master

Mmszjaasfeat:Support three steps pipeline rh2d

ca9b6e37创建于 7 天前历史提交

文件	最后提交记录	最后更新时间
BUILD.bazel	add bazel build	2 个月前
CMakeLists.txt	embeddedclient mmap	3 个月前
blocking_queue.h	Hi all, this is yuanrong datasystem	7 个月前
circular_queue.h	Hi all, this is yuanrong datasystem	7 个月前
priority_queue.h	Hi all, this is yuanrong datasystem	7 个月前
queue.h	Hi all, this is yuanrong datasystem	7 个月前
shm_circular_queue.cpp	feat:Support three steps pipeline rh2d Three steps pipeline: 1. worker2 -> worker1: urma write 2. worker1 -> kvclient: share memory queue 3. kvclient -> gpu: cuda The main idea diff from two step pipeline lays sending chunk info from worker1 to kvclient by share memory queue. Changes: 1. enlarge chunk maxsize from 16m to 4G 2. pin cuda memory in connecting stage 3. support timeout cancelling pipeline in kvclient and worker 4. make client request id atomic. 5. add vlog for pipeline debuging 6. auto find os-transport lib without installing rpm 7. support gpu id and gpu uuid as connection args 8. rename H2DParam to PiplnRh2dParam 9. worker don't need cuda lib 10. clean up worker rpc channel when worker lost Signed-off-by: mszjaas <mszjaas@qq.com>	7 天前
shm_circular_queue.h	feat:Support three steps pipeline rh2d Three steps pipeline: 1. worker2 -> worker1: urma write 2. worker1 -> kvclient: share memory queue 3. kvclient -> gpu: cuda The main idea diff from two step pipeline lays sending chunk info from worker1 to kvclient by share memory queue. Changes: 1. enlarge chunk maxsize from 16m to 4G 2. pin cuda memory in connecting stage 3. support timeout cancelling pipeline in kvclient and worker 4. make client request id atomic. 5. add vlog for pipeline debuging 6. auto find os-transport lib without installing rpm 7. support gpu id and gpu uuid as connection args 8. rename H2DParam to PiplnRh2dParam 9. worker don't need cuda lib 10. clean up worker rpc channel when worker lost Signed-off-by: mszjaas <mszjaas@qq.com>	7 天前