Implement IOCountersWithContext for AIX with CGO

[pgimalac]

Implement IOCountersWithContext for AIX with CGO.

Uses the disk info from github.com/power-devops/perfstat (already used by other functions).
https://github.com/power-devops/perfstat/blob/main/types_disk.go

[shirou]

Thanks for the implementation. I checked the perfstat source and have a few questions:

1. Unit of d.Time: The comment says "Time is already in milliseconds", but the perfstat Disk.Time comment only says "amount of time disk is active" without specifying the unit. Have you confirmed that dk_time is in milliseconds from the IBM perfstat_disk_t documentation? It seems inconsistent that Rserv/Wserv are in nanoseconds while Time is in milliseconds.
2. Unit of Rserv / Wserv / WqTime: These are converted as nanoseconds, but could you confirm this from the IBM documentation? They might be in microseconds instead.
3. QDepth: The perfstat.Disk struct has a QDepth field ("instantaneous service queue depth"). Could you map it to IopsInProgress?
   IopsInProgress: uint64(d.QDepth),

[pgimalac]

To be honest the documentation is somewhat lacking so it comes from experimentation...

1. On the host I have I had a value which was only credible when interpreted in milliseconds, that's why I assumed that.
   However since you asked I did some more checking and I now think it's actually in "kernel ticks" (10ms on my host).

I ran a program which writes to a file (without buffering) for 10 seconds, it also gets perfstat info for this disk at the start and end, and the time delta was 467.

I also ran iostat at the same time as the program ran, it showed an average use of 46.6% over the 10 second window.
If 467 was in milliseconds, the average use would be 4.7%, not 47%, so the time unit must be 10ms, which matches kernel ticks. On my host, CLK_TCK is 100, which means the tick is 1s/100 = 10ms.
I tried this a couple times, some runs had a few % off due to timing, but overall it matches.

I updated the PR. See _SC_CLK_TCK in https://www.ibm.com/docs/en/aix/7.2.0?topic=s-sysconf-subroutine.

2. I did a similar check for those fields in case I also got it wrong initially 😅

Two measures with a 10s sleep in between:

xfers=18472325 xrate=1274564
rserv=811487705067 wserv=1548373062628 min_wserv=0 max_wserv=0
wq_time=126732003901643

xfers=18508321 xrate=1274564
rserv=811487705067 wserv=1551492423966 min_wserv=40070 max_wserv=273794
wq_time=126732062753540

The deltas are

xfers=35996 xrate=0
rserv=0 wserv=3119361338
wq_time=58851897

Running iostat during the same window shows wps 3700, avgserv 0.2ms, minserv=0.1ms, maxserv=0.3~0.5ms, and queue avgtime 0.0ms (NB: units are not explicit but from this doc serv is in milliseconds https://docs.oracle.com/cd/E19455-01/805-7229/6j6q8svh7/index.html).

There are 35996 transfers (all writes), so the average time per write is 3119361338/35996=86658, which is only somewhat coherent with iostat (0.2ms) if the unit is nanoseconds.
Similarly min_wserv and max_wserv only match when interpreted as nanoseconds.
And finally wq_time per transfer is 58851897/35996=1635, which only matches 0.0ms if interpreted as nanoseconds.

Again I'm sorry I don't have much stronger arguments than that, the documentation is unfortunately not explicit so running things and comparing values is the best I can do.

3. Yes for sure ! I think that's correct

[shirou]

Thank you for the thorough investigation! The experimental data convincingly confirms:

• Time is in kernel ticks (using sysconf(_SC_CLK_TCK) to convert is the right approach)
• Rserv/Wserv/WqTime are in nanoseconds
• QDepth mapping to IopsInProgress added