Accuracy of VMSTAT metrics

[Dan_173]

I've been warned (rightly or wrongly) that the output of VMSTAT is inaccurate and that Glance should be used instead. However, I've seen no indication that VMSTAT metrics are unreliable.

We're using 11.0 (64b).

What do you think?

[Bill Hassell · Accepted solution]

vmstat is very accurate in reporting what it finds using claasic interfaces into the kernel, while Glance (via midaemon) has access to much more detailed (and MUCH better documented) metrics within the kernel. The problem is that the underlying mechanisms used in vmstat cannot be changed significantly as they are based on legacy code, while Glance is constantly updated to match kernel features that were totally unknown when vmstat was first written (dozens of processors, threads, ICOD, PRM, etc). So as with all statistics, understanding exactly how the measurement is made essentially validates the numbers.

[Dan_173]

Thanks Bill.

[A. Clay Stephenson]

THe values are quite accurate but of limited value. THe only vmstat value the I pay any significant attention to is the pageout rate. In a nutshell, if that metric is anything other than very low, you are experiencing memory pressure. A small number of pageouts can occur even in boxes with huge amounts of memory is memory-mapped files are in play - that's normal.

[Sridhar Bhaskarla]

Hi Dan,

It is to be generally noted that you have to keep up with reasonably latest patches.

There is a performance rule. "No single performance tool will provide all the data needed to solve the performance issues".

Glance is a multi-purpose tool from Glance+ that can provide good information with ease. Otherwise, you will have to use a combination of iostat, sar, vmstat, top etc., to get a good idea if not a complete picture.

There are two ways of collecting the performance data.

1. kernel counters: You do not need to turn anything on as these counters are continously updated by the kernel. You can access them using /dev/kmem or pstat calls. The default tools like iostat, sar, vmstat are based on them.

2. Kernel instrumentation: KI is based on the event traces generated by kernel through midaemon (part of Glance+ family). It can provide good depth of performance data than that of kernel counters. But the disadvantage which is acceptable is that there is an additional overhead. Per process detailed information is the advantage on the other hand.

-Sri