[profiler] Speed up postprocessing#58021

Closed

ilia-cher wants to merge 10 commits intogh/ilia-cher/106/basefrom

gh/ilia-cher/106/head

Contributor

ilia-cher commented May 11, 2021 •

edited

Loading

Stack from ghstack:

[profiler][small] Add skip_first parameter to the default schedule #58025 [profiler][small] Add skip_first parameter to the default schedule
[profiler] Speed up postprocessing #58021 [profiler][small] Speed up postprocessing
[profiler][small] Handle empty trace #58013 [profiler][small] Handle empty trace
[profiler][resend] Add cuda memory and distributed metadata #58010 [profiler][resend] Add cuda memory and distributed metadata

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:

import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))

After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s

Differential Revision: D28347217


          [profiler][small] Speed up postprocessing

fd1adcd

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id=0, thread=0, name="child", start_us=st+10, end_us=st+90))

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

[ghstack-poisoned]

ilia-cher requested review from albanD and soulitzer as code owners

May 11, 2021 07:44

This was referenced May 11, 2021

[profiler][resend] Add cuda memory and distributed metadata #58010

Closed

[profiler][small] Handle empty trace #58013

Closed

facebook-github-bot added the cla signed label

Contributor

facebook-github-bot commented May 11, 2021 •

edited

Loading

💊 CI failures summary and remediations

As of commit cdc3f46 (more details on the Dr. CI page):

💚 💚 Looks good so far! There are no failures yet. 💚 💚

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

Contributor Author

ilia-cher commented May 11, 2021

Contributor Author

ilia-cher commented May 11, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.


          Update on "[profiler][small] Speed up postprocessing"

c665237

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id=0, thread=0, name="child", start_us=st+10, end_us=st+90))

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

ilia-cher pushed a commit that referenced this pull request


          [profiler][small] Speed up postprocessing

35f98e5

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id=0, thread=0, name="child", start_us=st+10, end_us=st+90))

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

ghstack-source-id: 07f47c2
Pull Request resolved: #58021

Contributor Author

ilia-cher commented May 11, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ilia-cher mentioned this pull request

[profiler][small] Add skip_first parameter to the default schedule #58025

Closed


          Update on "[profiler][small] Speed up postprocessing"

dd3e4b4

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 11, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.


          Update on "[profiler][small] Speed up postprocessing"

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 11, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ilia-cher requested review from bitfort, chaekit, gdankel and xuzhao9

May 11, 2021 08:33

dgl-intel pushed a commit to dgl-intel/pytorch that referenced this pull request


          [profiler][small] Speed up postprocessing

f10e77c

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id=0, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id=0, thread=0, name="child", start_us=st+10, end_us=st+90))

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

ghstack-source-id: 1e2335d
Pull Request resolved: pytorch#58021


          Update on "[profiler][small] Speed up postprocessing"

cdcafa9

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 11, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.


          Update on "[profiler][small] Speed up postprocessing"

bb162e3

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 11, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.


          Update on "[profiler][small] Speed up postprocessing"

a344ee4

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 11, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.


          Update on "[profiler][small] Speed up postprocessing"

a8a110b

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 12, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.


          Update on "[profiler][small] Speed up postprocessing"

dd34d6a

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 12, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.


          Update on "[profiler][small] Speed up postprocessing"

cdc3f46

Summary:
Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Differential Revision: [D28347217](https://our.internmc.facebook.com/intern/diff/D28347217)

[ghstack-poisoned]

Contributor Author

ilia-cher commented May 12, 2021

@ilia-cher has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

facebook-github-bot closed this in

cdf161c

Contributor

facebook-github-bot commented May 12, 2021

@ilia-cher merged this pull request in cdf161c.

facebook-github-bot added the Merged label

facebook-github-bot deleted the gh/ilia-cher/106/head branch

May 15, 2021 14:23

krshrimali pushed a commit to krshrimali/pytorch that referenced this pull request


          [profiler][small] Speed up postprocessing (pytorch#58021)

Summary:
Pull Request resolved: pytorch#58021

Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Reviewed By: gdankel

Differential Revision: D28347217

Pulled By: ilia-cher

fbshipit-source-id: d62da3400009f1fa8cb41a11a828aa8307f190bf

ilia-cher changed the title ~~[profiler][small] Speed up postprocessing~~ [profiler] Speed up postprocessing

laurentdupin pushed a commit to laurentdupin/pytorch that referenced this pull request


          [profiler][small] Speed up postprocessing (pytorch#58021)

652bbc6

Summary:
Pull Request resolved: pytorch#58021

Improve complexity of _remove_dup_nodes function

Test Plan:
using trivial microbenchmark:
```
import torch
from torch.autograd.profiler import *
import time

evts = EventList()
id_cnt = 0
for r in range(10*1000):
    st = r * 1000
    evts.append(FunctionEvent(id_cnt, thread=0, name="parent", start_us=st, end_us=st+100))
    evts.append(FunctionEvent(id_cnt+1, thread=0, name="parent", start_us=st+1, end_us=st+99))
    evts.append(FunctionEvent(id_cnt+2, thread=0, name="child", start_us=st+10, end_us=st+90))
    id_cnt+=3

st = time.time()
evts._build_tree()
print("Elapsed: {:.3f}s".format(time.time() - st))
```

```
After:
python test_prof.py
Elapsed: 0.203s

Before:
python test_prof.py
Elapsed: 3.653s
```

Reviewed By: gdankel

Differential Revision: D28347217

Pulled By: ilia-cher

fbshipit-source-id: d62da3400009f1fa8cb41a11a828aa8307f190bf

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

albanD Awaiting requested review from albanD

soulitzer Awaiting requested review from soulitzer

gdankel Awaiting requested review from gdankel

xuzhao9 Awaiting requested review from xuzhao9

bitfort Awaiting requested review from bitfort

chaekit Awaiting requested review from chaekit

Labels

cla signed Merged