FAQs
- Q: What should I do if "libascendfaiss.so not found" is displayed during build?
A: Ensure that the path of libascendfaiss.so is stored in the lib package directory, and add it to the LD_LIBRARY_PATH environment variable.
- Q: Does the current FeatureRetrieval support multi-index retrieval in different threads?
A: Each processor can process only one request from the same process at a time and cannot process requests from two different processes on the host at the same time. However, for multiple threads of the same process, FeatureRetrieval allows different threads to share a device. But in the multi-thread scenario, if two indexes from two threads on the host are created on a processor, when the device processes the first index request, the request for processing the second index will be blocked.
- Q: Why is the training speed of large library algorithms such as IVFSQ on openEuler slow?
A: The CPU performance deterioration bug is introduced to openEuler 20.03 LTS. You can modify and adapt the system by visiting https://gitee.com/src-openeuler/openEuler-release/pulls/55 or use the latest openEuler system.
- Q: What should I do if the current retrieval performance does not meet the expectations?
A: Set the ratio of the Ctrl CPU to the AI CPU to 6:2 by referring to the DSMI API Reference. If an Atlas inference products-powered heterogeneous computing platform such as Atlas 300I Pro is used, set the ratio of the Ctrl CPU to the AI CPU to 7:1.