mir/criu
2
0
Fork 0
mirror of https://github.com/checkpoint-restore/criu synced 2025-08-21 17:37:39 +00:00
Code Issues Releases Wiki Activity

plugins/amdgpu: Update README.md and criu-amdgpu-plugin.txt

Browse Source 
Signed-off-by: Yanning Yang <yangyanning@sjtu.edu.cn>
This commit is contained in:
Yanning Yang 2024-12-10 12:44:35 +00:00 committed by Andrei Vagin
parent bfb4a3d842
commit 7c4bcdb2d4
2 changed files with 23 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
Documentation/criu-amdgpu-plugin.txt
View File

@ -15,6 +15,7 @@ Checkpoint / Restore inside a docker container
Pytorch
Tensorflow
Using CRIU Image Streamer
Parallel Restore
DESCRIPTION
-----------

23
plugins/amdgpu/README.md
View File

@ -3,7 +3,8 @@ Supporting ROCm with CRIU
_Felix Kuehling <Felix.Kuehling@amd.com>_<br>
_Rajneesh Bardwaj <Rajneesh.Bhardwaj@amd.com>_<br>
_David Yat Sin <David.YatSin@amd.com>_
_David Yat Sin <David.YatSin@amd.com>_<br>
_Yanning Yang <yangyanning@sjtu.edu.cn>_
# Introduction
@ -224,6 +225,26 @@ to resume execution on the GPUs.
*This new plugin is enabled by the new hook `__RESUME_DEVICES_LATE` in our RFC
patch series.*
## Restoring BO content in parallel
Restoring the BO content is an important part in the restore of GPU state and
usually takes a significant amount of time. A possible location for this
procedure is the `cr_plugin_restore_file` hook. However, restoring in this hook
blocks the target process from performing other restore operations, which
hinders further optimization of the restore process.
Therefore, a new plugin hook that runs in the master restore process is
introduced, and it interacts with the `cr_plugin_restore_file` hook to complete
the restore of BO content. Specifically, the target process only needs to send
the relevant BOs to the master restore process, while this new hook handles all
the restore of buffer objects. Through this method, during the restore of the BO
content, the target process can perform other restore operations, thus
accelerating the restore procedure. This is an implementation of the gCROP
method proposed in the ACM SoCC'24 paper: [On-demand and Parallel
Checkpoint/Restore for GPU Applications](https://dl.acm.org/doi/10.1145/3698038.3698510).
*This optimization technique is enabled by the `__POST_FORKING` hook.*
## Other CRIU changes
In addition to the new plugins, we need to make some changes to CRIU itself to