Stream: git-wasmtime

Topic: wasmtime / PR #8303 optimize perfermance of fd_read/fd_write

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 06:08):

liutao-liu requested wasmtime-core-reviewers for a review on PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 06:08):

liutao-liu opened PR #8303 from liutao-liu:opt_read_write to bytecodealliance:main:

Compared to native(C/C++), fd-read can only reach 20% of native. I will divide the performance hotspots into the following two categories:,

  1. Wasi implementation of fd-read(approximately 30% of the hotspots), such as the cost of redundant memory copies, the cost of inserting and deleting HostInputStream into WasiView, and the cost of accessing guest memory(merged PR);;
  2. Wasm program call host function(approximately 40% of the hotspots): I am not very familiar with this mechanism yet, and currently I cannot understand the reasons behind the hotspots here;

This PR is mainly optimized for the first type of performance hotspot:

  1. Avoid unnecessary copying;
  2. Avoid inserting and deleting HostInputStream intoWasiView;

The perfermance hotspots of fd_write is the same fd_read;

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 06:08):

liutao-liu requested alexcrichton for a review on PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 06:14):

liutao-liu edited PR #8303:

Compared to native(C/C++), fd-read can only reach 20% of native. I will divide the performance hotspots into the following two categories:,

  1. Wasi implementation of fd-read(approximately 30% of the hotspots), such as the cost of redundant memory copies, the cost of inserting and deleting HostInputStream into WasiView, and the cost of accessing guest memory(merged PR);
  2. Wasm program call host function(approximately 40% of the hotspots): I am not very familiar with this mechanism yet, and currently I cannot understand the reasons behind the hotspots here;

This PR is mainly optimized for the first type of performance hotspot:

  1. Avoid unnecessary copying;
  2. Avoid inserting and deleting HostInputStream intoWasiView;

The perfermance hotspots of fd_write is the same fd_read:
https://github.com/bytecodealliance/wasmtime/blob/86dcae460b5437703570e52253cc1b7d5bad8597/crates/wasi/src/preview1.rs#L1682

https://github.com/bytecodealliance/wasmtime/blob/86dcae460b5437703570e52253cc1b7d5bad8597/crates/wasi/src/preview1.rs#L244

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 06:27):

liutao-liu updated PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 06:28):

liutao-liu updated PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 06:42):

liutao-liu updated PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 18:25):

alexcrichton submitted PR review:

Thanks for the PR!

I've left some notes below, but I'll additionally point out that this is specifically ignoring the blocking_mode now. Previously that was "respected" to a certain degree, but now it's being ignored. I believe that's ok because that's what the previous implementation in wasi-common did anyway as OSes tend to ignore nonblocking on files. That being said the current state of the code is a bit confusing where the blocking_mode is sometimes used and sometimes isn't.

I mostly wanted to point this out in case anyone runs into this in the future. I'll follow up with a PR after this one to clean this up. You don't have to clean it up here yourself @liutao-liu

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 18:25):

alexcrichton submitted PR review:

Thanks for the PR!

I've left some notes below, but I'll additionally point out that this is specifically ignoring the blocking_mode now. Previously that was "respected" to a certain degree, but now it's being ignored. I believe that's ok because that's what the previous implementation in wasi-common did anyway as OSes tend to ignore nonblocking on files. That being said the current state of the code is a bit confusing where the blocking_mode is sometimes used and sometimes isn't.

I mostly wanted to point this out in case anyone runs into this in the future. I'll follow up with a PR after this one to clean this up. You don't have to clean it up here yourself @liutao-liu

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 18:25):

alexcrichton created PR review comment:

This .expect("...") will need to be handled (e.g. we can't panic here). The None happens here with shared memories at the core wasm level, meaning that this'll need to use the copy_from_slice method. (e.g. read data into a buffer on the host and then copy the contents of the buffer to the guest)

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 18:25):

alexcrichton created PR review comment:

It's ok to omit this line, this isn't actually doing anything. As-written this is copying data from the guest to the host, truncating it, and then deallocating it.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 18:25):

alexcrichton created PR review comment:

Also given the inner loops below the outer loop can be removed too.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 05 2024 at 18:25):

alexcrichton created PR review comment:

This limitation is only part of the Host*Stream traits, so it's ok to remove this NOTE as well as the .min(4096) here too.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 03:40):

liutao-liu submitted PR review.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 03:40):

liutao-liu created PR review comment:

got

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 03:40):

liutao-liu submitted PR review.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 03:40):

liutao-liu created PR review comment:

got

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 03:51):

liutao-liu updated PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 04:23):

liutao-liu updated PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 04:24):

liutao-liu submitted PR review.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 04:24):

liutao-liu created PR review comment:

got

view this post on Zulip Wasmtime GitHub notifications bot (Apr 06 2024 at 04:46):

liutao-liu updated PR #8303.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 10 2024 at 01:10):

liutao-liu commented on PR #8303:

@alexcrichton Your review comments have been completed，this pr can be merged.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 11 2024 at 15:11):

alexcrichton submitted PR review.

view this post on Zulip Wasmtime GitHub notifications bot (Apr 11 2024 at 15:38):

alexcrichton merged PR #8303.

Last updated: Dec 23 2024 at 13:07 UTC