Search

The emergence of Large Language Models (LLMs) has necessitated the adoption of parallel training techniques, involving the deployment of thousands of GPUs to train a single model. Unfortunately, we have found that the efficiency of current parallel training is often suboptimal, largely due to the following two main issues. Firstly, hardware failures are inevitable, leading to interruptions in the training tasks. The inability to quickly identify the faulty components results in a substantial waste of GPU resources. Secondly, since GPUs must wait for parameter synchronization to complete before proceeding to the next round of computation, network congestions can greatly increase the waiting time for GPUs. To address these challenges, this paper introduces a communication-driven solution, namely the C4. The key insights of C4 are two folds. First, in parallel training, collective communication exhibits periodic and homogeneous characteristics, so any anomalies are certainly due to some form of hardware malfunction. By leveraging this feature, C4 can rapidly identify the faulty components, swiftly isolate the anomaly, and restart the task, thereby avoiding resource wastage caused by delays in anomaly detection. Second, the predictable communication model of collective communication, involving few large flows, allows C4 to efficiently execute traffic planning, substantially reducing network congestion. C4 has been extensively implemented across our production systems, cutting error-induced overhead by roughly 30% and enhancing runtime performance by about 15% for certain applications with moderate communication costs.

In this paper, we conduct systematic measurement studies to show that the high memory bandwidth consumption of modern distributed applications can lead to a significant drop of network throughput and a large increase of tail latency in high-speed RDMA networks.We identify its root cause as the high contention of memory bandwidth between application processes and network processes. This contention leads to frequent packet drops at the NIC of receiving hosts, which triggers the congestion control mechanism of the network and eventually results in network performance degradation. To tackle this problem, we make a key observation that given the distributed storage service, the vast majority of data it receives from the network will be eventually written to high-speed storage media (e.g., SSD) by CPU. As such, we propose to bypass host memory when processing received data to completely circumvent this performance bottleneck. In particular, we design Lamda, a novel receiver cache processing system that consumes a small amount of CPU cache to process received data from the network at line rate. We implement a prototype of Lamda and evaluate its performance extensively in a Clos-based testbed. Results show that for distributed storage applications, Lamda improves network throughput by 4.7% with zero memory bandwidth consumption on storage nodes, and improves network throughput by up 17% and 45% for large block size and small size under the memory bandwidth pressure, respectively. Lamda can also be applied to latency-sensitive HPC applications, which reduces their communication latency by 35.1%.

Pre-training methods with contrastive learning objectives have shown remarkable success in dialog understanding tasks. However, current contrastive learning solely considers the self-augmented dialog samples as positive samples and treats all other dialog samples as negative ones, which enforces dissimilar representations even for dialogs that are semantically related. In this paper, we propose SPACE-2, a tree-structured pre-trained conversation model, which learns dialog representations from limited labeled dialogs and large-scale unlabeled dialog corpora via semi-supervised contrastive pre-training. Concretely, we first define a general semantic tree structure (STS) to unify the inconsistent annotation schema across different dialog datasets, so that the rich structural information stored in all labeled data can be exploited. Then we propose a novel multi-view score function to increase the relevance of all possible dialogs that share similar STSs and only push away other completely different dialogs during supervised contrastive pre-training. To fully exploit unlabeled dialogs, a basic self-supervised contrastive loss is also added to refine the learned representations. Experiments show that our method can achieve new state-of-the-art results on the DialoGLUE benchmark consisting of seven datasets and four popular dialog understanding tasks. For reproducibility, we release the code and data at https://github.com/AlibabaResearch/DAMO-ConvAI/tree/main/space-2., Comment: 17 pages, 6 figures. Accepted by COLING 2022

Deep learning has been used in a wide range of areas and made a huge breakthrough. With the ever-increasing model size and train-ing data volume, distributed deep learning emerges which utilizes a cluster to train a model in parallel. Unfortunately, the performance is often far from linear speedup due to the communication overhead between cluster nodes. To address this challenge, this paper designs and implements Libra, a network aggregator, that utilizes in-network computation to optimize the communication for distributed DL training in two aspects: 1) reduce active connections and 2) aggregate exchanged network packets. We implemented our Libra on Intel Tofino switches, customized a lightweight host stack and integrated it into an open-source training framework PS-lite. The experimental result shows that our Libra can achieve 1.5~4 times speedup., Comment: 13 pages, 18 figures

The development of personalized recommendation has significantly improved the accuracy of information matching and the revenue of e-commerce platforms. Recently, it has 2 trends: 1) recommender systems must be trained timely to cope with ever-growing new products and ever-changing user interests from online marketing and social network; 2) SOTA recommendation models introduce DNN modules to improve prediction accuracy. Traditional CPU-based recommender systems cannot meet these two trends, and GPU- centric training has become a trending approach. However, we observe that GPU devices in training recommender systems are underutilized, and they cannot attain an expected throughput improvement as what it has achieved in CV and NLP areas. This issue can be explained by two characteristics of these recommendation models: First, they contain up to a thousand input feature fields, introducing fragmentary and memory-intensive operations; Second, the multiple constituent feature interaction submodules introduce substantial small-sized compute kernels. To remove this roadblock to the development of recommender systems, we propose a novel framework named PICASSO to accelerate the training of recommendation models on commodity hardware. Specifically, we conduct a systematic analysis to reveal the bottlenecks encountered in training recommendation models. We leverage the model structure and data distribution to unleash the potential of hardware through our packing, interleaving, and caching optimization. Experiments show that PICASSO increases the hardware utilization by an order of magnitude on the basis of SOTA baselines and brings up to 6x throughput improvement for a variety of industrial recommendation models. Using the same hardware budget in production, PICASSO on average shortens the walltime of daily training tasks by 7 hours, significantly reducing the delay of continuous delivery.