[iccrg] 回复： New draft submitted for draft-pan-tsvwg-hpccplus-02.txt

["Rui, Miao" <[e-mail]>]

Hi Michael,

Thanks for your notes. Yes, I was talking about the capacity issue with Michael Scharf in another thread. Having switch-assisted congestion control does not mean that switches need to ``allocate'' bandwidth in a dual-mode. Instead, we can still implement a primal-mode algorithm with precise congestion information provided by switches for fast utilization convergence (though it will take longer for fairness to converge).

Yes, having switch-assisted is possible with newer hardware and many vendors now support this. But we are careful about the primitives supported. 

First, in HPCC++ we only ask the switch to support write-to-packet operation which is already supported in inband telemetry (like IOAM mentioned earlier). However, IMO adding a read-from-packet operation would trigger more interesting research ideas.

Second, there are many choices of load information like tx/rx rate, qlen, timestamp, etc. The key design of HPCC++ is to define ``inflight byte'' to quantify the load, which can be calculated by tx rate, queue length, timestamp. Because inflight byte can tell both overload or underload of the link for us to increase/decrease the rate multiplicatively. On contrary, traditional wisdom of having only the queue-length as the feedback can only guide the rate decrease but not increase. Besides, it subjects to fluctuation unless a certain threshold is held. We also compare RX and TX rate as well in the paper. HPCC++ uses TX rate while XCP uses RX rate. Generally, we view the TX rate is more precise because RX rate and queue_length are overlapped and thus it causes oscillation.

Thanks,
Rui



------------------------------------------------------------------
发件人：Michael Welzl <[e-mail]>
发送时间：2020年12月15日(星期二) 23:11
收件人："Pan, Rong" <[e-mail]>
抄　送："Miao, Rui(缪睿)" <[e-mail]>; "Liu, Hongqiang(洪强)" <[e-mail]>; jri.ietf <[e-mail]>; iccrg <[e-mail]>; "Lee, Jeongkeun" <[e-mail]>; Barak Gafni <[e-mail]>; Yuval Shpigelman <[e-mail]>
主　题：Re: [iccrg] New draft submitted for draft-pan-tsvwg-hpccplus-02.txt

Hi Rong,

On Dec 16, 2020, at 3:09 AM, Pan, Rong <[e-mail]> wrote:
Michael,

I read the paper that you referenced.

Thanks!  :-)


 Indeed, it is very similar to HPCC++.  Time has finally come to do switch-assisted congestion control 😊.

Yes, that’s what I thought when I first saw this - now there’s hardware supporting such mechanisms!


 I am not clear about your comment on “path segments”. Data center network should be able to support it end-to-end, right?

Oh sorry, is this the only context here? (It was, in the HPCC paper).  Then yes, of course!  I don’t see a problem with this for data centers.
(except for, possibly, some of the issues with “capacity" that Michael Scharf listed in his interesting email here - I see you were not in cc:
https://mailarchive.ietf.org/arch/msg/tsvwg/37SUp8JSE_tJXzKM6-tDcOhJaeA/   - but I don’t think that these *always* apply, under all conditions…  surely there must be ways around them, at least in datacenters)

Cheers,
Michael