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.

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

Speeding up v8 heap snapshots #702

Merged
merged 27 commits into from Jul 27, 2023
Merged
Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
27 commits
Select commit Hold shift + click to select a range
7d91e99
Speeding up v8 heap snapshots
jdapena Jun 20, 2023
485d04f
Apply suggestions from Rob Palmer
jdapena Jun 21, 2023
3dbf61b
Fix lint issues
jdapena Jun 22, 2023
b6860f4
Apply suggestions from Joyee Cheung
jdapena Jun 27, 2023
77c2744
More clarifications of Node.js specific parts
jdapena Jun 27, 2023
9eac65c
Added Joyee Cheung to the initial paragraph list of contributors
jdapena Jun 27, 2023
53d996a
More accurate description of --heapsnapshot-near-heap-limit implement…
jdapena Jun 28, 2023
a0518ff
More details of source position caching fix
jdapena Jun 28, 2023
f02741a
Reduce ambiguity of development and production JS code reference
jdapena Jun 28, 2023
bc0bcee
Rewrite description of V8 heap limit configuration.
jdapena Jun 28, 2023
91525cd
Apply suggestions from Joyee code review
jdapena Jul 11, 2023
69a958e
Remove unneeded reference from heap snapshot optimization post
jdapena Jul 11, 2023
1689958
Crop images to improve readability
jdapena Jul 11, 2023
3a1cdd9
Apply suggestions from @syg code review
jdapena Jul 12, 2023
91dc5a8
Apply suggestions from @syg in pull request.
jdapena Jul 12, 2023
3899164
Added @syg to the list of contributors in first paragraph
jdapena Jul 12, 2023
de70d28
Do not use specific names explaining the heap snapshot performance in…
jdapena Jul 12, 2023
3811c74
Remove sponsorship reference in heap optimizations post
jdapena Jul 12, 2023
c89f34f
Provided function snippet and value examples of new Hash function
jdapena Jul 13, 2023
239123a
speeding-up-heap-snapshots.md: apply suggestions from Joyee code review
jdapena Jul 18, 2023
3785dbe
Update speeding-up-v8-heap-snapshot.md: replace obscure reference to …
jdapena Jul 18, 2023
c603786
speeding-up-v8-snapshots.md: following recommendations from Joyee, si…
jdapena Jul 18, 2023
979a434
speeding-up-heap-snapshots.md: remove unused image
jdapena Jul 18, 2023
ca5bf19
speeding-up-v8-heap-snapshots.md: apply suggestions from @syg
jdapena Jul 19, 2023
bdd5b96
speeding-up-v8-heap-snapshots.md: manually apply @syg proposed changes
jdapena Jul 19, 2023
8f083a8
speeding-up-v8-heap-snapshots.md: update publication date
jdapena Jul 19, 2023
b065c43
speeding-up-v8-heap-snapshots.md: drop WPA session
jdapena Jul 26, 2023
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
jdapena marked this conversation as resolved.
Show resolved Hide resolved
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.
161 changes: 161 additions & 0 deletions src/blog/speeding-up-v8-heap-snapshots.md
@@ -0,0 +1,161 @@
---
title: 'Speeding up V8 heap snapshots'
description: 'This post about V8 heap snapshots presents some performance problems found by Bloomberg engineers, and how we fixed them to make JavaScript memory analysis faster than ever.'
author: 'Jose Dapena Paz'
date: 2023-06-23
tags:
- memory
- tools
---
*This blog post has been authored by José Dapena Paz (Igalia), with contributions from Jason Williams (Bloomberg), Ashley Claymore (Bloomberg), and Rob Palmer (Bloomberg).*

In this post about V8 heap snapshots, I will talk about some performance problems found by Bloomberg engineers, and how we fixed them to make JavaScript memory analysis faster than ever.

## The Problem

Ashley Claymore was working on diagnosing a memory leak in a JavaScript application. It was failing with *Out-Of-Memory* errors. Despite the process having access to plenty of system memory, V8 places a hard limit on the amount of memory dedicated to the garbage-collected heap from which all JavaScript objects are allocated. This V8 heap limit (~1400MB) was being hit.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

The standard way to debug a routine memory leak scenario like this is to capture a heap snapshot and then inspect the various summaries and object attributes using DevTools "Memory" tab to find out what is consuming the most memory. In DevTools, you click the round button marked "Take heap snapshot" to perform the capture. For Node.js applications, you can [trigger the snapshot](https://nodejs.org/en/docs/guides/diagnostics/memory/using-heap-snapshot) programatically using this API:
jdapena marked this conversation as resolved.
Show resolved Hide resolved

```javascript
require('v8').writeHeapSnapshot();
```

The desire was to capture several snapshots at different points in the application's life, so that DevTools Memory viewer could be used to show the difference between each. The problem was that capturing a single full-size (500MB) snapshot was taking **over 30 minutes** alone!
jdapena marked this conversation as resolved.
Show resolved Hide resolved

It was this slowness in the memory analysis workflow that we needed to solve.

## Narrowing down the Problem
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Jason Williams started investigating the issue using some V8 parameters. As described in the previous post, V8 has some nice command line parameters that can help with that. These options were used to curate the heap snapshots , simplify the reproduction, and improve observability:
jdapena marked this conversation as resolved.
Show resolved Hide resolved
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd just stick with "Bloomberg engineers" instead of specific names.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
Jason Williams started investigating the issue using some V8 parameters. As described in the previous post, V8 has some nice command line parameters that can help with that. These options were used to curate the heap snapshots , simplify the reproduction, and improve observability:
Jason Williams started investigating the issue using some V8 parameters. These options were used to create the heap snapshots, simplify the reproduction, and improve observability:

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Other than --max-old-space-size, all the flags below are Node-specific. What do you recommend folks do that are trying to debug memory leaks not on Node?

Also what does "improve observability" mean?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Re-reading this section after reading the entire post, I'm unclear on the point of listing the flags in detail. It seems almost incidental: here are the flags that we were using to capture heap snapshots, and where we observed surprising slowdowns. Could this section be shortened? Perhaps I'm missing the intention though?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Initially the problem was detected just extracting regular heap snapshots from DevTools.

But using these command line arguments allowed to get a detailed breakdown of what was happening, and also allowed to reproduce the heap snapshot performance problem several times, without requiring to use a remote DevTools connection.

So this is part of the investigation steps (increasing and simplifying reproducibility).

In general, the intent of the whole post is not only explaining the specific fixes, but the whole investigation procedure that led to them.


`--max-old-space-size=100`
: This limits the heap to 100 megabytes and helps to reproduce the issue much faster.

`--heapsnapshot-near-heap-limit=10`
jdapena marked this conversation as resolved.
Show resolved Hide resolved
: This tells V8 to produce a maximum of 10 snapshots, each time it comes close to running out of memory. This prevents thrashing where the memory-starved program spends a long time producing more snapshots than needed.

`--enable-etw-stack-walking`
: This allows tools such as ETW, WPA & xperf to see the JS stack which has been called in V8. (Node v20+)
jdapena marked this conversation as resolved.
Show resolved Hide resolved

`--interpreted-frames-native-stack`
: This flag is used in combination with tools like ETW, WPA & xperf to see the native stack when profiling. (Node v20+).
jdapena marked this conversation as resolved.
Show resolved Hide resolved

After each snapshot, V8 tries to force garbage collection to reduce memory usage and avoid hitting the limit. In the test case, the memory usage increases, but, after several iterations, garbage collection ultimately could not free up enough space and so the application terminated with an *Out-Of-Memory* error.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Jason took recordings using Windows Performance Analyzer (see below) in order to narrow down the issue. This revealed that most CPU time was being spent within the V8 Heap Explorer. Specifically, around 30 minutes would elapse just walking through the heap to visit each node and collect the name. This didn’t seem to make much sense - why would recording the name of each property take so long?
jdapena marked this conversation as resolved.
Show resolved Hide resolved

This is when I was asked to take a look.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

## Quantifying the Problem

We knew the snapshots were dramatically increasing execution time, so the first step was adding support in V8 to better understand the time used by heap snapshotting. The capture process itself is split into two phases: generation, then serialization.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

We landed [this patch](https://chromium-review.googlesource.com/c/v8/v8/+/4428810) upstream to introduce a new command line flag `--profile_heap_snapshot` to V8, which enables recording of both the generation and serialization times.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Using this flag, we learn some interesting things!
jdapena marked this conversation as resolved.
Show resolved Hide resolved

First, we could observe the exact time the CPU was using for each snapshot. In our reduced test case, the first took 5 minutes, the second took 8 minutes, and each subsequent snapshot kept on taking longer and longer. Nearly all of this time was spent in the generation phase.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

We also identified other widely-used JavaScript applications that demonstrated the same slowdown, in particular, running ESLint on TypeScript. The problem was not app-specific.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Furthermore, we found the problem happened on both Windows and Linux. The problem was also not platform-specific.

## Windows Performance Analyzer to the rescue
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure what this section adds to the article.


As the problem was initially reported on a Windows platform, I used [Windows Performance Toolkit](https://learn.microsoft.com/en-us/windows-hardware/test/wpt/), based on [ETW](https://learn.microsoft.com/en-us/windows-hardware/drivers/devtest/event-tracing-for-windows--etw-), for analysis. This is a powerful low-level expert tool to find out exactly what a program is doing on Windows.

To record the session, I followed these steps:
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This section seems overly detailed to me. The upshot seems to be that there's a Windows-specific profiler that gave you good leads, which I think suffices. The meat of this is what you found out from those leads and how you optimized the code.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

In this case I wanted to be explicit about the procedure used for the investigation, so other people can benefit from it investigating performance problems in V8 (and not only V8) scope.

Even when Windows Performance Toolkit documentation is not bad, the specific steps to investigate this V8 problem are not completely trivial. The actual solutions to the detected problems were not really complex, but identifying exactly where the overhead was has been more significant. And for that, the step-by-step looks valuable to me.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think the upside of this is that this teaches people how to address similar issues and would help future contributors. But one down side is that this is very Windows-specific. Maybe we can save the instructions about the exact buttons to click, because the links to the Microsoft documentation are already more than enough for people who wants to get their hands dirty. And we can focus on what to look (so people who are using e.g. Linux perf can apply it to their tool as well) instead, and perhaps touch on V8/Node.js flags to use for lack of better documentation on them. So maybe something like:

To analyze the performance of heap snapshot generation, I followed [these steps](https://learn.microsoft.com/en-us/windows-hardware/test/wpt/wpr-how-to-topics#start-a-recording) to run the Windows Performance Recorder with the failing script using `NODE_OPTIONS="--max-old-space-size=100 --heapsnapshot-near-heap-limit=10 --profile-heap-snapshot`. I had to modify Node.js to accept `--profile-heap-snapshot` in `NODE_OPTIONS`, as it uses an allowlist to filter V8 flags that can be configured through the environment variable.

To make this more succinct, we can also move the Narrowing the Problem section (or just the explanation about the flags) here - the first step was adding the flag, which gave us a rough picture of the time spent on snapshot generation. In terms of the readers' understanding, ETW does not come into the picture until we start the second step, so we do not have to explain about the flags or introduce ETW/Windows performance toolkit before we finish talking about the first step (the fact that we used ETW before we added the flag was a less relevant background info for readers, we really only put them into good use in the second step) . It also seems that the Narrowing the Problem left a lot of gaps that are not filled until we reach this section, and when we finally reach here we have to again repeat some information mentioned before, so we might as well just merge them here.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it is still interesting to add how to use the specific Windows tooling. It is not trivial to take the right steps with it to achieve a good recording that is informative enough.

Main issue I see with what I wrote here is that, one thing is explaining how to use the tool for this specific problem, and another different one is being a step-by-step manual of how to use WPR. This post should refer to the link you propose for the "step-by-step" guide, and only add hints of the specific configuration used.

I will rewrite this accordingly.

The gaps in "Narrowing the problem" are intentional, as they are explaining the investigation process. We come from scattered facts to getting more information (in this case first with logging, then profiling) that allows to resolve the problem. I would prefer to keep the flow like that.


1. Open [Windows Performance Recorder](https://learn.microsoft.com/en-us/windows-hardware/test/wpt/windows-performance-recorder) and selected CPU profiling, verbose detail level, general scenario, and file logging mode.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

![](/_img/speeding-up-v8-heap-snapshots/d62c099f-6218-4991-a298-c735afe3d6ce.png){.no-darkening}


2. After that, I started the recording session (pressing the Start button).
3. Then, I executed the failing script with `NODE_OPTIONS="--max-old-space-size=100 --heapsnapshot-near-heap-limit=10 --profile-heap-snapshot`. I had to modify Node.js to accept --profile-heap-snapshot, as its command line parameters filter does not allow passing it in `NODE_OPTIONS` right now.
jdapena marked this conversation as resolved.
Show resolved Hide resolved
4. I just let it run a couple of dumps (it would already take over 10 minutes!) and then I stopped the recording.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

## First Optimization: Improved StringsStorage hashing

When I opened the recording with [Windows Performance Analyzer](https://learn.microsoft.com/en-us/windows-hardware/test/wpt/windows-performance-analyzer), what did I find? This:
jdapena marked this conversation as resolved.
Show resolved Hide resolved

![](/_img/speeding-up-v8-heap-snapshots/345360c1-297e-43d2-901f-ea4fa81ede3c.png){.no-darkening}


One third of the samples was spent in `v8::internal::StringsStorage::GetEntry`:

```javascript
jdapena marked this conversation as resolved.
Show resolved Hide resolved
181 base::HashMap::Entry* StringsStorage::GetEntry(const char* str, int len) {
182 uint32_t hash = ComputeStringHash(str, len);
183 return names_.LookupOrInsert(const_cast<char*>(str), hash);
184 }
```

Where exactly? I added the source line numbers to the break down and found this:
jdapena marked this conversation as resolved.
Show resolved Hide resolved

![](/_img/speeding-up-v8-heap-snapshots/59293b17-6d52-4d9e-8737-5b23d038d50a.png){.no-darkening}

So, that inlined `ComputeStringHash` call is causing over 30% of the delay! But why?
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Let’s first talk about `StringsStorage`. Its purpose is to store a unique copy of all the strings that will be used in the heap snapshot. For fast access and handling uniqueness, this class uses a flatmap: a hashmap backed by an array, where collisions are handled by storing elements in the next positions of the array.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

I started to suspect the problem could be too many collisions, which could lead to long searches in the array. But I needed to prove it. So I added exhaustive logs to see the generated hash keys and, on insertion, see how far an entry could end after the expected position for its hash key.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Things were… not right: the offset of many items was over 20 entries. But some cases were really bad, in the order of thousands of positions off from their expected position!
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Part of the problem was caused by the scripts using strings for lots of numbers - and especially a big range of numbers from 1 to several hundreds without gaps. The hash key algorithm had two implementations, one for numbers and another for other strings. While the string hash function was quite classical, the problem came with the numbers implementation that would basically be the value of the number - with a prefix for the number of digits.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Some examples of problems with this hash function:

* Once we had strings with a hash key value in lower positions, then the storing of new numbers would offset all the positions.
jdapena marked this conversation as resolved.
Show resolved Hide resolved
* Or even worse: if we would introduce a string with a hash key value that would be a low number, and we already found several hundreds or thousands of consecutive numbers, then that value would be moved several hundreds or thousands of positions.

What did I do to fix it? As the problem comes mostly from numbers represented as strings that would fall in consecutive positions, I modified the hash function so we would rotate the resulting hash value 2 positions to the left. So, for each number, we would introduce 3 free positions. Why 2? Empirical testing across several work-sets showed this number was the best choice to minimize collisions.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
What did I do to fix it? As the problem comes mostly from numbers represented as strings that would fall in consecutive positions, I modified the hash function so we would rotate the resulting hash value 2 positions to the left. So, for each number, we would introduce 3 free positions. Why 2? Empirical testing across several work-sets showed this number was the best choice to minimize collisions.
What did I do to fix it? As the problem comes mostly from numbers represented as strings that would fall in consecutive positions, I modified the hash function so we would rotate the resulting hash value 2 positions to the left. So each pair of consecutive number would be placed with 3 free positions in between. Why 2? Empirical testing across several work-sets showed this number was the best choice to minimize collisions.

(I took me a while to realize that it was talking about ((n + 1) << 2) - (n << 2) = 3)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you perhaps add code snippets demonstrating the improved hash directly? V8 blog audience should be used to bitwise manipulation code.


[This hashing fix](https://chromium-review.googlesource.com/c/v8/v8/+/4428811) has landed in V8.

## Second Optimization: Caching source positions

After fixing the hashing, we re-profiled and found a further optimization opportunity that would take a significant part of the time. *An old friend*.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

For each allocation in the V8 heap, the heap snapshot tries to record the call stack responsible for the allocation. Therefore, for each stack frame, it needs to know the function name and its source location (filename, line number, column number). It was the fetching of this information that turned out to be super slow!
jdapena marked this conversation as resolved.
Show resolved Hide resolved

What was happening? It was something similar to [what I fixed in the ETW stack walk](https://chromium-review.googlesource.com/c/v8/v8/+/3669698) and that I explained in [this post](https://blogs.igalia.com/dape/2022/12/21/native-call-stack-profiling-3-3-2022-work-in-v8/).
jdapena marked this conversation as resolved.
Show resolved Hide resolved

To compute the source code lines of a function, V8 knows the linear position of the function in the file. But, to find the source line, it needs to traverse the whole file to identify where each newline occurs. This is expensive.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

When requesting line information, V8 already knows how to cache the source line positions per-script. But the heap snapshot implementation was not using a cache!
jdapena marked this conversation as resolved.
Show resolved Hide resolved

The solution? Before generating the heap snapshot, we now iterate over all the scripts in the V8 context to compute and cache the source line positions.

[This source position caching fix](https://chromium-review.googlesource.com/c/v8/v8/+/4538766) has also landed in V8.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

## Did we make it fast?

After enabling both fixes, we re-profiled. Both of our fixes only affect snapshot generation time, so, as expected, snapshot serialization times were unaffected.

When operating on a JS program containing...

* Unoptimized development JS, generation time is **50% faster** 👍
* Optimized production JS, generation time is **90% faster** 😮

Why such a massive difference between optimized and unoptimized code? It is mostly because of the source code line optimization. In code optimized using bundling and minification, there are fewer JS files and these files tend to be large, which means finding the newlines takes longer. And therefore, these larger files benefit most from the source position caching that only performs this work once.

The optimizations were validated on both Windows and Linux target environments.

For the particularly challenging problem originally faced by the Bloomberg engineers, the total end-to-end time to capture a 100MB snapshot was reduced from a painful 10 minutes down to a very pleasant 6 seconds. That is **a 100x win!** 🔥

The optimizations are generic wins that we expect to be widely applicable to anyone performing memory debugging on V8, Node, and Chromium. These wins shipped in V8 11.5.130, which means they are found in Chromium 115.0.5576.0. We look forward to Node gaining these optimizations once it adopts the newer V8 release.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

## What’s next?

First, it would be useful for Node.js to accept the new `--profile-heap-snapshot` flag. Today, it filters V8 command line options, and only allows a known subset.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Information accuracy in snapshots can be improved further. Today, each script source code line information is stored in a representation in the V8 heap itself. And that’s a problem because we want to measure the heap precisely without the observability infrastructure affecting the subject we are observing. Ideally, we would store the cache of line information outside the V8 heap in order to make heap snapshot information more accurate.
jdapena marked this conversation as resolved.
Show resolved Hide resolved

Finally, now that we improved the generation phase, the biggest cost is now the serialization phase. Further analysis may reveal new optimization opportunities in serialization.

## The end
jdapena marked this conversation as resolved.
Show resolved Hide resolved

This work was possible thanks to the sponsorship of [Igalia](https://www.igalia.com) and [Bloomberg](https://techatbloomberg.com).
jdapena marked this conversation as resolved.
Show resolved Hide resolved