[release-1.13] Fix gpu resource error - #4916
Conversation
…aringPredicateAndScore will add the used mem/core to the structure. For avoiding the pod allocated info updating latency, we use addResource it to add the pod the pod map. But it also add the used mem/core. To address this, we new a function that only add pod the map. Signed-off-by: chenw66 <chenw66@chinaunicom.cn>
Summary of ChangesHello @volcano-sh-bot, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request, which is an automated cherry-pick, addresses a GPU resource error by refining the mechanism for tracking GPU usage per pod. It introduces a dedicated function, Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
volcano-sh-bot
added
the
size/M
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request aims to fix a GPU resource error by introducing a new function, addToPodMap, to handle optimistic resource updates during allocation. While the intent is good, the new function has several issues. I've identified a critical flaw that could lead to GPU overcommitment, a high-severity risk of resource miscounting, and a minor logging issue. My review provides a single, comprehensive comment with a suggested rewrite of the function to address all these points, ensuring resource accounting is correct and robust.
| func (gs *GPUDevices) addToPodMap(annotations map[string]string, pod *v1.Pod) { | ||
| ids, ok := annotations[AssignedIDsAnnotations] | ||
| if !ok { | ||
| klog.Errorf("pod %s has no annotation volcano.sh/devices-to-allocate", pod.Name) | ||
| return | ||
| } | ||
| podDev := decodePodDevices(ids) | ||
| for _, val := range podDev { | ||
| for _, deviceused := range val { | ||
| for _, gsdevice := range gs.Device { | ||
| if strings.Contains(deviceused.UUID, gsdevice.UUID) { | ||
| podUID := string(pod.UID) | ||
| _, ok := gsdevice.PodMap[podUID] | ||
| if !ok { | ||
| gsdevice.PodMap[podUID] = &GPUUsage{ | ||
| UsedMem: 0, | ||
| UsedCore: 0, | ||
| } | ||
| } | ||
|
|
||
| gsdevice.PodMap[podUID].UsedMem += deviceused.Usedmem | ||
| gsdevice.PodMap[podUID].UsedCore += deviceused.Usedcores | ||
| } | ||
| } | ||
| } | ||
| } | ||
| } |
There was a problem hiding this comment.
This new function has a few issues that should be addressed:
- Critical: It only updates the
PodMapbut not the summary fields likeUsedMemandUsedCoreon thegsdevice. This can lead to stale resource usage data within a scheduling session, potentially causing overcommitment of GPU resources as subsequent scheduling decisions will not see the updated usage. - High: The inner loop over
gs.Devicedoes notbreakafter finding a match. A single allocated device should belong to exactly one physical GPU. Without abreak, its usage might be counted on multiple physical devices if UUIDs are not perfectly distinct, leading to incorrect resource accounting. - Medium: The error message on line 212 is misleading. It refers to
volcano.sh/devices-to-allocate, but the code checks forvolcano.sh/vgpu-ids-new(AssignedIDsAnnotations).
I'm providing a single suggestion that refactors the function to fix all three issues.
func (gs *GPUDevices) addToPodMap(annotations map[string]string, pod *v1.Pod) {
ids, ok := annotations[AssignedIDsAnnotations]
if !ok {
klog.Errorf("pod %s has no annotation %s", pod.Name, AssignedIDsAnnotations)
return
}
podDev := decodePodDevices(ids)
for _, val := range podDev {
for _, deviceused := range val {
for _, gsdevice := range gs.Device {
if strings.Contains(deviceused.UUID, gsdevice.UUID) {
podUID := string(pod.UID)
if _, ok := gsdevice.PodMap[podUID]; !ok {
gsdevice.PodMap[podUID] = &GPUUsage{
UsedMem: 0,
UsedCore: 0,
}
}
gsdevice.PodMap[podUID].UsedMem += deviceused.Usedmem
gsdevice.PodMap[podUID].UsedCore += deviceused.Usedcores
gsdevice.UsedMem += deviceused.Usedmem
gsdevice.UsedCore += deviceused.Usedcores
break
}
}
}
}
}
|
/approve |
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: JesseStutler The full list of commands accepted by this bot can be found here. DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
2 participants
This is an automated cherry-pick of #4873