Ghost-based benchmark have worse performance on node v20.3.0/v20.3.1

[Septa2112]

Version

v20.3.0/v20.3.1

Platform

Intel ICX Linux 5.15.0-72-generic
Intel SPR Linux 5.15.0-72-generic

Subsystem

No response

What steps will reproduce the bug?

Benchmark - An internal benchmark based on Ghost.js
Test version of node - v20.2.0 and v20.3.0/v20.3.1
OS - Linux
Test Platform - Intel ICX and SPR

We used the above two versions of node to test our benchmark.
And found that the performance of node v20.3.0 was not as good as that of v20.2.0.

How often does it reproduce? Is there a required condition?

No response

What is the expected behavior? Why is that the expected behavior?

No response

What do you see instead?

This value represents the RPS. It can be found that the version of node-v20.3.0 is nearly 40% lower than that of node-v20.2.0.

And after a rollback test, we found that this performance impact is related to the upgrade of libuv to 1.45.0, and the issue can be located: #48078

Additional information

Any thoughts on this issue?

[bnoordhuis]

What do those benchmarks test? Can you post profiles or flamegraphs of v20.2.0 and v20.3.0?

Even better would be to test the commits between libuv v1.44.2 and v1.45.0 (there are about 100) to pinpoint the exact commit.

If your benchmarks are amenable to something that works with git bisect, you can use this oneliner to land the libuv commits in this repo:

$ git checkout v20.2.0

$ for COMMIT in `cd ../libuv && git log --oneline v1.44.2..v1.45.0 | awk '{print $1}'`; do (cd ../libuv && git archive --prefix=deps/uv/ --format=tar $COMMIT) | tar x; git add deps/uv; git commit deps/uv -m "libuv: upgrade to $COMMIT"; done

Assumes libuv is in a directory adjacent to node.

[Septa2112]

Benchmark

About the benchmark, I create a ghost.js server based on node in a docker container, and use ab to send requests to test the throughput of the server.

Flamegraph

By comparing these two versions, it can be clearly found that the flame graph of node v20.3.0 has an additional io_sq_thread part, accounting for about 30%.

Rollback Test

I switch to v20.3.0 branch. I found that the decrease of throughput is related to this commit 851508b. The commit is about upgrade of libuv.

With testing of 1 instance, I found that the throughput after this commit was about 20% lower than before.

[bnoordhuis]

Can you test outside of docker? There's a known docker issue that will be fixed worked around in the next libuv release.

Replace ab with (for example) wrk if at all possible. ab is notoriously unreliable.

I found that the decrease of throughput is related to this commit 851508b.

How do you know for sure? Did you bisect or are you guessing?

[Septa2112]

Thanks, I'm following your suggestion to retest my benchmark. The results are expected soon and I will update here.

By the way, why is ab unreliable?

[bnoordhuis]

For one, it's too easy to misuse: it gets confused when the server sends back responses with different sizes, unless you pass -l (and that flag was only added in 2.4, before that it reported such responses as errors!)

It has plenty of other bugs and quirks, even though it's over 2 decades old.

[Septa2112]

Thank you for the explanation about ab.
I tested both inside and outside the docker container with wrk.

Inside the docker container

I just run 1 instance.

The test results of wrk and h2load both show that the performance of node-v20.3.0 is about 5% lower than that of node-v20.2.0.
Although not as drastic as ab's test results, the performance drop is still there.

Outside the docker container

Whether it is multi-instance or single-instance, the test results of wrk show that the performance of the two versions is similar.

Results

Maybe you are right, there may indeed be a bug related to docker. And it seems that ab is really not reliable.

I saw libuv has been upgraded to 1.46.0 and fixed a docker related bug. In addition, I saw that the latest version of node (v20.4.0) also upgraded libuv to 1.46.0.

I'll be retesting our benchmarks on node-v20.4.0 to see if that fixes the slowdown.

Additional

I found through top that the CPU utilization of node-v20.3.0 is nearly 80% higher than that of node-v20.2.0 when fully loaded. Is that OK? Maybe this is caused by sq_thread of libuv-1.45.0?

How I Tested

About the question

How do you know for sure? Did you bisect or are you guessing?

I rolled back the node-v20.3.0 commit in the container. When rolling back to this commit of upgrade to libuv 1.45.0, I found that the performance changed dramatically before and after applying this commit. So I'm sure that the upgrade of libuv had a negative impact on our benchmark.

[bnoordhuis]

Thanks for testing, that's very helpful. I'm also very interested in the results for v20.4.0.

I found through top that the CPU utilization of node-v20.3.0 is nearly 80% higher than that of node-v20.2.0 when fully loaded. Is that OK? Maybe this is caused by sq_thread of libuv-1.45.0?

That seems likely but can you check with perf record where the process spends its CPU time?

If that does indeed show it's the sqpoll thread, is it an option for you to test with a v6.x kernel?

[Septa2112]

Performance of Node v20.4.0

On our benchmark, the performance of node-v20.4.0 is similar to that of node-20.3.0, and both are much worse than node-v20.2.0. libuv upgrade to 1.46.0 doesn't seem to work. Maybe there is another bug?

Perf Data of v20.2.0 and v20.3.0

By the data of perf, we can find that in the 20.3.0 version, io_sq_thread accounts for about 30%.

Node V20.2.0

Node V20.3.0

[bnoordhuis]

Can you test with a newer kernel? I think you're looking at an io_uring performance bug in the 5.x kernels that's been fixed in newer kernels.

Another thing you may want to check is if the profile looks different with perf record -e cpu-clock. You can also split by kernel and user side CPU time with -e k:cycles and -e u:cycles (also works with cpu-clock, task-clock, etc.)

[Septa2112]

I have updated the kernel to v6.4.3 and is retesting all data. It will take a day or two. I will update my data as soon as possible.

[Septa2112]

After updating kernel to v6.4.3, here are the newest results.

Worse Performance on benchmark

Inside the docker container - Results of wrk

Here I only tested 1 instance. And here is the result I manually tested by running server and wrk in docker.

The baseline is node-v20.2.0 (100%) and the performance is

Version	%
v20.2.0	100%
v20.3.0	86%
v20.4.0	87%

Outside the docker container

All results were tested and collected by wrk.

1 instance

The performance of v20.2.0, v20.3.0 and v20.4.0 is about the same. Everything looks normal

full instance (The number is 128 on our ICX server)

I keep the CPU close to full load, and start an additional instance at this time.
I select the v20.2.0 as the baseline(100%). The performance is

Version	%
v20.2.0	100%
v20.3.0	82%
v20.4.0	78%

Abnormal CPU load

Results of Benchmark

I run 1 instance outside the docker container. The CPU utilization of node v20.4.0 and node v20.3.0 is similar. They are still about 80% higher than node v20.2.0 on kernel v6.4.3. But the performance of the three versions is very close.

Results of perf

I will update the results of perf later.

[bnoordhuis]

Thanks for the update. For my understanding: what does "full instance" (vis-a-vis "1 instance") mean?