InferAll/Possible improvement.txt at main · veerkumar/InferAll · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
Possible improvement:
1) Sometime scheduled number of tasks are lesser then the batch size, so we have reevlauate execution time. so it will reduce task execution time.
2) which VM to scale? 4gb, 8GB, 16GB etc? we can have different answer based on cost and latency. so need to choose one

  python    main_service_simulator.py --trace_type NONBATCH --trace_name test --slo 100 --scheduling_type 0 --execution_mode spock


 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 1000 --scheduling_type 0
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 500 --scheduling_type 0
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 300 --scheduling_type 0
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 200 --scheduling_type 0
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 100 --scheduling_type 0

 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 1000 --scheduling_type 1
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 500 --scheduling_type 1
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 300 --scheduling_type 1
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 200 --scheduling_type 1
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 100 --scheduling_type 1

 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 1000 --scheduling_type 2
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 500 --scheduling_type 2
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 300 --scheduling_type 2
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 200 --scheduling_type 2
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 100 --scheduling_type 2


 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 1000 --scheduling_type 0 --execution_mode spock
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 500 --scheduling_type 0 --execution_mode spock
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 300 --scheduling_type 0 --execution_mode spock
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 200 --scheduling_type 0 --execution_mode spock
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 100 --scheduling_type 0 --execution_mode spock

 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 1000 --scheduling_type 2 --execution_mode batch
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 500 --scheduling_type 2 --execution_mode batch
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 300 --scheduling_type 2 --execution_mode batch
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 200 --scheduling_type 2 --execution_mode batch
 time python    main_service_simulator.py --trace_type NONBATCH --trace_name tweet_load --slo 100 --scheduling_type 2 --execution_mode batch


we dont need any prior process arrival model


questions:
Do we need to consider cost for spinning up the VM ?
How do we handle Only VM case? what happen to job which cant be scheduled to VMs


******************************************************************************************************************************

why not fair comparision of spock on latest code
disadvantage 1) In spock VM has queues to while latest solution doesnt have, which make it to default more to lambda which will be costly in longer traces
advantage 2) here based on latency- calls to different models and efficiecny is going. for example lambda is always called with squeezenet while in spock imlemenation single mode with higher efficency.


In spock model, mem selection is not based on SLO.
Most Used VM models
		inception 102
		resnet200 51
		caffenet 0
		squeeznet 0
		vggnet16 0
	Most Used VM memory
		4096 51
		8192 51
		16384 51
	Most Used Top 5 VM Batch
		1 153
		2 0
		4 0
		8 0
		16 0
		32 0
	Most Used Lambda models
		squeeznet 2427
		caffenet 0
		vggnet16 0
		inception 0
		resnet200 0
	Most Used lambda memory
		2048 2427
		256 0
		512 0
		1024 0
		3008 0
	Most Used Top 5 Lambda Batch
		1 2427
		2 0
		4 0
		8 0
		16 0
		32 0

paralelizm in VM is depended on memory usage not CPU as RAM finished way before CPU.

 object-oriented modular discrete event
 https://doc.omnetpp.org/omnetpp/manual/