1. Mount a Memory Storage Directory on the Host Create a directory to mount 1 mkdir /mnt/memory_storage Mount the tmpfs filesystem 1 mount -t tmpfs -o size=800G tmpfs /mnt/memory_storage Storage is allocated on demand, so using 100G of storage only consumes 100G of memory. The host node has 2T of memory, and here 800G is allocated to store Elasticsearch data.

1. Install the Driver Create the HwHiAiUser user 1 2 groupadd -g 1000 HwHiAiUser useradd -g HwHiAiUser -u 1000 -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash Add directory permissions 1 2 chown -R HwHiAiUser /usr/local/Ascend chmod -R 755 /usr/local/Ascend Download the driver and firmware Go to https://www.hiascend.ru/hardware/firmware-drivers/community?product=1&model=30&cann=All&driver=1.0.26.alpha to find the matching driver and firmware.

1. Data Delivery Based on Object Storage As shown above, model R&D mainly involves three sub-platforms: Data platform The data platform is mainly responsible for data-related management, such as data ingestion and data processing, ultimately producing the data needed for training. The data platform stores raw data in object storage. During processing, it fetches data from object storage, processes it, and stores the result back in object storage.

1. Storage Tiering and Interconnect 2. Various Storage Types Storage Type Bandwidth Capacity Response Latency Cost per Unit of Storage Storage Principle Use Case CPU L1 Cache 256-512GB/s 32-64KB/core <1ns ~$2000/MB SRAM Closest cache layer to CPU core CPU L2 Cache 64-256GB/s 256KB-2MB/core group 2-5ns ~$1000/MB Partly SRAM, partly embedded DRAM Next-level cache for the CPU CPU L3 Cache 32-128GB/s 4-16MB/CPU 5-15ns ~$100-200/MB Embedded DRAM Last-level cache for the CPU GPU L1 Cache 8-32TB/s 16-96KB/SM <1ns ~$2000/MB SRAM Closest cache layer to GPU core GPU L2 Cache 1-4TB/s 512KB-4MB/SMC 5-15ns ~$500-1000/MB Partly SRAM, partly embedded DRAM Next-level cache for the GPU Main memory (RAM) 10-80GB/s 4-128GB 30-60ns ~$2-7/GB DDR4/5 DRAM System main storage space GPU VRAM 768GB/s-4TB/s 4-24GB 25-75ns ~$7-20/GB GDDR6/6X SDRAM Dedicated high-speed buffer storage for the GPU SSD 500MB/s-7000MB/s 128GB-8TB 0.

1. Storage Options There are three storage approaches: Isolate by directory and share a single JuiceFS Elasticsearch nodes share one JuiceFS instance, mounting different Elasticsearch nodes through subdirectories. /0/ corresponds to node Node-0 /1/ corresponds to node Node-1 /2/ corresponds to node Node-2 The main advantage of this approach is that it is easy to scale and simple to configure.