Jun Duan

Jun Duan

Auto-generated profile, based on publicly available information

Jun Duan

IBM Thomas J. Watson Research Center, USA

Research Focus & Metrics

Academic Career Length

11Years (2016-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 January 2021 in IEEE Transactions on Mobile Computing
Verified
Journal Article
Closed access
YHYaodong Huang
Yaodong Huang
JZJiarui Zhang
Jiarui Zhang
JDJun Duan
Jun Duan
BXBin XIAO
Bin XIAO
Fan YeFan Ye
Fan Ye
YYYuanyuan Yang
Yuanyuan Yang
Edge devices with sensing, storage, and communication resources are penetrating our daily lives. These resources make it possible for edge devices to conduct data transactions (e.g., micro-payments, micro-access control). The blockchain technology can be used to ensure transaction unmodifiable and undeniable. In this paper, we propose a blockchain system that adapts to the limitations of edge devices. The new blockchain system can fairly and efficiently allocate storage resources on edge devices, which makes it scalable. We find the optimal peer nodes for transaction data storage and propose a recent block storage allocation scheme for quick retrieval of missing blocks. We develop data migration algorithms to dynamically reallocate data and block storage to adapt topology changes in the network. The proposed blockchain system can also reach consensus with low energy consumption in edge devices with a new Proof of Stake mechanism. Extensive simulations show that our proposed blockchain system works efficiently in edge environments. On average, the new system uses 18.4 percent less time and consumes 87 percent less battery power when compared with traditional blockchain systems.

21 March 2018 in IEEE Transactions on Parallel and Distributed Systems
Verified
Journal Article
Closed access
JDJun Duan
Jun Duan
YYYuanyuan Yang
Yuanyuan Yang
Interconnection networks lie in the heart of all types of parallel architectures, because how processors or memory modules are connected to each other has a significant impact on the scalability, reliability, cost and performance. For example, a nonblocking interconnection network delivers guaranteed path availability to any connection requests, without interference to existing connections. Also, an interconnection network with multicast capability can distribute data from a single source to all the destinations in a one-shot manner, eliminating unnecessary duplications and minimizing communication delays. However, implementing nonblocking multicast networks imposes great challenges to designers because both of them demand high hardware cost. To deal with this problem, in this paper we propose a novel interconnection network, named Multicast Capable Low-cost Network (MCL), which is both nonblocking for multicast traffic and cost-efficient. We first design the topologies and routing algorithms for MCL, and then prove its nonblocking multicast properties. Most importantly, we show that MCL achieves the lowest hardware cost in terms of asymptotic number of crosspoints. Specifically, the theoretical upper bound on the crosspoints of an $N \times N$ MCL is $O(N\;\text{log}^{3.39}\;N)$). The explicitly constructed instance of an $N \times N$ MCL has a cost of $O(N^{\frac{5}{4}})$ and constant delay from the input ports to the output ports. These hardware costs are the lowest compared to previous schemes which deliver the same performance.

11 December 2023
Verified
Conference Paper
Open access
ZXZhengxu Xia
Zhengxu Xia
YHYitian Hao
Yitian Hao
JDJun Duan
Jun Duan
CWChen Wang
Chen Wang
JJJunchen Jiang
Junchen Jiang
With GPUs increasingly shared by DNN models at the edge, a crucial tradeoff arises between high GPU utilization and the ability of fast preemption when a high-priority request arrives. To reduce inference delay, an inference job can "burst" DNN kernels into the GPU to maintain high GPU utilization, but this also creates outstanding kernels internally queued in the GPU, causing a substantial preemption delay as the GPU must clear the queued kernels before the high-priority request can preempt. Unfortunately, while existing systems can alleviate the preemption delay by adding synchronization points, they fail to keep both inference delay and preemption delay low, because they cannot optimally insert the synchronization points for various workloads. Our measurements and analysis show that the impact of inserting synchronization points on the inference and preemption delays varies greatly with a range of workload characteristics, including the DNN architecture, input size, batch size of different requests, and GPU type, but most of these workload factors are overlooked in current shared-edge systems. Inspired by these findings, we make a case for a new module in shared edge systems to dynamically insert kernel synchronization points depending on the workload characteristics and service-level objective (SLO) deadlines. To examine its potential, we present Deft, a concrete prototype of concept to share a GPU for multiple DNN containers. The key component of Deft is a profiling-based delay predictor that estimates the impact of synchronization points on the inference delay and preemption delay and dynamically selects a frequency of inserting synchronization points that minimizes inference delay or SLO violations. Compared to the state-of-the-art GPU sharing schemes, Deft reduces inference delay and SLO violations by 28% and 14% respectively. While we intentionally keep Deft's design simple, it already shows early promise of adding synchronization points dynamically and highlights key questions for future research.

1 June 2016
Verified
Conference Paper
Closed access
JDJun Duan
Jun Duan
YYYuanyuan Yang
Yuanyuan Yang
Interconnects play a critical role in various networking environments including data center networks, high performance computing systems, networks-on-chip, etc. An important design concern of interconnects is the hardware cost, especially when the scale increases. In order to achieve lower hardware cost, a lot of previous work leverages a tradeoff between the cost and the performance. In this paper, we alternatively propose a new type of recursively constructed multi-stage interconnect which is both cost efficient and performance guaranteed. The proposed interconnect can deliver the same communication pattern support and path availability as previous designs, yet at the same time achieves lower hardware complexity. We first design a prototype (or base case) for the new interconnect, which consists of three stages of switching modules. Then we use this prototype to decompose and replace the switching modules repeatedly and obtain the recursive case of our interconnect. The hardware cost of the recursively constructed interconnect can be significantly reduced, meanwhile the communication pattern support and path availability are perfectly preserved. For both the base case and the recursive case, we study the properties of the network topology, propose routing algorithms, give performance analysis and calculate cost. We show that the interconnect reduces the hardware cost from the previously best known result O(N3/2), where N is the number of input or output nodes of the interconnect, to O(N11/1-2δ log 1.585 N), where δ is a small value approaching zero when N increases.

24 February 2019 in IEEE Transactions on Cloud Computing
Verified
Journal Article
Closed access
JDJun Duan
Jun Duan
YYYuanyuan Yang
Yuanyuan Yang
In this paper, we propose a versatile network architecture, named VFTree, for data centers. VFTree provisions unique versatility in the sense that its topology and performance can be tailored with great freedom. The wide dynamic range and fine granularity of its configurations allow this architecture to closely follow the demands of the running traffic. Also, this versatility facilitates the reuse of network devices, so that upgrades could be achieved by rearranging existing devices, avoiding the expenditure on new generations of hardware in each upgrade cycle. Furthermore, this versatility allows VFTree to be elaborately tuned into cost efficient zones. This way, compared to widely adopted industrial solutions including fat-tree, VFTree can provide even higher aggregate bandwidth and richer path availability at the same cost. In this paper, we first design the network topology of VFTree. We parameterize the topology so that it is flexible and generalized. For example, we can configure the numbers of upward and downward ports for the switches, which are fixed in the classic fat-tree. Based on that, we propose routing algorithms, implement the abstraction from packet level to the flow level, and handle the flow collisions. We also analyze VFTree’s performance, demonstrate its aforementioned features, and compare it with its predecessors. VFTree can be seen as a generalization of, and is backward-compatible to fat-tree. We believe it is ready to be deployed in industry, as a perfect replacement of the fat-tree DCNs.

22 October 2019
Verified
Book Chapter
Closed access
PDF file
JDJun Duan
Jun Duan
GLGuangcheng Li
Guangcheng Li
NANeeraj Asthana
Neeraj Asthana
SZSai Zeng
Sai Zeng
IDIvan Dell’eRa
Ivan Dell’eRa
RYRong Yan
Rong Yan