arXiv subjects

Peifeng Yu

Publications and source records attributed to Peifeng Yu.

23 records · Page 2Linked to original sources

arXiv2022

Orloj: Predictably Serving Unpredictable DNNs

Existing DNN serving solutions can provide tight latency SLOs while maintaining high throughput via careful scheduling of incoming requests, whose execution times are assumed to be highly predictable and data-independent. However, inference requests to emerging dynamic DNNs -- e.g., popular natural language processing (NLP) models and computer vision (CV) models that skip layers -- are data-dependent. They exhibit poor performance when served using existing solutions because they experience large variance in request execution times depending on the input -- the longest request in a batch inflates the execution times of the smaller ones, causing SLO misses in the absence of careful batching. In this paper, we present Orloj, a dynamic DNN serving system, that captures this variance in dynamic DNNs using empirical distributions of expected request execution times, and then efficiently batches and schedules them without knowing a request's precise execution time. Orloj significantly outperforms state-of-the-art serving solutions for high variance dynamic DNN workloads by 51--80% in finish rate under tight SLO constraints, and over 100% under more relaxed SLO settings. For well-studied static DNN workloads, Orloj keeps comparable performance with the state-of-the-art.

cs.DC↗

arXiv2022

Negative Chemical Pressure Effect on Superconductivity and Charge Density Wave of Cu0.5Ir1-xZrxTe2

This study demonstrates the design and synthesis of Cu0.5Ir1-xZrxTe2 system by partial substitution of Ir with Zr acting as a negative chemical pressure. With the doping of Zr, the cell parameters significantly expand, signifying an effective negative chemical pressure. The experimental results found evidence that the charge density wave (CDW)-like order is immediately quenched by subtle Zr substitution for Ir and a classical dome-shape Tc(x) that peaked at 2.80 K can be observed. The optimal Cu0.5Ir0.95Zr0.05Te2 compound is a BCS-type superconductor and exhibits type-II SC. However, high Zr concentration can provoke disorder, inducing the reappearance of CDW order. The present study shows that the Cu0.5Ir1-xZrxTe2 system may provide a new platform for further understanding of multiple electronic orders in transition metal dichalcogenides.

cond-mat.supr-con↗

arXiv2022

Superconductivity with the enhanced upper critical field in the Pt-Doping CuRh2Se4 spinel

We report the effect of Pt doping on the superconductivity in CuRh2Se4 spinel using a combined experimental and theoretical study. Our XRD results reveal that the Cu(Rh1-xPtx)2Se4 crystallizes in the structure with a space group of Fd3-m (No. 227), and the lattice parameter a increases with Pt doping. The resistivity and magnetic susceptibility measurement results verify that the superconducting transition temperature (Tc) forms a dome-like shape with a maximum value of 3.84 K at x = 0.06. It is also observed that the Pt-doping slightly reduces the lower critical magnetic field from 220 Oe in CuRh2Se4 to 168 Oe in Cu(Rh0.94Pt0.06)2Se4, while it significantly enhances the upper critical magnetic field, reaching the maximum of 4.93 T in the Cu(Rh0.94Pt0.06)2Se4 sample. The heat capacity result indicates that the sample Cu(Rh0.91Pt0.09)2Se4 is a bulk superconductor. First-principles calculations suggest that the Pt-doping leads to a red-shift of a density of state peak near the Fermi level, consistent with the dome-like Tc observed experimentally.

cond-mat.supr-con↗

arXiv2021

Enhanced superconductivity with possible re-appearance of charge density wave states in polycrystalline Cu$_{1-x}$Ag$_x$Ir$_2$Te$_4$ alloys

In this study, we determined the effect of doping with the noble metal Ag on Ir$_2$Te$_4$ superconductors. Based on the resistivity, magnetization, and heat capacity, we explored the changes in the superconductivity and charge density wave for Cu$_{1-x}$Ag$_x$Ir$_2$Te$_4$ as a function of isoelectric substitution. We assessed a complete set of competing states from suppressed charge density wave in the low doping region to superconductor in the middle doping region and re-entrant charge density wave in the high doping region, thereby obtaining an electronic phase diagram, where the superconducting dome was near bipartite charge density wave regions with a maximum superconducting temperature Tc of about 2.93 K at an Ag doping level of 12%. The lower H$_{c1}$ and upper H$_{c2}$ critical magnetic fields were determined for some representative samples in the Cu$_{1-x}$Ag$_x$Ir$_2$Te$_4$ series based on magnetization and resistivity measurements, respectively. We showed that H$_{c1}$ decreased whereas H$_{c2}$ increased as the doping content increased. The specific heat anomalies at the superconducting transitions $ΔC_{el}/γT_c$ for representative samples comprising Cu$_{0.92}$Ag$_{0.08}$Ir$_2$Te$_4$, Cu$_{0.88}$Ag$_{0.12}$Ir$_2$Te$_4$, and Cu$_{0.85}$Ag$_{0.15}$Ir$_2$Te$_4$ were approximately 1.40, 1.44, and 1.42, respectively, which are all near the Bardeen-Cooper-Schrieffer value of 1.43 and they indicate bulk superconductivity in these compounds.

cond-mat.supr-con↗

arXiv2019

Salus: Fine-Grained GPU Sharing Primitives for Deep Learning Applications

GPU computing is becoming increasingly more popular with the proliferation of deep learning (DL) applications. However, unlike traditional resources such as CPU or the network, modern GPUs do not natively support fine-grained sharing primitives. Consequently, implementing common policies such as time sharing and preemption are expensive. Worse, when a DL application cannot completely use a GPU's resources, the GPU cannot be efficiently shared between multiple applications, leading to GPU underutilization. We present Salus to enable two GPU sharing primitives: fast job switching and memory sharing, in order to achieve fine-grained GPU sharing among multiple DL applications. Salus implements an efficient, consolidated execution service that exposes the GPU to different DL applications, and enforces fine-grained sharing by performing iteration scheduling and addressing associated memory management issues. We show that these primitives can then be used to implement flexible sharing policies such as fairness, prioritization, and packing for various use cases. Our integration of Salus with TensorFlow and evaluation on popular DL jobs show that Salus can improve the average completion time of DL training jobs by $3.19\times$, GPU utilization for hyper-parameter tuning by $2.38\times$, and GPU utilization of DL inference applications by $42\times$ over not sharing the GPU and $7\times$ over NVIDIA MPS with small overhead.

cs.DC↗