Search for: All records

With the ever-increasing size of training models and datasets, network communication has emerged as a major bottleneck in distributed deep learning training. To address this challenge, we propose an optical distributed deep learning (ODDL) architecture. ODDL utilizes a fast yet scalable all-optical network architecture to accelerate distributed training. One of the key features of the architecture is its flow-based transmit scheduling with fast reconfiguration. This allows ODDL to allocate dedicated optical paths for each traffic stream dynamically, resulting in low network latency and high network utilization. Additionally, ODDL provides physically isolated and tailored network resources for training tasks by reconfiguring the optical switch using LCoS-WSS technology. The ODDL topology also uses tunable transceivers to adapt to time-varying traffic patterns. To achieve accurate and fine-grained scheduling of optical circuits, we propose an efficient distributed control scheme that incurs minimal delay overhead. Our evaluation on real-world traces showcases ODDL’s remarkable performance. When implemented with 1024 nodes and 100 Gbps bandwidth, ODDL accelerates VGG19 training by 1.6× and 1.7× compared to conventional fat-tree electrical networks and photonic SiP-Ring architectures, respectively. We further build a four-node testbed, and our experiments show that ODDL can achieve comparable training time compared to that of anidealelectrical switching network.

The 230 GHz lightcurves of Sagittarius A* (Sgr A*) predicted by general relativistic magnetohydrodynamics and general relativistic ray-tracing (GRRT) models by the Event Horizon Telescope Collaboration have higher variabilityM_ΔTcompared to observations. In this series of papers, we explore the origin of such large brightness variability. In this first paper, we performed large GRRT parameter surveys that span from the optically thin to the optically thick regimes, covering the ion-to-electron temperature ratio under strongly magnetized conditions,R_Low, from 1 to 60. We find that increasingR_Lowcan lead to either an increase or a reduction inM_ΔTdepending on the other model parameters, making it consistent with the observed variability of Sgr A* in some cases. Our analysis of GRRT image snapshots finds that the major contribution to the largeM_ΔTfor theR_Low= 1 models comes from the photon ring. However, secondary contributions from the accretion flow are also visible depending on the spin parameter. Our work demonstrates the importance of the electron temperature used for modeling radiatively inefficient accretion flows and places new constraints on the ion-to-electron temperature ratio. A more in-depth analysis for understanding the dependencies ofM_ΔTonR_Lowwill be performed in subsequent papers.

Two of the major factors that control the composition of herbaceous plant communities are competition for limiting soil resources and herbivory. We present results from a 14-year full factorial experiment in a tallgrass prairie ecosystem that crossed nitrogen (N) addition with fencing to exclude white-tailed deer,Odocoileus virginianus, from half the plots. Deer presence was associated with only modest decreases in aboveground plant biomass (14% decrease; −45 ± 19 g m⁻²) with no interaction with N addition. N addition at 5.44 and 9.52 g N m⁻² year⁻¹led to increases in biomass. There were weak increases in species richness associated with deer presence, but only for no or low added N (1 and 2 g N m⁻² year⁻¹). However, the presence of deer greatly impacted the abundances of some of the dominant perennial forb species, but not the dominant grasses. Deer presence increased the abundance of the forbArtemisia ludovicianaby 34 ± 12 SE g m⁻²(94%) and decreased the forbSolidago rigidaby 32 ± 13 SE g m⁻²(79%). We suggest that these changes may have resulted from trade-offs in plant competitive ability for soil N versus resistance to deer herbivory. Field observations suggest deer acted as florivores, mainly consuming the flowers of susceptible forb species. The preferential consumption of flowers of forbs that seem to be superior N competitors appears to create an axis of interspecific niche differentiation. The overpopulation of white-tailed deer in many tallgrass reserves likely structures the abundance of forb species.

It has been suggested that Ba3In2O6 might be a high-Tcsuperconductor. Experimental investigation of the properties of Ba3In2O6 was long inhibited by its instability in air. Recently epitaxial Ba3In2O6 with a protective capping layer was demonstrated, which finally allows its electronic characterization. The optical bandgap of Ba3In2O6 is determined to be 2.99 eV in-the (001) plane and 2.83 eV along the c-axis direction by spectroscopic ellipsometry. First-principles calculations were carried out, yielding a result in good agreement with the experimental value. Various dopants were explored to induce (super-)conductivity in this otherwise insulating material. Neither A- nor B-site doping proved successful. The underlying reason is predominately the formation of oxygen interstitials as revealed by scanning transmission electron microscopy and first-principles calculations. Additional efforts to induce superconductivity were investigated, including surface alkali doping, optical pumping, and hydrogen reduction. To probe liquid-ion gating, Ba3In2O6 was successfully grown epitaxially on an epitaxial SrRuO3 bottom electrode. So far none of these efforts induced superconductivity in Ba3In2O6, leaving the answer to the initial question of whether Ba3In2O6 is a high-Tcsuperconductor to be “no” thus far.

In 1977, Blandford and Znajek showed that the electromagnetic field surrounding a rotating black hole can harvest its spin energy and use it to power a collimated astrophysical jet, such as the one launched from the center of the elliptical galaxy M87. Today, interferometric observations with the Event Horizon Telescope (EHT) are delivering high-resolution, event-horizon-scale, polarimetric images of the supermassive black hole M87* at the jet launching point. These polarimetric images offer an unprecedented window into the electromagnetic field structure around a black hole. In this paper, we show that a simple polarimetric observable—the phase ∠β₂of the second azimuthal Fourier mode of the linear polarization in a near-horizon image—depends on the sign of the electromagnetic energy flux and therefore provides a direct probe of black hole energy extraction. In Boyer–Lindquist coordinates, the Poynting flux for axisymmetric electromagnetic fields is proportional to the productB^ϕB^r. The phase ∠β₂likewise depends on the ratioB^ϕ/B^r, thereby enabling an observer to determine the direction of electromagnetic energy flow in the near-horizon environment experimentally. Data from the 2017 EHT observations of M87* are consistent with electromagnetic energy outflow. Currently envisioned multifrequency observations of M87* will achieve higher dynamic range and angular resolution, and hence deliver measurements of ∠β₂closer to the event horizon as well as better constraints on Faraday rotation. Such observations will enable a definitive test for energy extraction from the black hole M87*.

Our symmetry-free model for spectrum allocation (SA) in networks of general topology leverages two properties: (1) SA is equivalent to a connection permutation problem, and (2) in assigning spectrum, it is sufficient to consider the allocation made by the first-fit (FF) algorithm. This model opens up algorithmic approaches that altogether sidestep spectrum symmetry, i.e., eliminate from consideration the exponential number of equivalent solutions resulting from spectrum slot permutations. Recursive FF (RFF) is such an algorithm; it applies FF recursively to search the connection permutation space and solve the SA problem optimally. Moreover, parallelism is inherent in the spectrum symmetry-free model, as the connection permutation space may be naturally decomposed into non-overlapping subsets that can be searched independently. Accordingly, RFF admits multi-threaded implementations that may be tailored to the computing environment at hand. In this work, we present two strategies for parallelizing the execution of RFF, and we evaluate them experimentally using a comprehensive set of metrics. Our experiments indicate that RFF explores a vast number of symmetry-free solutions, and for moderate-sized networks, it takes mere seconds to yield solutions that are either optimal or very close to the lower bound.

Epitaxial untwinned SrRuO3 thin films were grown on (110)-oriented DyScO3 substrates by molecular-beam epitaxy. We report an exceptional sample with a residual resistivity ratio (RRR), ρ [300 K]/ρ [4 K] of 205 and a ferromagnetic Curie temperature, TC, of 168.3 K. We compare the properties of this sample to other SrRuO3 films grown on DyScO3(110) with RRRs ranging from 8.8 to 205, and also compare it to the best reported bulk single crystal of SrRuO3. We determine that SrRuO3 thin films grown on DyScO3(110) have an enhanced TC as long as the RRR of the thin film is above a minimum electrical quality threshold. This RRR threshold is about 20 for SrRuO3. Films with lower RRR exhibit TCs that are significantly depressed from the intrinsic strain-enhanced value.

Images of supermassive black hole accretion flows contain features of both curved spacetime and plasma structure. Inferring properties of the spacetime from images requires modeling the plasma properties, and vice versa. The Event Horizon Telescope Collaboration has imaged near-horizon millimeter emission from both Messier 87* (M87*) and Sagittarius A* (Sgr A*) with very long baseline interferometry (VLBI) and has found a preference for magnetically arrested disk (MAD) accretion in each case. MAD accretion enables spacetime measurements through future observations of the photon ring, the image feature composed of near-orbiting photons. The ordered fields and relatively weak Faraday rotation of MADs yield rotationally symmetric polarization when viewed at modest inclination. In this letter, we utilize this symmetry along with parallel transport symmetries to construct a gain-robust interferometric quantity that detects the transition between the weakly lensed accretion flow image and the strongly lensed photon ring. We predict a shift in polarimetric phases on long baselines and demonstrate that the photon rings in M87* and Sgr A* can be unambiguously detected with sensitive, long-baseline measurements. For M87*, we find that photon ring detection in snapshot observations requires ∼1 mJy sensitivity on >15 Gλbaselines at 230 GHz and above, which could be achieved with space-VLBI or higher-frequency ground-based VLBI. For Sgr A*, we find that interstellar scattering inhibits photon ring detectability at 230 GHz, but ∼10 mJy sensitivity on >12 Gλbaselines at 345 GHz is sufficient and is accessible from the ground. For both sources, these sensitivity requirements may be relaxed by repeated observations and averaging.