State information and features about decima-sim

[larissayukimiz]

Also what's the dimension of the per-node embedding? What do you mean by R¹⁶?

[hongzimao]

You found the right place where we initialize the features for the nodes. Note that these features can get updated at each scheduling interaction. You can search for the names in the code space and see how they get called. For example, num_finished_tasks is updated here

decima-sim/spark_env/env.py

Line 266 in c010dd7

node.num_finished_tasks += 1

.

The main agent interaction loop is

decima-sim/train.py

Line 165 in c010dd7

obs, reward, done = env.step(node, use_exec)

, where the model is inovked in

decima-sim/train.py

Lines 28 to 36 in c010dd7

	# invoking the learning model
	node_act, job_act, \
	node_act_probs, job_act_probs, \
	node_inputs, job_inputs, \
	node_valid_mask, job_valid_mask, \
	gcn_mats, gcn_masks, summ_mats, \
	running_dags_mat, dag_summ_backward_map, \
	exec_map, job_dags_changed = \
	actor_agent.invoke_model(obs)

. You can see the features are parsed here (then they go into the policy network described in Figure 6 of the paper):

decima-sim/actor_agent.py

Lines 468 to 472 in c010dd7

	node_inputs, job_inputs, \
	job_dags, source_job, num_source_exec, \
	frontier_nodes, executor_limits, \
	exec_commit, moving_executors, \
	exec_map, action_map = self.translate_state(obs)

.

As in the paper, R^16 is the dimension of each vector of the (example) node feature. It just means there are 16 real numbers.