Neighbour: 6 neighbour, 10 samples
5/10 still yield good result, cite 2002
6/3 yields good result, cite last year
Control:
Inde:
Dep:
x axis / y axis
Computer:
HP Eliteone AIO
CentOS 7, Linux 3.10, x86-64
Intel(R) CoreTM i7-7700
8 GB memory
Google Chrome 61.0.3163.79 (Official Build) (64-bit)
Computer:
Clevo-based laptop
Ubuntu Gnome 17.04, Linux 4.13, x86-64
Intel(R) CoreTM i7-6700HQ
16 GB DDR4 memory (limited by browser)
Chromium 64.0.3282.119 (Official Build) (64-bit)
Computer used:
Dell AIO
CentOS 7, Linux 3.10, x86-64
Intel(R) CoreTM i5-3470S
8 GB memory
Google Chrome 61.0.3163.79 (Official Build) (64-bit)
Memory: which part uses ram DONE
Control: Dataset
Inde: Algorithm phases (each at its best setting unless non-affecting)
Dep: Ram use
Can cut out the last part?
How looks can affect interp DONE:
Control: Hybrid settings (enough to get good looks)
Inde: Dataset (small vs medium), multiple runs
Dep: Looks, when none of the class got selected
Values of refinement at the end of interp: Knows good refinement times, might depends on dataset
Control: Hybrid, dataset (large), settings
Inde: No of refinement at the end of interp
Dep: Time, looks, stress
Note down the sweet spot.
Pivot hitrate: Show how good pivot is vs time spnent
Control: Hybrid, dataset (large), settings
Inde: No of pivot, multiple runs
Dep: Hit rate, non-hit dist difference, looks
Note down the sweet spot.
Pivot time: Knows weather to bruteforce of pivot
10 runs, sufficient start
time spent finding parent (pivot and brute), time spent placing object, stress after / data size
How many full runs
10 runs, best settings so far
total time spent, looks, stress, average velo changes/ iterations
Multiple dataset: DONE
iterations / velocity changes, stress
Note down the sweet spot.
For hybrid only the end
Multiple dataset:
Dataset size / time
Talk about GC DONE
Talk about too much mem -> run time higher anyway. If u need fast, prolly use other languages.
Neighbour: 6 neighbour, 10 samples
5/10 still yield good result, cite 2002
6/3 yields good result, cite last year
- Control:
- Inde:
- Dep:
x axis / y axis
Computer:
- HP Eliteone AIO
- CentOS 7, Linux 3.10, x86-64
- Intel(R) CoreTM i7-7700
- 8 GB memory
- Google Chrome 61.0.3163.79 (Official Build) (64-bit)
Computer:
- Clevo-based laptop
- Ubuntu Gnome 17.04, Linux 4.13, x86-64
- Intel(R) CoreTM i7-6700HQ
- 16 GB DDR4 memory (limited by browser)
- Chromium 64.0.3282.119 (Official Build) (64-bit)
Computer **used**:
- Dell AIO
- CentOS 7, Linux 3.10, x86-64
- Intel(R) CoreTM i5-3470S
- 8 GB memory
- Google Chrome 61.0.3163.79 (Official Build) (64-bit)
---------------------
Memory: which part uses ram **DONE**
- Control: Dataset
- Inde: Algorithm phases (each at its best setting unless non-affecting)
- Dep: Ram use
Can cut out the last part?
---------------------
How looks can affect interp **DONE**:
- Control: Hybrid settings (enough to get good looks)
- Inde: Dataset (small vs medium), multiple runs
- Dep: Looks, when none of the class got selected
---------------------
Values of refinement at the end of interp: Knows good refinement times, might depends on dataset
- Control: Hybrid, dataset (large), settings
- Inde: No of refinement at the end of interp
- Dep: Time, looks, stress
Note down the sweet spot.
---------------------
Pivot hitrate: Show how good pivot is vs time spnent
- Control: Hybrid, dataset (large), settings
- Inde: No of pivot, multiple runs
- Dep: Hit rate, non-hit dist difference, looks
Note down the sweet spot.
---------------------
Pivot time: Knows weather to bruteforce of pivot
10 runs, sufficient start
time spent finding parent (pivot and brute), time spent placing object, stress after / data size
---------------------
How many full runs
10 runs, best settings so far
total time spent, looks, stress, average velo changes/ iterations
Multiple dataset: **DONE**
iterations / velocity changes, stress
Note down the sweet spot.
For hybrid only the end
Multiple dataset:
Dataset size / time
Talk about GC **DONE**
Talk about too much mem -> run time higher anyway. If u need fast, prolly use other languages.
Neighbour sampling cache distances?
Poker 5000, no render
With cache
15994ms
calc takes 3945ms
force takes 9724ms
JS heap reach 125MB before simulation started
128MB after sim + gc
Without cache
12510ms
calc takes 1995ms
force takes 12203ms (calc included)
JS heap mem never reach 50MB even without manually invoking GC.
Conclusion: やめて くれ
Neighbour sampling cache distances?
Poker 5000, no render
With cache
15994ms
calc takes 3945ms
force takes 9724ms
JS heap reach 125MB before simulation started
128MB after sim + gc
Without cache
12510ms
calc takes 1995ms
force takes 12203ms (calc included)
JS heap mem never reach 50MB even without manually invoking GC.
**Conclusion: やめて くれ**
Default 42805 ms
Ram 747MB peak, 418 MB at the end with GC
Tweaked 38605 ms
Ram 492MB peak, 238 MB at the end with GC
Tweaked without adding links 37841 ms
Ram 95MB peak, 59 MB at the end with GC
Link Poker 3000 no render
Default 42805 ms
Ram 747MB peak, 418 MB at the end with GC
Tweaked 38605 ms
Ram 492MB peak, 238 MB at the end with GC
Tweaked without adding links 37841 ms
Ram 95MB peak, 59 MB at the end with GC
10,000 Until 0.65
34894.06 ms (58 its)
Post Stress: 0.3019698281499924
100,000 Until 0.65
670033.4650000001 ms (98 its)
699007.5100000001 ms (102 its)
**Neighbour**
10,000 Until 0.65
34894.06 ms (58 its)
Post Stress: 0.3019698281499924
100,000 Until 0.65
670033.4650000001 ms (98 its)
699007.5100000001 ms (102 its)
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Neighbour: 6 neighbour, 10 samples
5/10 still yield good result, cite 2002
6/3 yields good result, cite last year
x axis / y axis
Computer:
Computer:
Computer used:
Memory: which part uses ram DONE
Can cut out the last part?
How looks can affect interp DONE:
Values of refinement at the end of interp: Knows good refinement times, might depends on dataset
Note down the sweet spot.
Pivot hitrate: Show how good pivot is vs time spnent
Note down the sweet spot.
Pivot time: Knows weather to bruteforce of pivot
10 runs, sufficient start
time spent finding parent (pivot and brute), time spent placing object, stress after / data size
How many full runs
10 runs, best settings so far
total time spent, looks, stress, average velo changes/ iterations
Multiple dataset: DONE
iterations / velocity changes, stress
Note down the sweet spot.
For hybrid only the end
Multiple dataset:
Dataset size / time
Talk about GC DONE
Talk about too much mem -> run time higher anyway. If u need fast, prolly use other languages.
Neighbour sampling cache distances?
Poker 5000, no render
With cache
15994ms
calc takes 3945ms
force takes 9724ms
JS heap reach 125MB before simulation started
128MB after sim + gc
Without cache
12510ms
calc takes 1995ms
force takes 12203ms (calc included)
JS heap mem never reach 50MB even without manually invoking GC.
Conclusion: やめて くれ
Link Poker 3000 no render
Default 42805 ms
Ram 747MB peak, 418 MB at the end with GC
Tweaked 38605 ms
Ram 492MB peak, 238 MB at the end with GC
Tweaked without adding links 37841 ms
Ram 95MB peak, 59 MB at the end with GC
^ continue with default
เล่าเรื่อง
Metric ที่จะหยุด run:
Stress วัดไม่ได้ นานเกิน
Static number of iterations: not good
Velocity changes เกี่ยวข้องกับ stress และ layout change
ใช้ Velo change เข้าถึง threshold แล้วเลิก, ค่าขึ้นกับ algo, data
Metric: Looks, Stress, Time, Memory
Memory: ดู graph bottleneck
Time: เพราะ hybrid เพื่อประหยัด neighbour ตอนจบ อาจจะ interp สวยนาน แต่สุดท้ายไม่ช่วย neighbour(show ด้วย)
Looks เพราะ Stress ดูน้อยแต่ไม่ได้แปลว่าดูดี
Test โลดด
Conclude best practice
Other things to talk about:
Neighbour
10,000 Until 0.65
34894.06 ms (58 its)
Post Stress: 0.3019698281499924
100,000 Until 0.65
670033.4650000001 ms (98 its)
699007.5100000001 ms (102 its)
Max data size:
470K neighbour is fine
600K neighbour ram full, swapping mem entire system. Hybrid without 3rd phase didn't.
1M crash on load