max load

[Hr. RJH Verhagen]

Hello,
Doe anybody knows what the maximal load for a hp-unix machine is? I made a tool witch reports me when the load is higher than 4 (average of 15 minutes). But this is only a guess. I do not know what the real maximum is. I have a (production) machine who always has a load around the 3, is this to much?
Thanks,
Robert Verhagen.

[Mark Greene_1]

If you are using "uptime" for your load number, I've seen systems in excess of 20. You are better off monitoring vmstat or glance as you will get more meaningful data.

HTH
mark

[Martin Johnson]

Max load is a relative thing based on what is acceptable response time. I have had loads in the 30+ range on batch systems. As long as the batch job finished within the batch window, there were no problems. Those type of loads would be unacceptable to interactive users.

in my experience, loads of 3-4 have not been a problem, but again it depends on what is acceptable response time. If your users are not complaining, then things are probably OK.

HTH
Marty

[Stefan Farrelly]

i dont know about maximum but recommended is 1 per cpu, so on a 4 cpu system a load of 4 is about the max before your performance will degrade.

However, when an HP servers develops a hardware or i/o problem the load average (from uptime) can skyrocket into the hundreds! ive seen them over 200 and yet users can still work. You will need a reboot to resolve once youve fixed the hardware or i/o problem.

[Brian Watkins]

What tool/utility is giving you that load value?

Typically, any load value of 3 or higher needs to be looked into, as you could be experiencing some kind of bottlnecking (memory, LAN, I/O, etc.)

If you have glance installed, I would suggest using it. If you don't have it installed, you can install an evaluation copy from you Application CDs.

The glance tool is much more robust than top or sar, as it has hooks directly into the kernel, which will give you a much more accurate "real time" (is there such a thing? 8^) ) reading of your system.

After using glance to analyze your system, you should be able to get a better idea of where your current/potential problems lie.

[Dave Chamberlin]

There isn't really a hard number for maximum load. It is a useful number to make relative comparisons on a given machine at different times. Your maximum load, I would think, is the point at which system resources (memory, CPU, IO bandwidth) are completely consumed. Using Glance or GlancePlus can show which of your resources are in the shortest supply. If you have problems they will show up there. If you still want a hard number, monitor the system and note what the load is when users start complaining. Lower that value by 1 or more and call that your max load.

[Tim D Fulford]

Robert

the load is the "number of running or runable processes in xxx minutes" as seen from say top.

there is not actual MAX load as you could have a load average of 100, which MIGHT mean 100 runable processes, but they are blocked on IO.

Personally I use MeasureWare to diagnose if you are CPU bound. The metrics to look at are
GBL_PRI_QUEUE & GBL_RUN_QUEUE. If both of these are high (the manual suggests 3) AND GBL_CPU_TOTAL_UTIL is high (>80%) then you can be fairly confident that you are CPU bound.

Also bear in mind that this situation can also occur when processes spawn new processes very quickly. So you might not be required to buy new CPU's just alter the way your apps work. It is worth looking at what you DB/application guys think before buying more CPUs.

Just my thoughts

Tim

[Sridhar Bhaskarla]

Hi Robert,

The maximum load usually defined is the number of CPUs on the system.

However, it is true if the following condition is true.

Do a sar 2 20 and observe the output. %usr should usually be far higher than %sys and %wio. If %sys and %wio look abnormal, then you mostly have a bottleneck on the system and the above max load formula is not valid.

-Sri

[Bill Hassell]

The load average reported by uptime and top is the average length of the run queue. With completely compute-bound programs, it indicates the total number of processes that are ready to run. In a 1-cpu system, a load of 1 means the CPU is fully utilized, in a 4-cpu system, the load can be 4.

But Unix is a multitasking operating system and the number of changes of programs during one second can be staggering, dozens to hundreds. So if you had 500 programs running and each program simply asked for a status over the LAN and got a short response, and these program ran as fast as possbile, you might have a run queue of 200 to 400, yet all is well, people can log in and read email, etc.

On the other hand, runaway programs could loop around asking the operating system to perform useless tasks that require enormous system overhead both on disk and LAN (tasks that require a spinlock on multi-cpu systems are the worst) and the system might have a run queue of 2x the number of CPUs and it will almost impossible to login.

Therefore, the run queue ("load" from top and uptime) is by itself, a meaningless value. It should only be used in conjunction with other metrics.

[MANOJ SRIVASTAVA]

Hi Robert


Also at higher loads like 12 and above the system utilities like mail , predective stop wroking , we have a system with 32 CPUs and I ahve seen it to go about 22 load factor thought that is very abnormal.


Manoj Srivastava