Hey there π, Itβs Yangyang from Microsoft working on Gluten Project. The purpose of this email is to collect ideas on improving the gluten native diagnose experience.
Gluten is indeed impressive, particularly in its significant contribution to boosting the performance of Spark ππ . but unfortunately, the current online diagnose experience doesn't quite match up. Right now, we're teaming up with the Intel Gluten team to plan some enhancements for the gluten native diagnose experience, the main goal is to make it easier for users to quickly locate issues when errors or any performance regression occur online. Our primary idea now is to utilize the gluten UI page as the main entry point and incorporate the necessary enhanced features directly onto this page, referred to as gluten UI enhancement, so that much enhancement work can directly go into OSS. Some ideas currently under consideration include: 1) Onboarding colorized DGA graph in Gluten UI like how Photon does to help users quickly identify which nodes are fallback or not and the context detail. 2) Onboarding execution time summary table in Gluten UI like How Photon does to help users quickly identify which stage/op is time-consuming. [A white paper with black text Description automatically generated] 3) Onboarding an async profile Flame graph in Gluten UI including CPU_FLAME_GRAPHS, CPU_STACK_TRACES, and WALL_CLOCK_STACK_TRACE to assist users pinpointing which native function calls are time-consuming. [cid:[email protected]] Our main goal with this email is to gather feedback and requirements from our community members. We value your input! Among the options mentioned, which one do you think is most essential right now? Additionally, we're open to hearing your fresh ideas and any new requirements you might have for improving the gluten native diagnose experience. Your contributions to this discussion are highly valued and appreciated! We look forward to your feedback. Thank you. [email protected]<mailto:[email protected]>
