[squid-dev] [RFC] CRUSH peer selection method

[Alex Rousskov]

On 04/19/2017 08:06 AM, Loic Dachary wrote:
> On 04/19/2017 03:53 PM, Alex Rousskov wrote:

>> On 04/18/2017 01:00 AM, Loic Dachary wrote:
>>
>>> It turns out [CARP] performs as well as CRUSH


>> On 04/19/2017 06:51 AM, Loic Dachary wrote:
>>
>>> I found one use case where CRUSH behaves significantly better than CARP.


>> FYI, here is how the above statements can be interpreted: "It is all a
>> mystery to us. Sometimes CRUSH and CARP are about the same, sometimes
>> one of the algorithms wins, but we cannot tell you when to use CARP or
>> CRUSH because we cannot generalize their differences. Our experimental
>> data does not suggest any clear trends."
>>
>> That unpredictability may be the nature of the beast, of course, but I
>> currently see no reasoning or experimental data to confirm or deny that
>> sad hypothesis.

> CRUSH is no mystery [...] CARP is somewhat mysterious

You may know exactly how each algorithm works, but that does not really
matter in this context. What matters here is the _relative_ strengths
and weaknesses of an algorithm. That part is still a mystery (to me).
When is CRUSH better than CARP? When is CARP better than CRUSH?


> Please let me know if that is of interest to Squid

As of 04/18/2017, my personal answer to that question is "why bother?".

As of 04/19/2017, my personal answer to that question is "maybe": One
out of 2+ random(?) tests shows about the same CARP and CRUSH  results.
Another test shows better CRUSH performance. It is not clear to me
whether some other tests will show better CARP performance.

I hope the above clarifies my position. Others may have a completely
different opinion, of course, and my current position does not block
CRUSH acceptance.


Thank you,

Alex.