only used in sim-gpu-demands

only used in sim-gpu-demands

If enough free gpus, allocate gpus Return: True, for success; False, for failure

If enough free gpus, allocate gpus Return: True, for success; False, for failure

antman allocates res from a single switch for resource-guarantee job: for opportunistic job: if no enough gpus, give

antman allocates res from a single switch for resource-guarantee job: for opportunistic job: if no enough gpus, give

calculate 1 iter time of job, whose priority is 1 ignore synchronization among different gpu and return the maximum time of each gpu

return the collocate job with job job is an opportunistic job

simulate antman with the following assumptions: 1. at most two jobs are packed 2. the memory is enough 3. resource guarantee job will

a means attained-service to that job gittins_index = P/E r_gi = E/P

a means attained-service to that job gittins_index = P/E r_gi = E/P

1-GPU, 2-GPU, 4-GPU, 8-GPU, 12-GPU, 16-GPU, 24-GPU, 32-GPU

1-GPU, 2-GPU, 4-GPU, 8-GPU, 12-GPU, 16-GPU, 24-GPU, 32-GPU

consolidate first, but randomly pick machines; if cross machines, still try to consolidate. if can't consolidate, consider sp

consolidate first, but randomly pick machines; if cross machines, still try to consolidate. if can't consolidate, consider sp

Job's executed time -- priority queue Q0:[0, 30min) Q1:[30min,1h) Q2:[1h, 2h) Q3:[2h, 00) in each queue, jobs are scheduled

Job's executed time -- priority queue Q0:[0, 30min) Q1:[30min,1h) Q2:[1h, 2h) Q3:[2h, 00) in each queue, jobs are scheduled

find gpus whose gpu util < gpu_util_upper

find gpus whose gpu util < gpu_util_upper

ns_num_gpu: num_gpu of job in this node_set node_list: node_set list to expend to num_ns: # of node_set to expend

ns_num_gpu: num_gpu of job in this node_set node_list: node_set list to expend to num_ns: # of node_set to expend

ns_num_gpu: num_gpu of job in this node_set node_list: node_set list to expend to num_ns: # of node_set to expend wh

ns_num_gpu: num_gpu of job in this node_set node_list: node_set list to expend to num_ns: # of node_set to expend wh

gandiva: nodes are put into class(1-GPU, 2-GPU, 4-GPU)

gandiva: nodes are put into class(1-GPU, 2-GPU, 4-GPU)

get the host ip elegantly https://www.chenyudong.com/archives/python-get-local-ip-graceful.html

get model tensor information by model_name return a dict{name, tensors(list)}

get model tensor information by model_name return a dict{name, tensors(list)}

greedy placement:

greedy placement:

Init and create cluster infration entities (switches, nodes) by using class _Switch, _Node

Init and create cluster infration entities (switches, nodes) by using class _Switch, _Node