arXiv subjects

Sanjoli Narang

Publications and source records attributed to Sanjoli Narang.

5 recordsLinked to original sources

arXiv2026

MLCC: A Congestion Control Technique to Accelerate ML Training

We present MLCC, a novel technique to augment today's congestion control algorithms to accelerate DNN training jobs in shared GPU clusters in a fully distributed manner. At the heart of MLCC lies a straightforward principle: DNN training flows should scale their sending rate to shift other flows' communication into their compute periods, achieving interleaving. We show that integrating this principle into today's congestion control protocols is simple (requiring less than 60 lines of code for a given protocol) and enables DNN jobs to interleave within a few training iterations, thereby reducing network contention and improving job completion times. Our testbed demonstrates that MLCC accelerates the average and 99th percentile training iteration times by up to 1.9x and 2.7x respectively. Through extensive packet-level simulations, we observe a 1.35x improvement in training throughput on a 36-node, 288 GPU fat-tree topology.

cs.NI↗

arXiv2026

Bring Your Own Objective: Inter-operability of Network Objectives in Datacenters

Datacenter networks are currently locked in a "tyranny of the single objective". While modern workloads demand diverse performance goals, ranging from coflow completion times, per-flow fairness, short-flow latencies, existing fabrics are typically hardcoded for a single metric. This rigid coupling ensures peak performance when application and network objectives align, but results in abysmal performance when they diverge. We propose DMart, a decentralized scheduling framework that treats network bandwidth as a competitive marketplace. In DMart, applications independently encode the urgency and importance of their network traffic into autonomous bids, allowing diverse objectives to co-exist natively on the same fabric. To meet the extreme scale and sub-microsecond requirements of modern datacenters, DMart implements distributed, per-link, per-RTT auctions, without relying on ILPs, centralized schedulers, or complex priority queues. We evaluate DMart using packet-level simulations and compare it against network schedulers designed for individual metrics, e.g., pFabric and Sincronia. DMart matches the performance of specialized schedulers on their own "home turf" while simultaneously optimizing secondary metrics. Compared to pFabric and Sincronia, DMart reduces deadline misses by 2x and coflow completion times by 1.6x respectively, while matching pFabric short-flow completion times.

cs.NI↗

arXiv2025

Development of a Compton Imager Setup

Hard X-ray photons with energies in the range of hundreds of keV typically undergo Compton scattering when they are incident on a detector. In this process, an incident photon deposits a fraction of its energy at the point of incidence and continues onwards with a change in direction that depends on the amount of energy deposited. By using a pair of detectors to detect the point of incidence and the direction of the scattered photon, we can calculate the scattering direction and angle. The position of a source in the sky can be reconstructed using many Compton photon pairs from a source. We demonstrate this principle in the laboratory by using a pair of Cadmium Zinc Telluride (CZT) detectors sensitive in the energy range of 20-200 keV, similar to those used in $\textit{AstroSat}$/CZT Imager (CZTI). The laboratory setup consists of the two detectors placed perpendicular to each other in a lead-lined box. The detectors are read out by a custom-programmed Xilinx PYNQ-Z2 FPGA board, and data are then transferred to a personal computer (PC)}. There are two key updates from CZTI: the detectors are read concurrently rather than serially, and the time resolution has been improved from $20~μ$s to $7.5~μ$s. We irradiated the detectors with a collimated $^{133}\mathrm{Ba}$ source and identified Compton scattering events for the 356 keV line. We run a Compton reconstruction algorithm to correctly infer the location of the source in the detector frame, with a location-dependent angular response measure of $16°-30°$. This comprises a successful technology demonstration for a Compton imaging camera in the hard X-ray regime. We present the details of our setup, the data acquisition process, and software algorithms, and showcase our results. We also quantify the limitations of this setup and discuss ways of improving the performance in future experiments.

astro-ph.IM↗

arXiv2024

Science with the Daksha High Energy Transients Mission

We present the science case for the proposed Daksha high energy transients mission. Daksha will comprise of two satellites covering the entire sky from 1~keV to $>1$~MeV. The primary objectives of the mission are to discover and characterize electromagnetic counterparts to gravitational wave source; and to study Gamma Ray Bursts (GRBs). Daksha is a versatile all-sky monitor that can address a wide variety of science cases. With its broadband spectral response, high sensitivity, and continuous all-sky coverage, it will discover fainter and rarer sources than any other existing or proposed mission. Daksha can make key strides in GRB research with polarization studies, prompt soft spectroscopy, and fine time-resolved spectral studies. Daksha will provide continuous monitoring of X-ray pulsars. It will detect magnetar outbursts and high energy counterparts to Fast Radio Bursts. Using Earth occultation to measure source fluxes, the two satellites together will obtain daily flux measurements of bright hard X-ray sources including active galactic nuclei, X-ray binaries, and slow transients like Novae. Correlation studies between the two satellites can be used to probe primordial black holes through lensing. Daksha will have a set of detectors continuously pointing towards the Sun, providing excellent hard X-ray monitoring data. Closer to home, the high sensitivity and time resolution of Daksha can be leveraged for the characterization of Terrestrial Gamma-ray Flashes.

astro-ph.HE↗

arXiv2024

Daksha: On Alert for High Energy Transients

We present Daksha, a proposed high energy transients mission for the study of electromagnetic counterparts of gravitational wave sources, and gamma ray bursts. Daksha will comprise of two satellites in low earth equatorial orbits, on opposite sides of earth. Each satellite will carry three types of detectors to cover the entire sky in an energy range from 1 keV to >1 MeV. Any transients detected on-board will be announced publicly within minutes of discovery. All photon data will be downloaded in ground station passes to obtain source positions, spectra, and light curves. In addition, Daksha will address a wide range of science cases including monitoring X-ray pulsars, studies of magnetars, solar flares, searches for fast radio burst counterparts, routine monitoring of bright persistent high energy sources, terrestrial gamma-ray flashes, and probing primordial black hole abundances through lensing. In this paper, we discuss the technical capabilities of Daksha, while the detailed science case is discussed in a separate paper.

astro-ph.IM↗