Abstract

1. Hardware and Software Environment

2. Swap Location Strategy

3. Step-by-Step Swap File Creation

3.1 Check Devices and Free Space

3.2 Create the Swap File

3.3 Secure and Format the Swap File

3.4 Enable the Swap File

4. Making Swap Persistent Across Reboots

5. Tuning Swappiness and zram for LLM Workloads

5.1 Adjust Swappiness and Cache Pressure

6. Relationship Between zram and /swapfile

7. Removing or Reconfiguring the Swap File

8. Practical Outcomes

9. Conclusions This document describes the process of creating, tuning, and managing a large swap file on an NVIDIA Jetson AGX Orin 64 GB running Ubuntu 22.04.5 LTS aarch64. The configuration is specifically optimized for running large language models (LLMs) alongside CUDA, cuMB, and TensorRT by leveraging a fast NVMe SSD as the primary swap backing store. The implementation was validated using a 50 GB swap file configuration alongside existing zram layers. The procedure successfully extended the usable memory capacity, allowing for the deployment of larger models without triggering immediate Out-Of-Memory (OOM) errors, provided the storage-to-RAM paging latency is acceptable. This tutorial serves as a technical reference for advanced Jetson and Linux users. It provides a reproducible method for extending virtual memory on edge AI hardware to support demanding 34B–70B parameter models. The target environment is an NVIDIA Jetson AGX Orin Developer Kit equipped with 64 GB of unified memory. The system runs Ubuntu 22.04.5 LTS on an aarch64 kernel (5.15.185-tegra). The installation includes JetPack 6.2.2, providing the necessary software stack for AI inference, including CUDA 12.6, cuDNN 9.3.0, and TensorRT 10.3.0. The primary storage for the swap file is the NVMe SSD, which serves as the root filesystem. This choice is critical for minimizing the performance penalty during memory paging operations. Table 1 — Jetson AGX Orin environment for swap configuration Effective swap placement is determined by the throughput and endurance of the underlying storage media. On the Jetson AGX Orin, the system utilizes eMMC for the boot partition and an NVMe SSD for the primary root filesystem. Table 2 — Recommended swap backing storage on Jetson AGX Orin For this configuration, the swap file is placed directly on the NVMe-backed root filesystem (/) at /swapfile. This ensures the highest possible I/O performance for paging operations. The following steps outline the allocation and initialization of a 50 GB swap file. Before allocation, verify the available space on the target partition. The lsblk command confirms the mount points, while df -h verifies the capacity. The current configuration shows approximately 636 GB of available space on /dev/nvme0n1p1, which is more than sufficient for a 50 GB allocation. The fallocate utility is used to pre-allocate the file space efficiently. Security is paramount; the swap file must be restricted to root-only access to prevent sensitive data leakage from memory to disk. Once formatted, the swap file must be activated in the running kernel. To ensure the swap file is automatically re-enabled upon system restart, an entry must be added to the /etc/fstab configuration file. Optimal performance for LLM inference requires tuning the kernel to prioritize physical RAM and the compressed zram layer over the disk-backed swap file. Lowering the swappiness value instructs the kernel to avoid swapping pages to the NVMe SSD unless absolutely necessary. Table 3 — Swappiness values and behavior for Jetson LLM use The Jetson system utilizes a tiered memory architecture. The zram-config service provides several compressed RAM-based swap devices (zram0 through zram11). The hierarchy of memory allocation is as follows: This tiered approach allows the kernel to handle small, compressible allocations within the highly efficient zram layer before resorting to the higher-latency NVMe disk-backed swap. If disk space needs to be reclaimed, the swap file can be decommissioned following these steps: Configuring a large, NVMe-backed swap file is a highly effective strategy for maximizing the utility of the NVIDIA Jetson AGX Orin 64 GB for large-scale AI workloads. By following the documented procedure of using fallocate, setting strict chmod 600 permissions, and tuning swappiness to 10, users can achieve a stable environment capable of handling models that exceed physical memory boundaries. While the performance penalty of disk-based swapping is unavoidable, the use of high-speed NVMe storage and a tiered zram approach minimizes the impact on inference latency, making it a viable solution for non-interactive or batch processing of 34B–70B parameter models. Templates let you quickly answer FAQs or store snippets for re-use. Hide child comments as well For further actions, you may consider blocking this person and/or reporting abuse