To meet the massive compute and high-speed communication demands of Large Language Model (LLM) training, modern datacenters typically adopt multipath topologies such as Fat-Tree and Clos to host parallel jobs across hundreds to thousands of GPUs. However, LLM training exhibits periodic, high-bandwidth communication patterns. Existing load-balancing schemes become misaligned under dynamic congestion and anomalous surges: they struggle to promptly mitigate iteration-peak congestion and lack effective isolation of anomalous traffic. To address this, we propose Security-Aware Load Balancing (SALB) for LLM training. SALB leverages a Deep Reinforcement Learning (DRL) controller with queue and delay signals for packet-level multipath load balancing and employs path binding to confine suspicious flows. By integrating data security into load balancing, SALB simultaneously achieves high throughput and robust traffic isolation. NS-3 simulation results show that, compared with CONGA, Hermes, and ConWeave, SALB reduces the 99th-percentile flow completion time (FCT) of short flows by an average of 65% and increases the throughput of long flows by an average of 54%. It further outperforms the baselines in aggregate throughput, path utilization, and packet loss rate, thereby significantly enhancing system stability, robustness, and data security.

Modern industrial datacenter networks employ multirooted tree topologies to accommodate a diverse range of cloud applications, which generate heterogeneous traffic with low-latency short flows and high-throughput long flows. Recently, the proposed learning-based load balancing mechanisms are resilient to dynamic network, but they are agnostic to heterogeneous traffic, resulting in large tail delay. In this article, we propose a new deep reinforcement learning (DRL) based load balancing called DRLB, which uses DRL with the distributed distributional deterministic policy gradients algorithm to make (re)routing for long flows, and adopts the weighted cost multipathing mechanism for short flows. Furthermore, this article introduces a traffic feature-based dynamic training cycle mechanism to adaptively adjust the training cycles. The experimental results show DRLB reduces the flow completion time of short flows by up to 58% and improves the throughput of long flows by 38% compared to the state-of-the-art load balancing mechanisms.

Increasingly large-scale models and rich data sets make communication overhead a key bottleneck for distributed Deep Neural Network (DNN) training, constantly attracting the attention of academia and industry. Despite continuous efforts, prior solutions such as pipelining computation/communication and in-network gradient compression/scheduling do not focus on how to accelerate DNN training through load balancing in datacenter networks (DCNs). However, the existing load balancing mechanisms are unaware of tensor integrity and priority for gradient parameter synchronization during the DNN training iterations, resulting in severe tensor tail latency and slow model convergence speed. In this paper, we present a Tensor-aware Load Balancing (TaLB) scheme to accelerate DNN training. Specifically, TaLB identifies the different priority tensors and makes (re)routing decisions based on the tensor-level granularity to cut the high-priority tensors tail delay. The testbed implementation and large-scale NS-3 simulation results show that TaLB effectively accelerates DNN training speed. For example, TaLB significantly reduces the average flow completion time (FCT) by up to 55%, and accelerates the model training speed up to 2.37× on VGG19, ResNet50 and AlexNet models.

Many existing load balancing mechanisms work effectively in lossy datacenter networks (DCNs), but they suffer from serious packet reordering in lossless Ethernet DCNs deployed with the hop-by-hop Priority-based Flow Control (PFC). The key reason is that the prior solutions are not able to perceive PFC triggering correctly and in a timely manner when making load balancing decisions. Once the forwarding path pauses transmission due to PFC triggering, the packets allocated on it are blocked, inevitably leading to out-of-order packets and retransmission. In this paper, we present an Reordering-robust Load Balancing (RLB) scheme with PFC prediction in lossless DCNs. At its heart, RLB leverages the derivative of ingress queue length to predict PFC triggering and proactively notifies the upstream switches to choose an appropriate rerouting path or perform packet recirculation to avoid reordering. Furthermore, under switch failure scenarios, RLB adjusts the recirculation threshold adaptively to mitigate the risk of packets over-recirculation. We have implemented RLB in the hardware programmable switch. As a building block for existing load balancing mechanisms, we have integrated RLB into Presto, LetFlow, Hermes and DRILL. The evaluation results show that the RLB-enhanced solutions deliver significant performance by avoiding packet reordering. For example, it reduces the$99^{th}$percentile flow completion time (FCT) by up to 72%, 67%, 58% and 54% over DRILL, Presto, LetFlow and Hermes, respectively.

Many existing load balancing mechanisms work effectively in lossy datacenter networks (DCNs), but they suffer from serious packet reordering in lossless Ethernet DCNs deployed with the hop-by-hop Priority-based Flow Control (PFC). The key reason is that the prior solutions are not able to correctly and timely perceive PFC triggering when making load balancing decisions. Once the forwarding path pauses transmission due to PFC triggering, the packets allocated on it are blocked, inevitably leading to out-of-order packets and retransmission. In this paper, we present a Reordering-robust Load Balancing (RLB) scheme with PFC prediction in lossless DCNs. At its heart, RLB leverages the derivative of ingress queue length to predict PFC triggering and proactively notifies the upstream switches to choose an appropriate rerouting path or perform packet recirculation to avoid reordering. As a building block for existing load balancing mechanisms, we have integrated RLB into Presto, LetFlow, Hermes and DRILL. The test results show that the RLB-enhanced solutions deliver significant performance by avoiding packet reordering. For example, it reduces the 99th percentile flow completion time (FCT) by up to 58%, 67%, 72% and 54% over Presto, LetFlow, Hermes and DRILL, respectively.