Subscribe

Wrong QoS counters

Hi,

 

I set up several QoS policies with a limit to INF.

It worked fine for a while (monitored with Harvest).

 

Yesterday I noticed that some counters are providing nonsense values:

 

qos statistics performance show -refresh-display true -rows 20

Policy Group             IOPS      Throughput    Latency
-------------------- -------- --------------- ----------
-total-               3654789    61343.16MB/s    11.55ms
User-Best-Effort      3315730    56823.18MB/s    11.14ms
SPLUNK                 141786      520.79MB/s    29.78ms
_System-Work            51393        6.29MB/s   151.00us
T1RESI                  44773     1990.19MB/s   323.00us
D1ECM                   41502     1875.59MB/s     2.53ms
saelkes3565             36951           0KB/s    19.66ms
W5E                      6972       60.17MB/s     6.94ms
WE5                      3542       10.88MB/s    10.75ms
WQ5                      3169       10.76MB/s    10.86ms
I2I                      3132        9.42MB/s     9.24ms
W4Q                      2124       18.85MB/s     3.35ms
WQ4                      1487        9.15MB/s    15.28ms
T1MODEL                   814        3.11MB/s  1029.00us
_System-Best-Effort       750           0KB/s        0ms
S4D                       361        4.63MB/s   725.00us
T1INSTRA                  250       20.57KB/s   588.00us
MDMT                       18       36.00KB/s     1.73ms
TEO                        13           0KB/s   149.00us
P1ACRMDB                   12       16.00KB/s        0ms
T1WFM                       6       16.00KB/s   777.00us

The whole node is doing 61GB/s? With SATA Smiley Frustrated

saelkes3565 is limited to 1000iops (100% getattr).

SPLUNK is limited to 1000iops (90% getattr).

In Grafana I see most of the time no values for the nonsense counters.

 

We are running 8.3.1P1 here.

 

The cluster is heavy loaded, is this the problem?

 

Marcus

 

 

Re: Wrong QoS counters

Hi @marcusgross

 

I haven't seen this strange behavior from the CLI statistics command before and they should be accurate regardless of cluster load.  Could it be that you have nested QoS policies defined?  So maybe a policy applied at SVM level and then also volume or lun/file? Such a config is not supported and might cause oddness like this.

 

For data not showing up in Grafana, if it is very low IO it could be the latency_io_reqd feature is kicking in.  See here for more on it:

http://community.netapp.com/t5/OnCommand-Storage-Management-Software-Discussions/Harvest-Graphite-quot-spotty-quot-data/td-p/118734

 

 

Sorry I don't have a better answer.  If the problem persists I recommend to open a support case.

 

Cheers,
Chris Madden

Storage Architect, NetApp EMEA (and author of Harvest)

Blog: It all begins with data

 

If this post resolved your issue, please help others by selecting ACCEPT AS SOLUTION or adding a KUDO or both!

Re: Wrong QoS counters

Hi Chris,

 

we don't have nested Qos groups.

 

I open a ticket towards Netapp.

 

Marcus

Re: Wrong QoS counters

Hi,

 

sometimes the QoS counters showing the right values, sometimes not. OCPM works well.

 

I also noticed that there are some spikes on normal Harvest counters:

 

 

 

spikes.png

Re: Wrong QoS counters

Hi @marcusgross

 

According to the Counter Manager system documentation counters must be monotomically increasing, or in other words it must only increase.  It's kind of like the odometer in a car; you check the value, wait a bit, check it again, and calculate the rate of change from the time passed and the change in the odometer.  If the odometer goes backwards, well, that doesn't happen unless you are up to no good.

 

Anyway, back to ONTAP, if you ever check and the rate of change is negative you are then to assume a reset occurred, likely from a rollover of the counter (i.e. it reached the max size of the data type) or a reset (like a system reboot).  In this case you drop the negative sample and on the next one you can compute your change again.

 

When I see the massive numbers like in your screenshot it appears if the values went down temporarily, so something like this:

 

Time:                T1           T2                   T3                                  T4

NFS OPS:  122400, 123400,             100,                           123600

Calc'd:             N/A   ,   1000,        -123300 (discard),         123500

 

I've seen it before sporadically at customer sites but haven't had enough to open a bug.  If you run Harvest with the -v flag it will record all the raw data received and we can verify this behavior.  Next to figure out is what system event caused it.  Did anything happen at those timestamps?  SnapMirror updates maybe?  Cloning?

 

 

 

OPM uses archive files from the system which is a different collection method.  It also uses presets which are less granular.  Since this is a timing issue I could imagine that those differences somehow avoid the problem.

 

Cheers,
Chris Madden

Storage Architect, NetApp EMEA (and author of Harvest)

Blog: It all begins with data

 

If this post resolved your issue, please help others by selecting ACCEPT AS SOLUTION or adding a KUDO or both!