logs

based on results so far

series 2
- good results for g
outcome:
- fix other rewards for PG, IMPALA and PPO and then find a bad case for DQN and you are done

series 3, 4, 5:
test of reduced space
outcome: should be repeated because of the wrong reward

series 6, 7, 8:
repeat of 3, 4, 5
outcome:    "backlog_services_requests",
            "nodes_requests",
            "nodes_usages"
seems to be enough
TODO_1: test this obs space

series 9:
test of reward v
outcome: all zero
TODO_2: reward v meaning should be understood and then re-test

series 10:
TODO_1: 
test of reward p with reduced search space
outcome: we can continue with this observation space

series 11:
test of reward c
outcome: Seems to work - more elaboration is needed
TODO_4: test reward c on benchmark cluster 7

series 12:
test of reward p on smaller dataset - cluster 7
outcome: working completely

series 13:
test of reward g on smaller dataset - cluster 7
outcome: not working at all
TODO_5: test reward g on cluster 7 after checking search space

series 14:
series 12 (test of reward p) with the following space
"backlog_services_requests_frac",
            "nodes_requests",
            "nodes_usages"
outcome: new state space is validated

series 15:
series 13 with the following space, reward g
"backlog_services_requests_frac",
            "nodes_requests",
            "nodes_usages"
outcome:
TODO_6: validate the new state-space in code

sereis 15.1:
TODO_6: validate the new state-space in code
outcome: done

series 16:
TODO_4: test reward c on cluster 7
outcome: seems to work
TODO_7: test it without normalization
TODO_8: test it with changed cap

series 17:
TODO_5: test reward g on cluster 7 with changed cap
outcome: Interestingly it reduces reward cv but not the g itself
TODO_9: test it without normalization

series 18:
TODO_3: one experiment with reward cv on cluster 7
outcome: Seems not working
TODO_10: test it without normalization

series 19:
TODO_7: test reward c on cluster 7 without normalization

series 20:
TODO_8: test reward c on cluster 7 with changed cap and without normalization

series 21:
TODO_9: Test reward cv on cluster 7 without normalization
outcome: there was a bug and should be repeated
TODO_11: repeat withouth the bug

series 22:
TODO_10: Test reward g on cluster 7 without normalization
outcome: it is working, the results should be compared with g after TODO_11

series 23:
Test reward p on cluster 7 without normalization
outcome: mistake - not fixed
TODO_12: repeat

series 24:
TODO_11: Test reward cv on cluster 7 without normalization with bug fix
outcome: fixed on reward cv as it is performing much better than cv
Fixed: reward cv

series 25:
TODO_12: Test reward p on cluster 7 without normalization with bug fix
outcome: works well

series 26:
a two node minimal cluster with similar sized jobs one reward p

series 27:
a two node minimal cluster with similar sized jobs one reward g
outchome: substitute cv with g -> this is wrong g is differnt from cv TODO_13

series 28:
a two node minimal cluster with similar sized jobs one reward cv
outcome: works perfectly

series 29:
a two node minimal cluster with similar sized jobs one reward c
outcome: we left reward c

series 30:
a four node minimal cluster with similar sized jobs one reward p
outcome: works perfectly

series 31:
a four node minimal cluster with similar sized jobs one reward cv
outcome: works even better

series 32:
a four node minimal cluster with similar sized jobs one reward c
outcome: we left reward c

series 33:
a eight node minimal cluster with similar sized jobs one reward p
outcome: works perfectly

series 34:
a eight node minimal cluster with similar sized jobs one reward cv
outcome: works perfectly

series 35:
a eight node minimal cluster with similar sized jobs one reward c
outcome: we left reward c

series 36:
TODO_13: a two node minimal cluster with similar sized jobs one reward g
outcome: not working
TODO_14: try it with normalization

series 37:
TODO_13: a four node minimal cluster with similar sized jobs one reward g
outcome: not working
TODO_14: try it with normalization

series 38:
TODO_13: a eight node minimal cluster with similar sized jobs one reward g
outcome: not working
TODO_14: try it with normalization

series 39:
a four node on reward cv with shorter steps
outcome: perfect but shorter steps was forgotten
TODO_15: repeat with shorter steps

series 40:
a four node on reward p with shorter steps
outcome: wrong reward, repeat
TODO_16: repeat

series 41:
TODO_14: a two node minimal cluster with similar sized jobs one reward g with normalization
outcome: not working
TODO_16: in depth analysis with debug of reward g

series 42:
TODO_14: a four node minimal cluster with similar sized jobs one reward g with normalization
outcome: not working
TODO_16: in depth analysis with debug of reward g

series 43:
TODO_14: a eight node minimal cluster with similar sized jobs one reward g with normalization
outcome: not working
TODO_16: in depth analysis with debug of reward g

series 44:
TODO_15: a two node on reward cv with shorter steps
outcome: good but small
TODO_17: run on four and eight cluster

series 45:
TODO_15: a two node on reward p with shorter steps
outcome: good but small
TODO_17: run on four and eight cluster

series 46:
TODO_17: a four on reward cv with shorter steps
outcome: fixed for session

series 47:
TODO_17: a four node on reward p with shorter steps
outcome: fixed for session

series 48:
TODO_17: an eight node on reward cv with shorter steps
outcome: fixed for session

series 49:
TODO_17: an eight node on reward p with shorter steps
outcome: fixed for session

series 50:
combination of reward cv and p to find a middle ground with normalizing on four node cluster - reward cv only
outcome: worked as expected

series 51:
combination of reward cv and p to find a middle ground with normalizing on four node cluster- reward cv and p
outcome: worked as expected

series 52:
combination of reward cv and p to find a middle ground with normalizing on four node cluster - reward p only
outcome: worked as expected

series 53:
combination of reward cv and p to find a middle ground with normalizing on eight node cluster - reward cv only
outcome: not working

series 54:
combination of reward cv and p to find a middle ground with normalizing on eight node cluster - reward cv and p
outcome: not working

series 55:
combination of reward cv and p to find a middle ground with normalizing on eight node cluster- reward p only
outcome: working but useless as 53 and 54 are also not working

series 56:
none-fixed sized job on the cluster of size four nodes reward p
outcome: working as expected
TODO_18: more diverse set -> 8 types of jobs

series 57:
none-fixed sized job on the cluster of size four nodes reward cv and p
outcome: working as expected
TODO_18: more diverse set -> 8 types of jobs

series 58:
none-fixed sized job on the cluster of size four nodes reward cv
outcome: working as expected
TODO_18: more diverse set -> 8 types of jobs

series 59:
TODO_18: 8 type none-fixed sized job on the cluster of size four nodes reward p
outcome: works

series 60:
TODO_18: more diverse set -> 8 types of jobs on reward cv and p
outcome: wrong cluser
TODO_19: repeat


series 61:
TODO_18: more diverse set -> 8 types of jobs on reward cv
outcome: works

series 62:
TODO_19: repeat 60
outcome: works

series 63:
TODO_19: repeat 60
double weights of consolidation
outcome: not work

series 64:
check different weighting between cv and p to find the best hyperparameters
weighting that make a difference:
outcome: equal wights are still the bests
TODO_20: smaller servers and multiples of two
TODO_21: more testing jobs
TODO_22: PPO
TODO_23: IMPALA

series 65:
TODO_20: Make servers smaller and multipls of two to be consistent with cloud resources and with all three reward weightenings

series 66:
TODO_22: PPO

series 67:
TODO_23: IMPALA

series 68:
DQN

series 69:
A2C

series 70:
TODO_20: Make servers smaller and multipls of two to be consistent with cloud resources and with all three reward weightenings



series 71-75:
trying to sync values
PG, PPO, IMPALA, DQN

series 76-79:
repeat of the former series



series 80-83:
make training batch bigger to make DQN worse
outcome: no use


series 84-85:
8 node cluster and test for IMPALA and DQN
outcome: no use:

series 86-87:
8 node cluster with smaller servers and test for IMPALA and DQN
outcome: no use:


series 88-89:
4 node cluster with smaller servers setting and test for IMPALA and DQN
outcome: DQN

series 90-91:
4 node cluster with previous setting and test for IMPALA and DQN
outcome: no use

series 92-93:
4 node cluster and test for IMPALA and DQN with longer episode length
outcome: still DQN is better

series 94-95:
8 node cluster and test for IMPALA and DQN w100ith with longer episode length
outcome: still DQN is better but IMPALA is better in finding the balance

series 96-97:
reeat of 92-93 with shuffling - 4 node cluster and test for IMPALA and DQN with longer episode length
outcome: without shuffle showed better curves in DQN no use

series 98-99:
reeat of 94-95 with shuffling - 8 node cluster and test for IMPALA and DQN with with longer episode length
outcome: without shuffle showed better curves in DQN, no use

series 100-101:
8 node cluster for IMPALA and DQN with smaller episode length but bigger, tighter jobs
outcome: shows some promising curve, DQN is paper ready wit hreward g and u

series 102:
follow the DQN path of 100 and make the span clearer
gridsearch for big and small servers

series 103:
follow the DQN path of 100 and make the span clearer
gridsearch on differnt wieghting between rewards
outcome: 0.75, 0.75 optimal

series 104:
follow the DQN path of 100 and make the span clearer
gridsearch on differnt episode lenghth
outcome: no real different, do it on 100

series 105:
not using homogeneous weighting on eight node 2, 0.75,
outcome: no use

series 106:
same result on 8 node cluster but only the optimal
outcome: finalized for paper
FINALIZED

series 107:
same resutls on 4 node cluster
outcome: great convergance, but needs to play with reward value to get to a tradeoff

series 108:
same resutls on 16 node cluster
outcome: great convergance, but needs to play with reward value to get to a tradeoff

series 109:
Redo 107 for best tradeoff
outcome: weighting does not give tradeoff, I should go for normalization variables
TODO_20: work on the normalizing

series 110:
Redo 108 for best tradeoff
outcome: weighting does not give tradeoff, I should go for normalization variables
TODO_20: work on the normalizing

series 111:
An experiment with 32 servers
outcome: promising but needs to work on the tradeoff with weighthing
TODO_20: work on the normalizing

series 112:
redo 109 for TODO_20 on 4 node
outcome: finalized for the paper
FINALIZED

series 113:
redo 110 for TODO_20 on 16 node
outcome: promissing just neeed some more playing with variables
TODO_21

series 114:
redo 111 for TODO_20 on 32 node
outcome: promissing just neeed some more playing with variables
TODO_21

series 115:
redo of 113 for TODO_21 for better curve DQN on 16 nodes and deacreased consolidation effect
outcome: finalized for the paper

series 116:
test on IMPALA with new scheme, all cluster sizes
outcome: not bad but not enough

series 117:
test on PG with new scheme, all cluster sizes
outcome: bye bye PG

series 118:
test on PPO with new scheme, all cluster sizes
outcome: bye bye PPO

series 119:
test on DQN with new scheme, all cluster sizes
outcome: redo for mistake in experiments

series 120:
test on IMPALA with reduced rewards new scheme, all cluster sizes
outcome: not good, going for impala parameter search

series 121:
test on DQN with new scheme, all cluster sizes
outcome: finalized for the paepr
FINALIZED

series 122:
gridsearch on impala algorithms on entropy_coeff and entropy_coeff_schedule 
outcome: not good, going for smaller values

series 123:
gridsearch on impala algorithms on entropy_coeff and entropy_coeff_schedule 
repeat of 122 with smaller values
outcome: no result

series 124:
test on other algorithms
outcome: no result

series 125:
test on impala with filter
outcome: no result

series 126:
test ppp with filter
outcome: no result

series 127:
gridsearch on ppo
outcome: no result

series 128:
Conclude here on one type and go for tests - only continue if there is a very clear difference, otherwise do this just as a hobby with a cap of 30 min per day

series 129:
Ask question about Apex and run one experiements with each of them but only one!

series 120:
making the workloads tighter by making the service times longer

series 121:
test on other algorithms


series 114:
finalize with 4, 8, 16, 32 on the paper

series 114:
brute force all other algorithms for one better answer until the deadline


series 103:
Either another algorithm
criteria to check:
1. weighting
2. episode length
3. training iterations

series 103:
Or go with the DQN with smaller sizes

-----------------

series 96:
what factor in https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9591490 for IMPALA better PPO


series 96:
check how the system works


Fixed:
we are fixed on the reward cv and p for their performance
show 84-87-101 to Joseph

regx:
runs: ((92/)|(93/)).*.(?=experiments/2)
reward to search: (tune/custom_metrics/num_consolidated_avg_mean)|(tune/custom_metrics/reward_p_mean)|(tune/custom_metrics/reward_cv_mean)|(tune/custom_metrics/reward_u_mean)|tune/custom_metrics/reward_g_mean