> On Dec 21, 2020, at 9:53 PM, changhuaixin <changhuai...@linux.alibaba.com>
> wrote:
>
>
>
>> On Dec 17, 2020, at 9:36 PM, Peter Zijlstra <pet...@infradead.org> wrote:
>>
>> On Thu, Dec 17, 2020 at 03:46:17PM +0800, Huaixin Chang wrote:
>>> In this patch, we introduce the notion of CFS bandwidth burst. Unused
>>> "quota" from pervious "periods" might be accumulated and used in the
>>> following "periods". The maximum amount of accumulated bandwidth is
>>> bounded by "burst". And the maximun amount of CPU a group can consume in
>>> a given period is "buffer" which is equivalent to "quota" + "burst in
>>> case that this group has done enough accumulation.
>>
>> Oh man, Juri, wasn't there a paper about statistical bandwidth
>> accounting somewhere? Where, if you replace every utilization by a
>> statistical variable, the end result is still useful?
>>
>> That is, instead of something like; \Sum u_i <= 1, you get something
>> like: \Sum {avg(u),var(u)}_i <= {1, sqrt(\Sum var_i^2)} and you can
>> still proof bounded tardiness etc.. (assuming a gaussian distribution).
>>
>> The proposed seems close to that, but not quite, and I'm afraid it's not
>> quite strong enough to still provide any guarantees.
>
> After reading some papers on statistical bandwidth sharing, it occurs to me
> that statistical bandwidth sharing is about the way bandwidth is shared
> between competitors. I wonder if the paper you mentioned was "Insensitivity
> results in statistical bandwidth sharing" or some paper referenced, which
> showed the end result is insensitive to maybe the distribution or the arrival
> pattern.
>
> I am sorry that I cannot prove using statistical bandwidth sharing theory
> now. However, I wonder if it is more acceptable to put rate-based cfsb after
> share fairness, because the input stream of cfsb account is the output stream
> of fairness. And that is in contrast with what statistical bandwidth sharing
> does, in which fairness is used to share bandwidth upon the output stream of
> rate-based control. In this way, maybe guarantees can be provided by share
> fairness, and cfsb may use a larger buffer to allow more jitters.
>
> A buffer, which is several times of quota, is able to handle large bursts,
> and throttle threads soon when overloaded. The present cfsb, however, has to
> be configured several times larger to handle jitters, which is ineffective
> and does not provide the designed rate-based control.
Hi, peter.
Will you please have a look at my earlier reply.
The guarantee provided by cfsb is max-min fairness with strict upper bound, I
think. And after this modification, it is still roughly that. As for longer
periods, each cfsb group is still limited to their quota on average.
When users configure a huge burst buffer, cfsb doesn't really throttle threads
and upper bound is exceeded. In that case, guarantees are provided by share,
which provides proportional fairness.
Thanks,
Huaixin
