linux-kernel - Re: [Question] load balance move tasks not suitable ?

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 30 Jul 2018 21:42:27 +0800
From:   "weiqi (C)" <weiqi4@...wei.com>
To:     Peter Zijlstra <peterz@...radead.org>
CC:     <linux-kernel@...r.kernel.org>,
        Zhanghailiang <zhang.zhanghailiang@...wei.com>
Subject: Re: [Question] load balance move tasks not suitable ?

in latest  linux kernel version,  load-balance's  detach_tasks() keeps  
"load/2"  behavior:

"""

         if (( load / 2 ) > env->imbalance)
                 goto next;
"""

It  means  a task  "p",  which  load  is 1.5 times than "env->imbalance" 
can move from "src_rq" to "dst_rq".
after this move,  imbalance between  "src_rq"  and "dst_rq" may larger 
than  before.
but if use "load*2" here,  It means we can pick a task "p" , which load 
is half of "env->imbalance"， and after move，
"src_rq" and "dst_rq"  is balanced.

Am I right?


在 2018/7/30 20:56, Peter Zijlstra 写道:
> On Mon, Jul 30, 2018 at 08:08:55PM +0800, weiqi (C) wrote:
>> Hi Peter,
>>
>> After  Commits 5d6523ebd(sched: Fix load-balance wreckage)，
> That's a _6_ year old patch... and I can barely remember last week.
>
>> The jugement whether imbalance reached changed from twice to one-half.
>>
>> from
>>
>> (1) if((load * 2)> rem_load_move)
>>          goto next;
>>
>> to
>>
>> (2) if((load / 2)> env-> load_move)
>>          goto next;
>>
>> I'm confused about this change.
>>
>> "load*2" may be more appropriate, because if a task whose load more than
>> env->imbalance is moved from high load cpu to low load cpu,
>>
>> will make more imbalance.
> It basically goes back to what it was before (see 367456c), in that
> patch I (inadvertently) replaced (load.weight >> 1) with (load * 2)
> instead of (load / 2).
>
> I'm not entirely sure I can explain that logic at this time. Doing
> software archeology on it might clarify where it came from.
> .
>