Subscribe to the gold package and get unlimited access to Shamra Academy
Register a new userNetReduce: RDMA-Compatible In-Network Reduction for Distributed DNN Training Acceleration
153
0
0.0
(
0
)
Added by
Shuo Liu
Publication date
2020
fields
Informatics Engineering
and research's language is
English
Authors
Shuo Liu
Ask ChatGPT about the research
No Arabic abstract
We present NetReduce, a novel RDMA-compatible in-network reduction architecture to accelerate distributed DNN training. Compared to existing designs, NetReduce maintains a reliable connection between end-hosts in the Ethernet and does not terminate the connection in the network. The advantage of doing so is that we can fully reuse the designs of congestion control and reliability in RoCE. In the meanwhile, we do not need to implement a high-cost network protocol processing stack in the switch, as IB does. The prototype implemented by using FPGA is an out-of-box solution without modifying commodity devices such as NICs or switches. For the coordination between the end-host and the switch, NetReduce customizes the transport protocol only on the first packet in a data message to comply with RoCE v2. The special status monitoring module is designed to reuse the reliability mechanism of RoCE v2 for dealing with packet loss. A message-level credit-based flow control algorithm is also proposed to fully utilize bandwidth and avoid buffer overflow. We study the effects of intra bandwidth on the training performance in multi-machines multi-GPUs scenario and give sufficient conditions for hierarchical NetReduce to outperform other algorithms. We also extend the design from rack-level aggregation to more general spine-leaf topology in the data center. NetReduce accelerates the training up to 1.7x and 1.5x for CNN-based CV and transformer-based NLP tasks, respectively. Simulations on large-scale systems indicate the superior scalability of NetReduce to the state-of-the-art ring all-reduce.
rate research
Read More
The advent of RoCE (RDMA over Converged Ethernet) has led to a significant increase in the use of RDMA in datacenter networks. To achieve good performance, RoCE requires a lossless network which is in turn achieved by enabling Priority Flow Control (PFC) within the network. However, PFC brings with it a host of problems such as head-of-the-line blocking, congestion spreading, and occasional deadlocks. Rather than seek to fix these issues, we instead ask: is PFC fundamentally required to support RDMA over Ethernet? We show that the need for PFC is an artifact of current RoCE NIC designs rather than a fundamental requirement. We propose an improved RoCE NIC (IRN) design that makes a few simple changes to the RoCE NIC for better handling of packet losses. We show that IRN (without PFC) outperforms RoCE (with PFC) by 6-83% for typical network scenarios. Thus not only does IRN eliminate the need for PFC, it improves performance in the process! We further show that the changes that IRN introduces can be implemented with modest overheads of about 3-10% to NIC resources. Based on our results, we argue that research and industry should rethink the current trajectory of network support for RDMA.
Communication overhead poses an important obstacle to distributed DNN training and draws increasing attention in recent years. Despite continuous efforts, prior solutions such as gradient compression/reduction, compute/communication overlapping and layer-wise flow scheduling, etc., are still coarse-grained and insufficient for an efficient distributed training especially when the network is under pressure. We present DLCP, a novel solution exploiting the domain-specific properties of deep learning to optimize communication overhead of DNN training in a fine-grained manner. At its heart, DLCP comprises of several key innovations beyond prior work: e.g., it exploits {em bounded loss tolerance} of SGD-based training to improve tail communication latency which cannot be avoided purely through gradient compression. It then performs fine-grained packet-level prioritization and dropping, as opposed to flow-level scheduling, based on layers and magnitudes of gradients to further speedup model convergence without affecting accuracy. In addition, it leverages inter-packet order-independency to perform per-packet load balancing without causing classical re-ordering issues. DLCP works with both Parameter Server and collective communication routines. We have implemented DLCP with commodity switches, integrated it with various training frameworks including TensorFlow, MXNet and PyTorch, and deployed it in our small-scale testbed with 10 Nvidia V100 GPUs. Our testbed experiments and large-scale simulations show that DLCP delivers up to $84.3%$ additional training acceleration over the best existing solutions.
Distributed storage systems provide reliable access to data through redundancy spread over individually unreliable nodes. Application scenarios include data centers, peer-to-peer storage systems, and storage in wireless networks. Storing data using an erasure code, in fragments spread across nodes, requires less redundancy than simple replication for the same level of reliability. However, since fragments must be periodically replaced as nodes fail, a key question is how to generate encoded fragments in a distributed way while transferring as little data as possible across the network. For an erasure coded system, a common practice to repair from a node failure is for a new node to download subsets of data stored at a number of surviving nodes, reconstruct a lost coded block using the downloaded data, and store it at the new node. We show that this procedure is sub-optimal. We introduce the notion of regenerating codes, which allow a new node to download emph{functions} of the stored data from the surviving nodes. We show that regenerating codes can significantly reduce the repair bandwidth. Further, we show that there is a fundamental tradeoff between storage and repair bandwidth which we theoretically characterize using flow arguments on an appropriately constructed graph. By invoking constructive results in network coding, we introduce regenerating codes that can achieve any point in this optimal tradeoff.
A good parallelization strategy can significantly improve the efficiency or reduce the cost for the distributed training of deep neural networks (DNNs). Recently, several methods have been proposed to find efficient parallelization strategies but they all optimize a single objective (e.g., execution time, memory consumption) and produce only one strategy. We propose FT, an efficient algorithm that searches for an optimal set of parallelization strategies to allow the trade-off among different objectives. FT can adapt to different scenarios by minimizing the memory consumption when the number of devices is limited and fully utilize additional resources to reduce the execution time. For popular DNN models (e.g., vision, language), an in-depth analysis is conducted to understand the trade-offs among different objectives and their influence on the parallelization strategies. We also develop a user-friendly system, called TensorOpt, which allows users to run their distributed DNN training jobs without caring the details of parallelization strategies. Experimental results show that FT runs efficiently and provides accurate estimation of runtime costs, and TensorOpt is more flexible in adapting to resource availability compared with existing frameworks.
Deep Neural Nets have hit quite a crest, But physical networks are where they must rest, And here we put them all to the test, To see which network optimization is best.
Log in to be able to interact and post comments
comments
Fetching comments
Sign in to be able to follow your search criteria
