On Thu, Sep 25, 2014 at 10:02 AM, ramki Krishnan <[email protected]> wrote: > Hi Gorry, > > Many thanks for your detailed comments. > > The goal of this draft is to describe the commonly used traffic management > techniques in switches and routers. AQM (RED, WRED etc.) is an advanced > traffic management technique which we plan on addressing in a separate > draft. Tail drop is the commonly used technique in switches and routers for > addressing congestion control. Please see references to Cisco and Brocade > configuration guides - > (http://www.cisco.com/c/en/us/td/docs/routers/10000/10008/configuration/guides/qos/qoscf/10qqueue.html#wp1041728; > > http://www.brocade.com/downloads/documents/product_manuals/B_NetIron/NetIronUnified_05400a_ConfigGuide.pdf). > > BTW, David asked us at the last IETF meeting about AQM and the above response > was well received by the working group.
I was delighted to see this new wg attempting to tackle more complex scenarios for characterizing switch and router behavior, up to, and *almost* including fq and aqm techniques. A lot of the test ideas are still relevant at that level, and I'd hoped some of it would feed forward into the aqm evaluation guidelines. And while I do hope to one day live in a world where fq and aqm technologies are deployed by default, my principal hope from the bmwg is that the latency gets measured and reported everywhere appropriate and weighted more heavily than loss. I also note that the lightweight techniques netperf-wrapper uses, and the common data result format, could be extended to cover many of the tests discussed there so far, and it would be a boon to be able to gather and share detailed metrics rather than single point ones. netperf-wrapper has now been tested successfully up to 40GigE, and some folk are now using it to analyze the effects of shaving nanoseconds off of host generated traffic at the lowest layer interaction between the tx ring and qdisc... https://patchwork.ozlabs.org/patch/393034/ So I wouldn't mind a mention of netperf, d-itg, owamp, in the draft. netperf is used, rather than iperf, in the linux kernel networking community, because A) it's trusted at high rates, and B) the author responds to email. > We will address all your non AQM comments in the next version of the draft. > > Thanks, > Ramki > > -----Original Message----- > From: bmwg [mailto:[email protected]] On Behalf Of [email protected] > Sent: Thursday, September 25, 2014 12:33 AM > To: [email protected]; MORTON, ALFRED C (AL) > Cc: [email protected] > Subject: Re: [bmwg] [aqm] first WGLC on draft-ietf-bmwg-traffic > > > I have a few comments from a quick pass through this document. I think it is > useful to define such tests, and I suspect those performing the tests may > understand the wider context of testing a specific mechanism, but I have a > concern that the tests would produce strange results if deployed with > sophisticated traffic management, and this is not really discussed. I think > this needs to be addressed, and if necessary the cope of the tests clearly > stated. > > I am worried that while the tests may help characterise a single > shaper/policer, there is a need to extend the document at least to comment on > more complex interactions that can occur. > > QoS Frameworks define policies (e.g. DS) that can map DSCPs (or MF classified > traffic) to treatments, possibly aggregating behaviours for different traffic > aggregates. Treatments are not independent - traffic with one marking can and > does traffic with a different marking (well its intended to). Testing > multiple DSCPs in DS at the same time is more complex, it should not be > assumed that each of these can be characterised separately. > > AQM seeks to manage buffering, and this reaction is likely to be triggered by > test sequences. When triggered, the response (loss, delay) can impact the > flow's treatment long after the traffic burst that triggered AQM in the first > place. > > Many AQM methods are also used in combination with scheduling, again this can > be and often is non-trivial and depends on the number of flows and sometimes > the hashing of the flows to queues, etc within the scheduling framework. > > Other specific comments follow. > > Gorry > > --- > > Section 1: > Introduction needs to define terms when first used. > - would be good to define VLAN and DSCP > > > /sent onto to the next network/sent to the next network/ > - is it also possible to say "forwarded" rather than "sent"? > > Active Queue Management (AQM): > - would be good to cite: > https://datatracker.ietf.org/doc/draft-ietf-aqm-recommendation/ > > Please define MEF > > /of it's egress ports/of its egress ports/ > > /The Network Delay Emulator (NDE) is a requirement for the TCP > stateful tests/ > - probably true of any control-loop based congestion control (i.e. TCP is an > example, SCTP, DCCP and other congestion controls layered on UDP would also > need this (including some uses of RTP). > > Please define BDPs > > Section 3: > > /Another goal is to devise methods that utilize flows with > congestion-aware transport (TCP) / > - I think TCP is an example here, and needs an e.g., to avoid it seeming like > this is the only test case. Note also this comment applies later in the same > para. > > When it comes to an active technique, the concept of Burst Size Achieved > (BSA) and Lost packets (LP) is less clear, since won't this also depend on > the recent history of the flow traffic patterns and could be significantly > influenced by scheduling mechanisms between flows (i.e. > whether the flow is hashed to a separate queue, whether the flow is > classified as "new" or "old" and the total number of flows being serviced). > Performance could be quite different for a single flow test. > > With Active management, Packet Delay (PD) also becomes a function of traffic. > > Section 4.2 - I think the document should mention the need to test for other > non TCP congestion-controlled traffic, although I could see that for > simplicity the tests are defined only for TCP. > > 5.1.1 Burst Hunt with Stateless Traffic > - A cautionary note here could be helpful, that advises that if AQM is used, > the bursts must be sufficiently separated in time that they do not > cumulatively trigger a response - since AQM algorithms can have history, > treatment of a subsequent burst can be influenced by an earlier burst. > > 5.2.1. and 6.4: > - could note that multiple flows may also be queued and scheduled separately > within the device. > > Informative References > -missing > > >> BMWG (and AQM): >> >> A WG Last Call period for the Internet-Draft on Traffic Management >> Benchmarking: >> >> http://tools.ietf.org/html/draft-ietf-bmwg-traffic-management >> >> will be open from 23 September 2014 through October 21 2014. >> >> These drafts are beginning the BMWG Last Call Process. See >> http://www1.ietf.org/mail-archive/web/bmwg/current/msg00846.html >> >> which says: >> ... >> The primary change is a requirement that approximately four WG members >> complete a review template for the draft (or set of drafts), upon >> entry to the Last Call Process. The WG reviewers should be outside the >> body of active contributors (editors + authors). >> The goal is to produce a quality review with valuable feedback on the >> draft(s), and not to set a minimum number of people willing to >> complete the review template. Directed feedback, regardless of >> quantity, is a better condition by which BMWG can more clearly assess >> a draft's readiness to advance. >> (the template is appended to the archived message) >> >> Please read and express your opinion on whether or not these >> Internet-Drafts should be forwarded to the Area Directors for >> publication as Informational RFCs. Send your comments to this list or >> [email protected] and [email protected] >> >> Al >> bmwg co-chair >> >> _______________________________________________ >> aqm mailing list >> [email protected] >> https://www.ietf.org/mailman/listinfo/aqm >> > > _______________________________________________ > bmwg mailing list > [email protected] > https://www.ietf.org/mailman/listinfo/bmwg > > _______________________________________________ > aqm mailing list > [email protected] > https://www.ietf.org/mailman/listinfo/aqm -- Dave Täht https://www.bufferbloat.net/projects/make-wifi-fast _______________________________________________ aqm mailing list [email protected] https://www.ietf.org/mailman/listinfo/aqm
