[ClusterLabs] Antw: Re: Antw: Re: Resources start serial, not parralel

Ulrich Windl Ulrich.Windl at rz.uni-regensburg.de
Wed Dec 16 14:29:42 UTC 2015


>>> Michal Koutný <mkoutny at suse.com> schrieb am 16.12.2015 um 13:43 in
Nachricht
<56715C56.60409 at suse.com>:
> Hi Oleg.
> 
> On 12/16/2015 11:31 AM, Oleg Ilyin wrote:
>> So, main point of my issue  is  jobs = 1
>> 
>> Please,  does it possibly to increase number of jobs through  throttle
high?
> The parameter you are looking for is 'load_threshold' cluster property.
> It defaults to 0.8, which is IMO quite reasonable. So if you didn't

Hmm,

I'd like to object strongly (as I pointed out in a former message): Linux adds
the load of all CPUs, adn does not average them, AND I/O waits add to the load
(as opposed to HP-UX for example). Thus on a 2*6 Core Intel machin (24 logical
CPUs) a load of 10 means some CPUs are bored. Recently we had a load of 60, and
nobody noticed.

This means such a system is more or less idle:
---
top - 15:27:07 up 4 days, 19:07,  5 users,  load average: 2.45, 2.73, 2.37
Tasks: 1017 total,   4 running, 1013 sleeping,   0 stopped,   0 zombie
Cpu(s): 11.5%us,  2.1%sy,  0.0%ni, 81.1%id,  4.6%wa,  0.0%hi,  0.7%si, 
0.0%st
Mem:    129058M total,   120777M used,     8281M free,     1319M buffers
Swap:    20465M total,      166M used,    20299M free,    93065M cached
---

Maybe the default should be 0.8 times the number of logical CPUs, considering
"hot plug CPUs" that work nicely already (e.g. in Xen PVMs).

> change it, I'd suggest rather looking how the actual load could be
> reduced (pinpoint and optimize the consumers or more performant HW),
> instead of "forcing" Pacemaker to intentionally overload the node.

Dropping useless load and improving efficiency is the best solutions of all
;-)

Regards,
Ulrich





More information about the Users mailing list