Linux Pressure Stall Information (PSI) by Example

[tyingq]

Wow. This would have totally changed my job of being a Unix sysadmin in the 90's.

Funny how such a basic capability doesn't appear as a batteries included thing until 40+ years after Unix appears. Lots of futzing around with uptime, ps, top, iostat, vmstat, ntop, sar, lsof, free, etc. All of that replaced with a simpler tool to get the high level answer.

That first level filter is the key to narrow in your troubleshooting. Bravo. I bet Adrian Cockroft (Sun, Netflix, now AMZN) approves.

[akira2501]

> doesn't appear as a batteries included thing until 40+ years after Unix appears

64kB of RAM in 1980: $405. $1,335 if you adjust for inflation.

64GB of RAM in 2020: $175.

In my mind, that has the most to do with it.

[heavenlyblue]

>> Funny how such a basic capability doesn't appear as a batteries included thing until 40+ years after Unix appears.

Is there anyone on HN who knows why it didn’t appear earlier?

Lack of need? This doesn’t seem like a hard thing to implement - am I missing something in terms of implementation complexity?

[navinsylvester]

Ubercool.

Wasn't aware of PSI - thanks to the author. The load avg alerts are mostly late to our liking but tooling around this should really help.

[the8472]

Another advantage of PSI is that there also are per-cgroup monitors. Unlike load indicators which are system-global and thus pointless when your cgroup is limited by quotas.

[danw1979]

This is going to cone in very useful indeed.

It's been around since kernel 4.2 according to the article. Do any of the mainstream container orchestration tools use this data, I wonder...

[kees99]

In-kernel resource-specific counters are great, but those are limited to 10/60/300 seconds averages.

For historic resource utilization, sar/sadc tool is still a go-to.

[shuss]

kees99, you should check the "total" fields. They have information in microsecond resolution.

[kccqzy]

If I have a cron job and reads the pressure file and simply stores its contents with a time stamp for later analysis, would that be enough to determine historic resource utilization?

[magoon]

Major improvement over loadavg for troubleshooting, given the difficulty in identifying when a system is io-bound.

[rwha]

This is an excellent write-up. The code example for polling events is going to keep me entertained long enough to get in trouble. Well done shuss!

[markandrewj]

This is pretty useful and cool.