Gaussian random field skewers into skewers of transmitted flux fraction via a number of fast approximations. In this work we explain the methods of producing mock datasets used in LyaCoLoRe, and then measure correlation functions on a suite of realisations of such data. We demonstrate that we are able to recover the correct BAO signal, as well as large-scale bias parameters similar to literature values. Finally, we briefly describe methods to add further astrophysical effects to our skewers - high column density systems and metal absorbers - which act as potential complications for BAO analyses.

astro-ph.CO↗

arXiv2020

Discovery of a Candidate Binary Supermassive Black Hole in a Periodic Quasar from Circumbinary Accretion Variability

Binary supermassive black holes (BSBHs) are expected to be a generic byproduct from hierarchical galaxy formation. The final coalescence of BSBHs is thought to be the loudest gravitational wave (GW) siren, yet no confirmed BSBH is known in the GW-dominated regime. While periodic quasars have been proposed as BSBH candidates, the physical origin of the periodicity has been largely uncertain. Here we report discovery of a periodicity (P=1607$\pm$7 days) at 99.95% significance (with a global p-value of ~$10^{-3}$ accounting for the look elsewhere effect) in the optical light curves of a redshift 1.53 quasar, SDSS J025214.67-002813.7. Combining archival Sloan Digital Sky Survey data with new, sensitive imaging from the Dark Energy Survey, the total ~20-yr time baseline spans ~4.6 cycles of the observed 4.4-yr (restframe 1.7-yr) periodicity. The light curves are best fit by a bursty model predicted by hydrodynamic simulations of circumbinary accretion disks. The periodicity is likely caused by accretion rate modulation by a milli-parsec BSBH emitting GWs, dynamically coupled to the circumbinary accretion disk. A bursty hydrodynamic variability model is statistically preferred over a smooth, sinusoidal model expected from relativistic Doppler boost, a kinematic effect proposed for PG1302-102. Furthermore, the frequency dependence of the variability amplitudes disfavors Doppler boost, lending independent support to the circumbinary accretion variability hypothesis. Given our detection rate of one BSBH candidate from circumbinary accretion variability out of 625 quasars, it suggests that future large, sensitive synoptic surveys such as the Vera C. Rubin Observatory Legacy Survey of Space and Time may be able to detect hundreds to thousands of candidate BSBHs from circumbinary accretion with direct implications for Laser Interferometer Space Antenna.

astro-ph.GA↗

arXiv2020

A binary-activation, multi-level weight RNN and training algorithm for ADC-/DAC-free and noise-resilient processing-in-memory inference with eNVM

We propose a new algorithm for training neural networks with binary activations and multi-level weights, which enables efficient processing-in-memory circuits with embedded nonvolatile memories (eNVM). Binary activations obviate costly DACs and ADCs. Multi-level weights leverage multi-level eNVM cells. Compared to existing algorithms, our method not only works for feed-forward networks (e.g., fully-connected and convolutional), but also achieves higher accuracy and noise resilience for recurrent networks. In particular, we present an RNN-based trigger-word detection PIM accelerator, with detailed hardware noise models and circuit co-design techniques, and validate our algorithm's high inference accuracy and robustness against a variety of real hardware non-idealities.

cs.LG↗

arXiv2020

Cross-Stack Workload Characterization of Deep Recommendation Systems

Deep learning based recommendation systems form the backbone of most personalized cloud services. Though the computer architecture community has recently started to take notice of deep recommendation inference, the resulting solutions have taken wildly different approaches - ranging from near memory processing to at-scale optimizations. To better design future hardware systems for deep recommendation inference, we must first systematically examine and characterize the underlying systems-level impact of design decisions across the different levels of the execution stack. In this paper, we characterize eight industry-representative deep recommendation models at three different levels of the execution stack: algorithms and software, systems platforms, and hardware microarchitectures. Through this cross-stack characterization, we first show that system deployment choices (i.e., CPUs or GPUs, batch size granularity) can give us up to 15x speedup. To better understand the bottlenecks for further optimization, we look at both software operator usage breakdown and CPU frontend and backend microarchitectural inefficiencies. Finally, we model the correlation between key algorithmic model architecture features and hardware bottlenecks, revealing the absence of a single dominant algorithmic component behind each hardware bottleneck.

cs.AR↗

arXiv2020

Cheetah: Optimizing and Accelerating Homomorphic Encryption for Private Inference

As the application of deep learning continues to grow, so does the amount of data used to make predictions. While traditionally, big-data deep learning was constrained by computing performance and off-chip memory bandwidth, a new constraint has emerged: privacy. One solution is homomorphic encryption (HE). Applying HE to the client-cloud model allows cloud services to perform inference directly on the client's encrypted data. While HE can meet privacy constraints, it introduces enormous computational challenges and remains impractically slow in current systems. This paper introduces Cheetah, a set of algorithmic and hardware optimizations for HE DNN inference to achieve plaintext DNN inference speeds. Cheetah proposes HE-parameter tuning optimization and operator scheduling optimizations, which together deliver 79x speedup over the state-of-the-art. However, this still falls short of plaintext inference speeds by almost four orders of magnitude. To bridge the remaining performance gap, Cheetah further proposes an accelerator architecture that, when combined with the algorithmic optimizations, approaches plaintext DNN inference speeds. We evaluate several common neural network models (e.g., ResNet50, VGG16, and AlexNet) and show that plaintext-level HE inference for each is feasible with a custom accelerator consuming 30W and 545mm^2.

cs.CR↗