Xinle Du
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Xinle Du has a total of 41 co-authors. Here are the top co-authors by number of shared publications:
- Ke Xu
Tsinghua University, Beijing, China
5 shared publications - Jingbin Zhou
HUAWEI, Shenzhen, China
3 shared publications - Qi Li
Tsinghua University, Beijing, China
3 shared publications - Kun Tan
HUAWEI, Shenzhen, China
2 shared publications - Shuihai Hu
HUAWEI, Shenzhen, China
2 shared publications - Songtao Fu
Department of Computer Science and Technology, Beijing National Research Center for Information Science and Technology (BNRist), Tsinghua University, Beijing, China
2 shared publications
Latest Publications
- Open Access
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 612-627
Concord: Airtime-Aware Contention Control for Taming Tail Latency from Wi-Fi Frame Bursting
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 612-627
Open AccessConcord: Airtime-Aware Contention Control for Taming Tail Latency from Wi-Fi Frame Bursting
- Open Access
- 24 November 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 3 (CoNEXT4) , 1-20
Harmonics: Scalable Collective Scheduling in Multi-Tenant GPU Clusters
Distributed machine learning (DML), such as large language model (LLM) training, has become one of the most critical services in multi-tenant cloud computing. However, communication contention among concurrent DML jobs significantly degrades overall GPU utilization, leading to inefficient training cluster performance. Existing approaches either achieve high performance at the cost of long scheduling runtime or reduce scheduling time at the expense of poor performance. We present Harmonics, a novel two-tier scheduling framework that strikes a balance between scheduling latency and performance. It coordinates decisions between Local Schedulers and a lightweight Global Coordinator to enable scalable and adaptive scheduling. By combining rack-level epoch-based optimization with global coordination, Harmonics alleviates communication contention and improves resource efficiency. We implement and evaluate Harmonics on real distributed ML workloads running on a GPU testbed. Compared to state-of-the-art methods such as fair sharing, optimal scheduling, Crux, and Cassini, Harmonics reduces training time by up to 33% and communication time by up to 48%. Large-scale simulations show that it reduces scheduling time by up to 91× while improving training time by 26% in large-cluster settings.
- 24 November 2025
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 3 (CoNEXT4) , 1-20
Open AccessHarmonics: Scalable Collective Scheduling in Multi-Tenant GPU Clusters
Distributed machine learning (DML), such as large language model (LLM) training, has become one of the most critical services in multi-tenant cloud computing. However, communication contention among concurrent DML jobs significantly degrades overall GPU utilization, leading to inefficient training cluster performance. Existing approaches either achieve high performance at the cost of long scheduling runtime or reduce scheduling time at the expense of poor performance. We present Harmonics, a novel two-tier scheduling framework that strikes a balance between scheduling latency and performance. It coordinates decisions between Local Schedulers and a lightweight Global Coordinator to enable scalable and adaptive scheduling. By combining rack-level epoch-based optimization with global coordination, Harmonics alleviates communication contention and improves resource efficiency. We implement and evaluate Harmonics on real distributed ML workloads running on a GPU testbed. Compared to state-of-the-art methods such as fair sharing, optimal scheduling, Crux, and Cassini, Harmonics reduces training time by up to 33% and communication time by up to 48%. Large-scale simulations show that it reduces scheduling time by up to 91× while improving training time by 26% in large-cluster settings.
- Open Access
- 6 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 172-177
FlexSpark: Robust and Efficient Multi-Device Collaborative Inference over Wireless Network
- 6 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 172-177
Open AccessFlexSpark: Robust and Efficient Multi-Device Collaborative Inference over Wireless Network
- Open Access
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 7891-7899
Toward Timeliness-Enhanced Loss Recovery for Large-Scale Live Streaming
Due to the limited permissions for upgrading dual-side (i.e., server-side and client-side) loss tolerance schemes from the perspective of CDN vendors in a multi-supplier market, modern large-scale live streaming services are still using the automatic-repeat-request (ARQ) based paradigm for loss recovery, which only requires server-side modifications. In this paper, we first conduct a large-scale measurement study with up to 50 million live streams. We find that loss shows dynamics and live streaming contains frequent on-off mode switching in the wild. We further find that the recovery latency, enlarged by the ubiquitous retransmission loss, is a critical factor affecting live streaming's client side QoE (e.g., video freezing). We then propose an enhanced recovery mechanism called AutoRec, which can transform the disadvantages of on-off mode switching into an advantage for reducing loss recovery latency without any modifications on the client side. AutoRec also adopts an online learning-based policy to fit the dynamics of loss, balancing the tradeoff between the recovery latency and the incurred overhead. We implement AutoRec upon QUIC and evaluate it via both testbed and real-world commercial services deployments. The experimental results demonstrate the practicability and profitability of AutoRec, in which the average times and duration of client-side video freezing can be lowered by 11.4% and 5.2%, respectively.
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 7891-7899
Open AccessToward Timeliness-Enhanced Loss Recovery for Large-Scale Live Streaming
Due to the limited permissions for upgrading dual-side (i.e., server-side and client-side) loss tolerance schemes from the perspective of CDN vendors in a multi-supplier market, modern large-scale live streaming services are still using the automatic-repeat-request (ARQ) based paradigm for loss recovery, which only requires server-side modifications. In this paper, we first conduct a large-scale measurement study with up to 50 million live streams. We find that loss shows dynamics and live streaming contains frequent on-off mode switching in the wild. We further find that the recovery latency, enlarged by the ubiquitous retransmission loss, is a critical factor affecting live streaming's client side QoE (e.g., video freezing). We then propose an enhanced recovery mechanism called AutoRec, which can transform the disadvantages of on-off mode switching into an advantage for reducing loss recovery latency without any modifications on the client side. AutoRec also adopts an online learning-based policy to fit the dynamics of loss, balancing the tradeoff between the recovery latency and the incurred overhead. We implement AutoRec upon QUIC and evaluate it via both testbed and real-world commercial services deployments. The experimental results demonstrate the practicability and profitability of AutoRec, in which the average times and duration of client-side video freezing can be lowered by 11.4% and 5.2%, respectively.
- 26 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Networking
- Vol. 32 (6) , 5310-5325
Stable Byzantine Fault Tolerance in Wide Area Networks With Unreliable Links
With the increasing demand for blockchain technology in various industry sectors, there has been a growing interest in the Byzantine Fault Tolerance (BFT) consensus that is the backbone of most of these blockchains. However, many state-of-the-art algorithms that require reliable connections can only offer limited throughput in wide-area networks (WANs), where participants are connected over long distances and may experience unpredictable network failures. The partially-connected BFTs are designed for unreliable and highly dynamic networks yet impose exponential communication complexity. This paper proposes Stable Byzantine Fault Tolerance (SBFT), a BFT communication abstraction that can sustain high throughput and low latency in WAN. SBFT separates the leader from consensus in pipelined BFT consensus and uses an adaptive consensus mechanism to resist dynamic faulty links, maintaining consensus efficiency when network connectivity is high while adapting to dynamic networks with low connectivity. We implemented a prototype of SBFT and tested it on the WAN. The results demonstrate that SBFT has a throughput similar to HotStuff in a fault-free environment but can reduce about 80% of consensus latency. Besides, SBFT retains 40% of the original throughput when the link failure probability is $0.4$ , while the baseline HotStuff retains less than 40% when the link failure probability is only $0.1$ .
- 26 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Networking
- Vol. 32 (6) , 5310-5325
Stable Byzantine Fault Tolerance in Wide Area Networks With Unreliable Links
With the increasing demand for blockchain technology in various industry sectors, there has been a growing interest in the Byzantine Fault Tolerance (BFT) consensus that is the backbone of most of these blockchains. However, many state-of-the-art algorithms that require reliable connections can only offer limited throughput in wide-area networks (WANs), where participants are connected over long distances and may experience unpredictable network failures. The partially-connected BFTs are designed for unreliable and highly dynamic networks yet impose exponential communication complexity. This paper proposes Stable Byzantine Fault Tolerance (SBFT), a BFT communication abstraction that can sustain high throughput and low latency in WAN. SBFT separates the leader from consensus in pipelined BFT consensus and uses an adaptive consensus mechanism to resist dynamic faulty links, maintaining consensus efficiency when network connectivity is high while adapting to dynamic networks with low connectivity. We implemented a prototype of SBFT and tested it on the WAN. The results demonstrate that SBFT has a throughput similar to HotStuff in a fault-free environment but can reduce about 80% of consensus latency. Besides, SBFT retains 40% of the original throughput when the link failure probability is $0.4$ , while the baseline HotStuff retains less than 40% when the link failure probability is only $0.1$ .
- Open Access
- 3 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 88-94
Revisiting Congestion Control for WiFi Networks
WiFi networks, widely utilized by wireless devices, have become increasingly complex and congested environments, leading to noticeable delays, jitter, and throughput degradation for end-to-end network flows in today’s Internet. Through detailed experimental observations, we identified the TCP victim problem in WiFi, where TCP erroneously detects congestion and interacts with WiFi routers, resulting in a significant throughput decrease for certain hosts. In this paper, we introduce Cupid, a novel congestion control algorithm that relies on receiver-side WiFi physical layer measurements. Cupid accurately assesses congestion by measuring parameters such as airtime utilization, concurrency, and rates, allowing for precise rate adjustments. Our results demonstrate that whether employed alone or alongside other congestion controls, Cupid effectively mitigates the TCP victim problem. Furthermore, when used independently, Cupid can also reduce latency.
- 3 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 88-94
Open AccessRevisiting Congestion Control for WiFi Networks
WiFi networks, widely utilized by wireless devices, have become increasingly complex and congested environments, leading to noticeable delays, jitter, and throughput degradation for end-to-end network flows in today’s Internet. Through detailed experimental observations, we identified the TCP victim problem in WiFi, where TCP erroneously detects congestion and interacts with WiFi routers, resulting in a significant throughput decrease for certain hosts. In this paper, we introduce Cupid, a novel congestion control algorithm that relies on receiver-side WiFi physical layer measurements. Cupid accurately assesses congestion by measuring parameters such as airtime utilization, concurrency, and rates, allowing for precise rate adjustments. Our results demonstrate that whether employed alone or alongside other congestion controls, Cupid effectively mitigates the TCP victim problem. Furthermore, when used independently, Cupid can also reduce latency.
- 22 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 22 (1) , 205-220
Secure Fault Localization in Path Aware Networking
Secure data forwarding is critical for users to meet their requirements. In this paper, we propose D3 (Demon Detector in Data Plane), a source-driven, secure fault localization mechanism, which empowers the source to localize faulty link in Path Aware Networking, thus circumventing faulty link to guarantee secure data forwarding. D3 utilizes the source to instruct the on-path routers, thus empowering it to detect whether the on-path routers forward the packet as expected. Compared with existing schemes that are difficult to be deployed in practice due to the heavy storage, computation, and communication overhead, D3 offloads most of the on-path router's storage and computation overhead, thus dramatically improving the deployment efficiency. Particularly, the length of the additional packet header in D3 is 2-5 times less than the state-of-the-art mechanisms, thus having a low communication overhead. Besides that, the destination in D3 could keep stateless processing, thus having backward compatibility and eliminating the opportunity for DoS attacks toward a stateful destination. The BMv2 and Barefoot Tofino hardware evaluations show that D3 could achieve high fault localization accuracy and process the packet at line rate.
- 22 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dependable and Secure Computing
- Vol. 22 (1) , 205-220
Secure Fault Localization in Path Aware Networking
Secure data forwarding is critical for users to meet their requirements. In this paper, we propose D3 (Demon Detector in Data Plane), a source-driven, secure fault localization mechanism, which empowers the source to localize faulty link in Path Aware Networking, thus circumventing faulty link to guarantee secure data forwarding. D3 utilizes the source to instruct the on-path routers, thus empowering it to detect whether the on-path routers forward the packet as expected. Compared with existing schemes that are difficult to be deployed in practice due to the heavy storage, computation, and communication overhead, D3 offloads most of the on-path router's storage and computation overhead, thus dramatically improving the deployment efficiency. Particularly, the length of the additional packet header in D3 is 2-5 times less than the state-of-the-art mechanisms, thus having a low communication overhead. Besides that, the destination in D3 could keep stateless processing, thus having backward compatibility and eliminating the opportunity for DoS attacks toward a stateful destination. The BMv2 and Barefoot Tofino hardware evaluations show that D3 could achieve high fault localization accuracy and process the packet at line rate.
- 24 November 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Networking
- Vol. 31 (4) , 1478-1493
MASK: Practical Source and Path Verification Based on Multi-AS-Key
The source and path verification in Path-Aware Networking considers the two critical issues: (1) end hosts could verify that the network follows their forwarding decisions, and (2) both on-path routers and destination host could authenticate the source of packets and filter the malicious traffic. Unfortunately, the state-of-the-art mechanisms require heavy communication overhead in the network and computation overhead in the router; moreover, it is difficult to meet the dynamic requirements of the end host. We propose a user-driven mechanism, source and path verification based on Multi-AS-Key (MASK). MASK decreases the communication overhead by a short additional packet header and reduces the computation overhead by separating the control and data plane in terms of the cryptographic operation. Furthermore, it utilizes the stateful user to instruct the stateless routers to process the packet with a user-driven policy, thus satisfying the user’s requirements such as detecting the packet drop and replay attack. With the plausible design, the communication overhead for realistic path lengths is 1/2 to 1/10 compared with the state-of-the-art mechanisms. We implement MASK in the BMv2 environment and commodity Barefoot Tofino programmable switch, testify that MASK introduces significantly less overhead than the state-of-the-art mechanisms, and demonstrate that MASK could achieve the verification in the programmable switch at line rate.
- 24 November 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Networking
- Vol. 31 (4) , 1478-1493
MASK: Practical Source and Path Verification Based on Multi-AS-Key
The source and path verification in Path-Aware Networking considers the two critical issues: (1) end hosts could verify that the network follows their forwarding decisions, and (2) both on-path routers and destination host could authenticate the source of packets and filter the malicious traffic. Unfortunately, the state-of-the-art mechanisms require heavy communication overhead in the network and computation overhead in the router; moreover, it is difficult to meet the dynamic requirements of the end host. We propose a user-driven mechanism, source and path verification based on Multi-AS-Key (MASK). MASK decreases the communication overhead by a short additional packet header and reduces the computation overhead by separating the control and data plane in terms of the cryptographic operation. Furthermore, it utilizes the stateful user to instruct the stateless routers to process the packet with a user-driven policy, thus satisfying the user’s requirements such as detecting the packet drop and replay attack. With the plausible design, the communication overhead for realistic path lengths is 1/2 to 1/10 compared with the state-of-the-art mechanisms. We implement MASK in the BMv2 environment and commodity Barefoot Tofino programmable switch, testify that MASK introduces significantly less overhead than the state-of-the-art mechanisms, and demonstrate that MASK could achieve the verification in the programmable switch at line rate.
- 23 August 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Networking
- Vol. 31 (2) , 526-541
R-AQM: Reverse ACK Active Queue Management in Multitenant Data Centers
TCP incast has become a practical problem for high-bandwidth, low-latency transmissions, resulting in throughput degradation of up to 90% and delays of hundreds of milliseconds, severely impacting application performance. However, in virtualized multi-tenant data centers, host-based advancements in the TCP stack are hard to deploy from the operators’ perspective. Operators only provide infrastructure in the form of virtual machines, in which only tenants can directly modify the end-host TCP stack. In this paper, we present R-AQM, a switch-powered reverse ACK active queue management (R-AQM) mechanism for enhancing ACK-clocking effects through assisting legacy TCP. Specifically, R-AQM proactively intercepts ACKs and paces the ACK-clocked in-flight data packets, preventing TCP from suffering incast collapse. We implement and evaluate R-AQM in NS-3 simulation and NetFPGA-based hardware switch. Both simulation and testbed results show that R-AQM greatly improves TCP performance under heavy incast workloads by significantly lowering packet loss rate, reducing retransmission timeouts, and supporting 16 times (i.e., 60 to 1000) more senders. Meanwhile, the forward queuing delays are also reduced by 4.6 times.
- 23 August 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Networking
- Vol. 31 (2) , 526-541
R-AQM: Reverse ACK Active Queue Management in Multitenant Data Centers
TCP incast has become a practical problem for high-bandwidth, low-latency transmissions, resulting in throughput degradation of up to 90% and delays of hundreds of milliseconds, severely impacting application performance. However, in virtualized multi-tenant data centers, host-based advancements in the TCP stack are hard to deploy from the operators’ perspective. Operators only provide infrastructure in the form of virtual machines, in which only tenants can directly modify the end-host TCP stack. In this paper, we present R-AQM, a switch-powered reverse ACK active queue management (R-AQM) mechanism for enhancing ACK-clocking effects through assisting legacy TCP. Specifically, R-AQM proactively intercepts ACKs and paces the ACK-clocked in-flight data packets, preventing TCP from suffering incast collapse. We implement and evaluate R-AQM in NS-3 simulation and NetFPGA-based hardware switch. Both simulation and testbed results show that R-AQM greatly improves TCP performance under heavy incast workloads by significantly lowering packet loss rate, reducing retransmission timeouts, and supporting 16 times (i.e., 60 to 1000) more senders. Meanwhile, the forward queuing delays are also reduced by 4.6 times.