Hello! I see the source code of NPKit for GPU events tracing and found that how you discern one device 's events from another is through rank index. However, the same GPU could be mapped to as different ranks in different communicators. So if used in multiple communicators, would it be the case that GPU event buffer with rank index r documents events of GPU 1 (rank r in comm 1) and events of GPU 2 (rank r in comm 2). So we cannot distinguish which events are from GPU 1 and which are from GPU 2. Thanks for any help!
Hello! I see the source code of NPKit for GPU events tracing and found that how you discern one device 's events from another is through rank index. However, the same GPU could be mapped to as different ranks in different communicators. So if used in multiple communicators, would it be the case that GPU event buffer with rank index
rdocuments events of GPU 1 (rankrin comm 1) and events of GPU 2 (rankrin comm 2). So we cannot distinguish which events are from GPU 1 and which are from GPU 2. Thanks for any help!