Performance - %wio to be considered as available ?

[Ninad_1]

Hi All,

When determining performance and available resources for a system, should the %wio be considered as available or only the %idle be considered as available from the sar -u output.
I am aware that %wio means cpu waiting idle for IO to complete - so should this be considered as idle and available or idle but unavailable.
So when I determine what is the system load - should I consider
1. %usr+%sys as system usage and %wio+%idle = available capacity
OR
2. %usr+%sys+%wio as system usage and %idle= available capacity.

Please provide your reasoning for your suggestion.

Thanks,
Ninad

[Steven E. Protter]

Shalom Ninad,

You are trying to boil down to a number something that may not be quanitifiable.

How much system do you have available? If you have any significant wio, wait i/o your system can have very light cpu use and not be available for any more work.

You can have wait i/o whether the system is busy cpu wise or not.

Seems you are giving me two choices here if I wish to answer.

I would choose to state capacity differently.

I'd use idle to show spare cpu capacity.

I'd show a constant wait io figure or percentage of the system being at i/o capacity.

Wait i/o is something you can tune.

You can spread i/o over more disks and make wio% go away. So I don't think wait i/o percentage is a valid measure of system capacity. You can add i/o use to idle disks on a system and not effect overall capacity.

So its more comples than a formula.

SEP

[Ninad_1]

SEP,

Sorry, but I think I didnt fully understand what you mean. Please correct me if I am wrong -
The %wio CPU time is actually wasted by CPU neither doing any constructive work nor is available for other processes - Is that right ?
I understand that %wio can be tuned by distributing IO on several disks etc etc., I will divulge into it a bit later. First let me understand how to show system usage and available resources. Suppose I have to draw graphs on server CPU usage Vs time , I have collected stats from sar -u , now can you guide me on how I should show the graphs ?
Should I show usage=%usr+%sys
OR
usage=%usr+%sys+%wio [ Even though system is not using the %wio time to process its neither available for processing - so should I show it under usage - rather than idle ]
OR any other way you can suggest - not giving you only 2 choices :)
How to represnt best.

Thanks for your guidance,
Ninad

[Ninad_1]

SEP,

I have sorted out the question of representing the data in graphs.
Now let me ask the question a bit differently.
If I have %usr =60% %sys=10% %wio=30% and %idle=0% , and if there are some runnable processes , will they get CPU time to execute or will not execute - leave aside the priority and scheduling bit.

vmstat idle = sar %idle + sar %wio

So does this mean that the actual idle CPU = %idle + %wio and can be used by processes.

2nd thing - of reducing %wio.
Suppose I have a Oracle database using raw volumes and my sar output is showing %wio in the range 20-40% . Now I want to understand
Suppose there is a sql query which needs to say search through the database according to some conditions - then the process will fetch data from disks , it will get some data which it will process, check if condition satisfies and again proceeds for next records and so on. Now will the time the process has waited to get the data from the disks be determined as %wio ? If I have already distributed the data as far as possible and have low avserv times for the disks (am using XP1024 array LUNs with striping at LVM level), will the large amount of data (if my database size is large say 1-1.5 TB) cause the %wio to be higher rather than distribution of data across disks ?
Then is the only way out is to archive data and reduce database size ? Or still there are any other ways to reduce %wio ?
I guess I am not the only one with large database size, so what are real world experiences with huge databases and %wio and overall performance. Will increasing cache of XP1024 array help in reducing %wio

Please enlighten me.

Thanks,
Ninad

[Chan 007]

Hi Ninad,

CPU usage = %usr+%sys+%wio+%idle
Idle = usr + sys + wio

see below

00:00:00 %usr %sys %wio %idle
00:10:00 5 16 11 67

wio = stands for the CPU has porcessed the request but unable to send back to the application/db (Requester who requested for a process)

e.g you write a program to calculate / find / perf some thing, the CPU performs its activity, but unable to send the result back to the application/log due to some io activity happenning. May you have have a fast CPU & slow controller or your disks are busy doing a heavy application with more priority. This unwanted waiting time is termed as wio (Waiting for i/o). This means that CPU still holds your result in its ALU (Arith Logic Unit), due to which the next request to the ALU is held in q queue. This time is just wasted.

In theory you should not waste more than 10 to 20%, during a bottleneck , I assume 25% is a bottleneck and you should do some thing to elivate the problem.

Hope you got it.

Chan

[Ninad_1]

Chan,

I have a slightly different opinion on this.
If you say the time is wasted and hence shown as %wio then does vmstat report wrongly ?? because vmstat calculations are
%idle = 100 - (%usr + %sys)
whereas sar reports
%idle = 100 - (%usr + %sys + %wio)

So either It is possible for a process to run even if %idle reported by sar is 0% IF %usr + %sys not = 100 (that means there is some component of %wio shown)
OR
vmstat reports wrongly.
or for that matter even glance reports wrongly - because glance also shows CPU Util as %usr + %sys and hence
idle=100 - (%usr + %sys)

But definitely glance must be showing correctly as I trust it very much :)
So then can %usr + %sys + %wio be > 100% if %idle = 0

I am so confused.
Please help.

Regards,
Ninad

[Alzhy]

Peace and Clean Earth Ninad.

ALL of my custom Performance monitoring scripts on all UNIX Dialects deal only with %sys & %user as true indicators of CPU state.

On an extremely busy/CPU-bound system - you will hardly anymore see %wio anyway.

I've used this scheme for so many years now and my stats/system historicals have been used for capacity planning, system sizsing and CPU forecasting activities -- with much success.

[Tim Nelson]

I have found over the years as cpus and storage have increased their speed that %wio in SAR is a extremely outdated.

Compare information using glance and you will see.

There have been a number of discussions in these forums and the metrics in the glance tools are currently the best for isolating and reporting.

%wio includes all IO. disk, streams, etc..

Only use %wio to lead you to reviewing other stats ( not just disk ).

Many 3rd party reporting products still use %usr + %sys + %wio . This again is old, outdated and misleading.