Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Support interruption for NNdescent #3173

Closed
mdouze opened this issue Dec 13, 2023 · 1 comment
Closed

Support interruption for NNdescent #3173

mdouze opened this issue Dec 13, 2023 · 1 comment
Labels
enhancement

Comments

@mdouze
Copy link
Contributor

mdouze commented Dec 13, 2023

NNdescent cannot be interrupted. It would be nice to support the InterruptCallback

faiss/faiss/impl/AuxIndexStructures.cpp

Line 223 in be12427

bool InterruptCallback::is_interrupted() {

@mdouze
Copy link
Contributor Author

mdouze commented Apr 10, 2024

Context: most slow operations have an explicit way of checking for interruptions.
This is to avoid having to kill the process or handle signals that arrive asynchronously.
The main usage now is in Python to handle keyboard interrupts.
On the C++ side, this is done by periodically checking am InterruptCallback (between 1 and 100 times per second).
See for example

https://github.com/facebookresearch/faiss/blob/main/faiss/IndexHNSW.cpp#L239

faiss/faiss/utils/distances.cpp

Line 256 in 366a814

InterruptCallback::check();

mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent
568a47b 
Summary:
Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: mdouze

Differential Revision: D57300514
@mengdilin mengdilin mentioned this issue May 14, 2024
Closed
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
d14e700 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
cb53b35 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
08267f8 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
1e8b0f1 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
b8b1a57 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
06b39df 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
399388e 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
583db16 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
9443e8e 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
ac2521b 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
6ca732e 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
56b9aa9 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
c6ef289 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
28244cc 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
84e1a28 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
8711090 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
d2c7a11 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
mengdilin added a commit to mengdilin/faiss that referenced this issue May 14, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (facebookresearch#3432)
55e3266 
Summary:

Addresses the issue in facebookresearch#3173 

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514
facebook-github-bot pushed a commit that referenced this issue May 15, 2024
@mengdilin @facebook-github-bot
interrupt for NNDescent (#3432)
558a7c3 
Summary:
Pull Request resolved: #3432

Addresses the issue in #3173

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514

fbshipit-source-id: d343e0a292c35027ffdb8cbd0131e945b9881d63
abhinavdangeti pushed a commit to blevesearch/faiss that referenced this issue Jul 12, 2024
@mengdilin @abhinavdangeti
interrupt for NNDescent (facebookresearch#3432)
787fa29 
Summary:
Pull Request resolved: facebookresearch#3432

Addresses the issue in facebookresearch#3173

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514

fbshipit-source-id: d343e0a292c35027ffdb8cbd0131e945b9881d63
aalekhpatel07 pushed a commit to aalekhpatel07/faiss that referenced this issue Oct 17, 2024
@mengdilin @aalekhpatel07
interrupt for NNDescent (facebookresearch#3432)
932405d 
Summary:
Pull Request resolved: facebookresearch#3432

Addresses the issue in facebookresearch#3173

for `IndexNNDescent`, I see that there is already interrupt implemented for it's [search](https://fburl.com/code/iwn3tqic) API, so I looked into it's `add` API.

For a given dataset nb = 10 mil, iter = 10, K =  32, d = 32 on a CPU only machine reveals that bulk of the cost comes from [nndescent](https://fburl.com/code/5rdb1p5o). For every iteration of `nndescent` takes around ~12 seconds, ~70-80% of the time is spent on `join` method (~10 seconds per iteration) and ~20-30% spent on `update` (~2 second per iteration). Adding the interrupt on the `join` should suffice on quickly terminating the program when users hit ctrl+C (happy to move the interrupt elsewhere if we think otherwise)

Reviewed By: junjieqi, mdouze

Differential Revision: D57300514

fbshipit-source-id: d343e0a292c35027ffdb8cbd0131e945b9881d63
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
enhancement
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@asadoughi @mdouze