test: test-buffer-creation-regression intermittent failure on smartos

[mhdawson]

• Version: master
• Platform: smartos
• Subsystem: buffer

Failure on smartos15-64 during regression run for change not related to this test:

https://ci.nodejs.org/job/node-test-commit-smartos/5631/nodes=smartos15-64/console

not ok 96 parallel/test-buffer-creation-regression
  ---
  duration_ms: 61.210
  severity: fail
  stack: |-
    timeout
  ...

[mhdawson]

@nodejs/platform-smartos

[cjihrig]

Related: #10161

[Trott]

Fixed in #10161.

[mhdawson]

Still failed in sequential run here: https://ci.nodejs.org/job/node-test-commit-smartos/5739/nodes=smartos15-64/console

not ok 1271 sequential/test-buffer-creation-regression
  ---
  duration_ms: 369.121
  severity: fail
  stack: |-
    timeout
  ...

[bnoordhuis]

Suggestion: mark the test as flaky on smartos and move on? I suspect this is caused by the infamously terrible virtual memory management that smartos/illumos inherited from solaris. It's not the first time we've hit this.

[bcantrill]

Counter-suggestion: don't debug by superstition or prejudice. Could we please get a little more detail on what it is you're talking about? If there's a SmartOS or illumos issue here, we'd like to understand and fix it. And if you are unable or unwilling to debug it please let us know, and we will find someone to do it.

[misterdjules]

In the description of the second failure, duration_ms: 369.121 means that it took 369 seconds (not milliseconds) for tools/test.py to notice that this specific test did not complete.

However, the default timeout is 60 seconds, and tools/test.py actively polls its child processes' status every at least as frequently as 1/10th of a second.

It seems to suggest that the system was heavily loaded and that the test harness process couldn't be scheduled as frequently as usual, thus making this failure possibly caused by the environment more than by the test itself or a bug in node on SmartOS.

Given that this test doesn't seem to fail that frequently, and that I haven't been able to reproduce that failure manually, I would suggest the following way forward:

1. Pass --abort-on-timeout to tools/test.py when running tests on SmartOS as described in Enable --abort-on-timeout for node-test-commit-smartos Jenkins job build#613.

2. Continue adding instrumentation features to the test harness so that we can get processes' and system stats when a test times out.

3. Close this issue without marking this test as flaky, and reopen it when the next failure happens. The reason for not marking this test as flaky is that we currently don't know if the test itself is actually flaky, and that it is possible that failures were due to the environment. Marking it as flaky would hide future actual failures. Hopefully, when the next failure occurs, we'll be able to use the instrumentation described in 1) and 2) to have a better idea of what the problem is.

How does that sound?

[Trott]

SGTM

[misterdjules]

thus making this failure possibly caused by the environment more than by the test itself or a bug in node on SmartOS.

After further testing, it seems that the performance of this test on SmartOS could be related to the version of the platform on which the VM running the test is provisioned.

To make a long story short, after looking at various kstats (notably the n_pf_throttle_usec in the memory_cap module that represents the time spent by the process being throttled when page faulting) and microstate counters with prstat -m, I couldn't come to a definite conclusion about what's going on.

However, I did provision two VMs using the same image as the current test VMs used to run test Jenkins jobs, using a 8GBs RAM package. One was provisioned on a server using a 6 months old platform, the other on a server using a more recent platform. The VM provisioned on the more recent platform was able to run this test in ~1 second consistently, whereas the other was running the same test in ~10 seconds consistently, with some peaks ~50 seconds.

So instead of moving forward with what I suggested previously, I'll investigate further and update this issue as soon as I have more details.

[thefourtheye]

If needed, we can reduce the size of that test

const size = 8589934592;

to

const size = 4294968296;

I'll submit a PR soon.