Lcore Guide

This guide is outdated and should only be referenced if the use of the Manager service is not possible.

1. Introduction

In DPDK (Data Plane Development Kit), an “lcore” stands for “logical core,” and it represents a logical CPU core on a multi-core processor. Binding a thread to a specific logical core (lcore) is a technique used to achieve better control over the execution of packet processing tasks and to optimize the performance of networking applications.

It minimizes the impact of OS scheduler decisions, reduces cache-related issues, and allows for fine-grained control over CPU resources, all of which are critical for meeting the stringent performance requirements of networking workloads. Media Transport Library scheduler used pinned lcore also as performance consideration.

2. Lcore manager for multi process

MTL supports multi-process deployment based on SR-IOV VF isolation. To manage the dispatching of lcores among multiple processes, it introduces shared memory mapping to maintain the status of lcore usage. Each process searches the mapping and allocates a new lcore slot from it, freeing the slot when it is no longer in use. The last user attached to the shared memory will reset the mapping to an initial status during the shmdt operation as part of the release routine.

MTL also provides a tool which can be find from ./build/app/LcoreMgr for manually cleaning up lcore entries. This tool is typically used when a process fails to release lcores due to panic issues or other failures that prevent the regular free routine from running. Below is the usage information:

  --info: Print lcore shared manager detail info
  --clean_pid_auto_check: Clean the dead entries if PID is not active
  --clean_lcore <lcore id>: Clean the entry by lcore ID

When you use the --clean_pid_auto_check option, the tool will perform a loop check for all the active entries in the map. It checks whether the hostname and user match the current login environment and then verifies if the PID is still running. If the tool detects that the PID is no longer active, it will proceed to remove the lcore from the mapping.

If you are deploying in a multi-container environment, the PID check becomes less useful as each container has its own process namespace. In such cases, you can use the --clean_lcore option to remove an entry based on the lcore ID. However, it’s important to confirm that the lcore is not active before using this option.

Before using the --clean_lcore option, please ensure that you double-check the current lcore status by running ./build/app/LcoreMgr --info. An example output list is provided below. Pay close attention to the lcore usage details, including user, host, PID, process name, allocation information, and, most importantly, CPU usage. Typically, an active lcore allocated by the MTL library should utilize 100% of the CPU resources.

MT: 2023-11-14 15:00:03, mtl_lcore_shm_print, MTL used lcores 3
MT: 2023-11-14 15:00:03, mtl_lcore_shm_print, lcore 28 active by xxx@xxx, pid: 236759(comm: RxTxApp) type: app_allocated
MT: 2023-11-14 15:00:03, mtl_lcore_shm_print, lcore 29 active by xxx@xxx, pid: 236759(comm: RxTxApp) type: lib_sch
MT: 2023-11-14 15:00:03, mtl_lcore_shm_print, lcore 30 active by xxx@xxx, pid: 236759(comm: RxTxApp) type: lib_sch
MT: 2023-11-14 15:00:03, mtl_lcore_shm_print, collecting cpu usage
MT: 2023-11-14 15:00:04, mtl_lcore_shm_print, lcore 28 cpu usage 4.08%
MT: 2023-11-14 15:00:04, mtl_lcore_shm_print, lcore 29 cpu usage 100.00%
MT: 2023-11-14 15:00:04, mtl_lcore_shm_print, lcore 30 cpu usage 100.00%