Repository navigation
Node CPU Profiling Roadmap #148
Description
We’d like to get feedback on the status of the profiling tools landscape in Node.js today. In particular -- we want to get alignment on a roadmap which will provide a free, open source, and cross-platform set of tools that are part of the node/v8 API i.e. maintained across LTS versions that can provide a powerful suite to debug and diagnose Node.js issues in production.
Production Challenges
There are some challenges that are unique to debugging and diagnosing issues in production. Specifically for large critical production deployments. In particular here are some of the constraints due to the production nature of the deployments:
- Generally, most solutions need to be low impact on the process. Attaching debuggers or tracers is often impractical when the application is taking production traffic. Debuggers will pause the process, causing all inflight requests to hang. Tracers and debuggers often introduce large performance penalties, which can impact customers.
- Tools need to be safe and not cause kernel panics or process crashes.
- It’s often hard to reproduce issues seen at production scale -- therefore programmatic access of tooling via APIs can help diagnose problems as they occur. e.g. start a profile based on some condition since it’s often impossible to reproduce
Profiling
One of the most useful methodologies to optimize CPU performance in a running application is by sampling the CPU stack frames (CPU profiling) and then visualizing the samples, typically using a flamegraph. This technique will show hot code paths on CPU -- which gives you the opportunity to optimize the relevant source code.
The approach can be done in production with OS level profilers such as (perf, DTrace, systemtap, eBPF) with very low overhead. The profilers lack the information to resolve the JS frames, resulting in unhelpful memory addresses for the JS frames. V8 solves this problem by dumping a mapping of native frame addresses to JS source and line number.
It’s important to mention here that having access to all stack frames, whether native (v8, libc, syscalls, libuv, native modules) or JS is important. Problems can occur anywhere in the stack, and we want to be able to profile Node with complete stack frames. E.g. We heavily use gRPC -- which is a native module -- so without access to native frames we would not be able to get visibility into this critical part of our application.
There are a few issues with this implementation:
perf(1)support is now deprecated in V8 and will not be supported starting in Node 8 -- which effectively means we’re losing the ability to profile JS stacks.- Lack of cross-platform tooling. Even if perf support wasn’t deprecated, this solution only works on Linux.
We’d like to contribute and collaborate on a set of comprehensive, cross-platform, and open source CPU profiling tools with the Node and V8 team. The V8 team has advised us that they plan to support the v8 profiler and the v8 cpu profiler API going forward, and we want to unlock CPU profiling capabilities in Node using these supported frameworks.
Roadmap:
- Determine which tools -- v8 profiler (--prof), cpu profiler or something else -- to support and maintain for profiling Node applications.
- Currently, these tools do not support native c/c++ frames which includes v8, native modules, libuv or syscall stack frames. Note because these tools are in process, they will never be able to show syscalls as those can only be sampled from within the kernel. Supporting native frames will require implementing a stackwalker in the v8 profiler or the cpu profiler.
- Determine the overhead of profiling using these tools. With the now deprecated OS level profilers, the overhead was quite low -- and thus suitable to use in a production environment without outsized impact to customers.
- Dynamically toggle profiling in a production environment. It’s not practical to profile all the time in production due to the performance impact -- being able to dynamically enable and disable profiling via an API would be ideal.
- Provide an API to programmatically consume profiling data once enabled.
- Add documentation so these tools can be adopted by users.
We’re looking for feedback and alignment with the community on this subject before proceeding with the design and implementation -- please let us know your thoughts.
Activity
mmarchini commentedon Feb 9, 2018
Great explanation and summary of the current scenario.
Suggestion: We could add "Write tests to those profiling tools" to the roadmap, as I think one of the goals is to have these tools as first-class citizens in the future.
joyeecheung commentedon Feb 9, 2018
Another suggestion: a better standardized format for the profiles, and tools that convert legacy formats to it. The current v8 CPU profile format is quite limiting considering it's JSON so it cannot be concatenated easily and has to be in the memory as a whole before being serialized.
davidmarkclements commentedon Feb 9, 2018
In 0x (flamegraph tool) I've re-implemented stack tracing on top of v8 profiler (
--prof) and the v8 tick processor (--prof-process --preprocess) – see davidmarkclements/0x#99 - it's currently behind a flag but next major version this will be default with a--kernel-tracingflag for OS level tracing.Pros
Cons
TimeTicks::HighResolutionNowwhich is an internal v8 function which has different output depending on system. In a particular configuration of OS X for instance, the TimeTicks::HighResolutionNow output is based on amount of ticks (it just adds one for each tick) – this makes it incredibly difficult to reliably cross reference stacks capture by--profto stacks captured by anything else. Not only that but how do you match the sample rate so that you get one-to-one on the samples?What I would love is the ability to cross reference prof stacks with OS level tracing stacks. If we can do that, then perf being deprecated, and bytecode handler problem (and any other problems) would be resolved with --prof while lower level tracing could be used to compliment --prof.
davidmarkclements commentedon Feb 9, 2018
In terms of Node internal implementation, absolutely agree on tests.
The way we currently expose
--prof-processisn't ideal, the stacks processor is concatenated JavaScript based on some JS files in deps/v8/tools - .. with additional injected code to make log output work with Node (see nodejs/node#14966)danielkhan commentedon Feb 9, 2018
We recently added production grade continuous sampling to our product. For that we had to find a sweet spot between sampling frequency and duration. Additionally there was a memory leak in v8::CpuProfiler (nodejs/node#14894) we discovered.
Here is a blog post with screen shots https://www.dynatrace.com/news/blog/introducing-code-level-visibility-for-node-js/
As already mentioned, this approach won't give you native frames - still we considered it to be sufficient for many use cases.
@Hollerberg who implemented it will be at the summit and we are happy to discuss the validity of our approach and be part of an initiative to improve CPU sampling capabilities in node.
aalexand commentedon Feb 9, 2018
We use the CPU profiler in the profiling agent here that is an experimental way to continuously collect the data from production Node.js apps. Some notes on that experience so far:
mhdawson commentedon Feb 9, 2018
I believe it is important that Node.js have first-class tooling for problem investigation and CPU profiling is part of that along with tracing, debugging, heap dumps, node-report and being able to introspect core files. Ideally, a base level of core tooling would be available with the runtime without having to install additional components as that can often be an issue in production.
That's a long way of saying that I support your effort and will look to see how I can help in respect to getting ci testing and other supporting pieces in place and also how I might be able to pull in people from IBM to help collaborate.
mmarchini commentedon Feb 13, 2018
For reference: #150
hashseed commentedon Feb 16, 2018
Another issue that @bmeurer mentioned to me is that there is no way for Linux perf and any other perf tool outside of V8 to discern inlined functions from its caller, since both occupy a single stack frame. The CpuProfiler doesn't provide this either, but @FranziH is addressing this.
mmarchini commentedon Feb 28, 2018
Updates from the Diagnostics Summit
External profilers
External profilers are working well on Node 6, even though they are not officially supported. They are also working on Node 8+, but the information collected by them can be misleading due to the introduction of Turbofan and Ignition. Also, the current method used by most external profilers to resolve JITed functions (
--perf-basic-prof) is not officially supported on V8 and might break/be removed in the future.To support external profilers in the future, we need two things:
Interpreted Frames
After the introduction of Turbofan and Ignition, Interpreted Frames on the stack don’t reflect JavaScript function calls since only the interpreter appears in the stack, as we can see in the image below.
As a consequence, understanding the data collected by external profilers can be tricky, and the data can be misleading since there's no way to distinguish between different JavaScript function calls when they are running in interpreted mode. As soon as those functions are compiled and optimized by Turbofan, they will appear in the stack as before.
During the Summit, we came up with three different approaches to have more meaningful information on the call stack for external profilers. All of them must be implemented on the V8 interpreter, and they basically change the current flow to add a unique stack frame for each JS function.
Intermediate function before calling the interpreter
Add an intermediate frame to the stack which points to a JIT-function with the purpose of keeping track of which JS function is being called.
Duplicate the interpreter code for each JS function
Copy the InterpreterEntryTrampoline code for each Interpreted Function, this way each Interpreted Frame will have a unique entry point. Apparently, ChakraCore is implemented this way.
Change the stack on runtime to replace the interpreter with a unique address representing the JS function
Hack the call stack at runtime, replacing InterpreterEntryTrampoline's return address with the address to a small JIT function which will return to InterpreterEntryTrampoline later.
API with information to resolve JIT function addresses
External profilers can’t resolve names for JITed functions (including V8 Builtins) without help from the runtime. Also, most of the time those names are resolved with post-processing tools. Today we have
--perf-basic-prof, which is used by a variety of tools to post-process the output from profilers and extract useful information from that.As
--perf-basic-profis very limited and is not officially supported by V8, the suggestion is to create an API which listens to code creation events and expose all information needed by external tools to resolve names for JITed functions.V8 Profiler & CpuProfiler
V8 builtin profilers are officially supported, but they can only sample JavaScript frames. We discussed the possibility to add native frames to those profilers as well in the future. Some challenges are 1) sampling frames from other threads; 2) sampling syscalls frames.
Roadmap
mmarchini commentedon Feb 28, 2018
Related issue: #150
mhdawson commentedon Mar 1, 2018
Thanks for the great write up of the discussion from the summit.
ofrobots commentedon Mar 2, 2018
OP said:
AFAIK,
perf(1)support is not deprecated. Like any of the other profiling APIs, it doesn't come with an official support expectation from the V8. I do not think this perf support is going anywhere anytime soon.59 remaining items
Flarna commentedon Apr 5, 2020
The problem I reported in #23070 was based on a real crash with a real application from a customer. That application had ~230000 functions. I increased it to 1M in my reproducer to get a similar memory pattern as the real application which requires a few hundred MB already without Profiler.
In that application quite some functions were GCed after startup initialization of the app was done.
That time the memory overhead was the main problem which looks far better now. But it still seems that the amount of work to be done by v8 main thread depends on the number of functions and it is synchron. Most likely because functions coluld move in memory otherwise.
I agree that such a lot functions is not the typical setup. The main issue for us was that that there is no mechanism to estimate the overhead nor to limit it therefore we stopped using the profiler.
Our use case is to continously profile with low interval to get a long term pattern not to profile/optimize something or analyse an app in a bad state.
mmarchini commentedon Apr 6, 2020
Those are valid use cases, but most of the improvements as I mentioned above should be implemented on V8. So I guess we need to start to discuss how to get this upstream.
mmarchini commentedon Apr 6, 2020
Added to the agenda so we can discuss how to proceed.
mmarchini commentedon Jun 17, 2020
Removed from the agenda until someone has time to work on it (reach out to V8 folks, gather data, etc.)
github-actions commentedon Sep 16, 2020
This issue is stale because it has been open many days with no activity. It will be closed soon unless the stale label is removed or a comment is made.
RafaelGSS commentedon Sep 17, 2020
I would like to discuss it and help. Re-adding to the agenda
RafaelGSS commentedon Nov 11, 2020
Hi folks.
As previously discussed on WG Meetings, since it issue was created in 2018 a lot of work was been done and a lot of contexts were mentioned here.
Seems fair to close that issue and create another one about the performance of
Profiler.start()andProfiler.end().We are working also on expose the Deep Dive documents through the
documentations/folder. You can see it here.RafaelGSS commentedon Nov 19, 2020
Closing in favor of #444