1. Data Processing Architecture It is mainly divided into four layers: Processing capability layer: Spark on Kubernetes provides streaming data processing capability Data management layer: Iceberg provides dataset access operations such as ACID and tables Storage layer: Hive MetaStore manages Iceberg table metadata, PostgreSQL serves as the storage backend for

1. What GDS (GPUDirectStorage) Is GDS lets an RDMA NIC access GPU memory directly, which helps increase the IO bandwidth for GPU applications reading and writing files, reduces IO latency, and lowers their CPU load. Once the client has GDS enabled, files are opened in O_DIRECT mode and the client no longer caches file data.

The DLRover version used in this article is 0.3.7 1. DLRover Operator 1.1 Starting the ElasticJob and ScalePlan Controllers Implementation code: 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 // 创建 ElasticJob 的控制器 if err = controllers.NewElasticJobReconciler(mgr, masterImage).SetupWithManager(mgr); err

1. Parameter Server Architecture In the Parameter Server architecture, the nodes in the cluster are divided into two categories: parameter server nodes (Parameter Server) and worker server nodes (Worker). 1.1 Parameter Server The Parameter Server is used to store the model’s parameters. Each parameter server node is responsible for managing and updating one part of the model’s parameters, while each worker node only processes the subset of data corresponding to it.

1. Problems Facing Distributed Training Estimating training resources is difficult and cannot be automated How much compute, how much time, how much bandwidth, how many CPUs, how much memory — without enough accumulated experience it is hard to estimate accurately. The result is over-requesting and over-allocation, causing enormous resource waste.