分布式训练ps负载不均衡
Open
- Dominant language
- C++
- Stars
- 2.9k
- Forks
- 553
- PR merge metrics
- No merged PRs in 30d
Description
我用自己写的模型,在分布式下跑,10个分片,9千万结点,启用2个ps会oom,但是启用3个ps时第三个启用了但是没有数据进入。
测试了用line去跑一份简单的数据,3个ps1个worker,只有第一个ps工作,第23个ps没有参与计算,该如何解决?
Contributor guide
No contributing guide indexed for this repository
Research direction
Start by reproducing the reported setup with 10 shards, 90 million nodes, and two versus three parameter servers, then compare it with the simpler Line dataset using three parameter servers and one worker. Done means identifying why only the first parameter server receives work and documenting a validated fix or clear reproduction findings.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- cpp
- Domain
- distributed-systems, machine-learning
- Issue type
- Bug
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 25/100