alloc job resource

alloc job resource

select placement scheme

select placement scheme

['even_time', 'job_id', 'num_gpu', 'submit_time', 'start_time', 'end_time', 'executed time', 'real executed time', 'jct', 'duration', 'pendin

Record cluster, and job information, including: time idle_node busy_node: gpu running full_node: all gpus are

sync for communication

forward, calculate loss and backward

get data

['even_time', 'job_id', 'num_gpu', 'submit_time', 'start_time', 'end_time', 'executed time', 'real executed time', 'jct', 'duration', 'pendin

prepare dataloader, model, optimizer for training

input is node_dict from placement {'id':xx, 'num_gpu':xxx, 'num_cpu': xxx, 'network': xxxx, 'tasks': [w2, ps2]}

job gets into the system: pending or running, and finally END

input is node_dict from placement {'id':xx, 'num_gpu':xxx, 'num_cpu': xxx, 'network': xxxx, 'tasks': [w2, ps2]}

under this switch, there is only one need used {'switch': xx, 'nodes': [{'id':xx, 'num_gpu':xxx, 'num_cpu': xxx, 'network': xxxx, 'ta

under this switch, there is only one need used {'switch': xx, 'nodes': [{'id':xx, 'num_gpu':xxx, 'num_cpu': xxx, 'network': xxxx, 'ta

job gets into the system: pending or running, and finally END

Record cluster, and job information, including: time queue length blocking index gpu util cpu util

new jobs are added to the end of the ending queue but in the queue, shortest (gpu) job first be served and pack other jobs with the same

new jobs are added to the end of the ending queue but in the queue, shortest (gpu) job first be served and pack other jobs with the same

release using cpus back to free list

release using cpus back to free list

input is gpu and cpu

input is gpu and cpu