In message <4552f0907735844e9204a62bbdd325e732aff...@nkgeml508-mbx.china.huawei.com> Mingui Zhang writes: > Hi Curtis, > > I am very happy that you know GreenTE so well. Very good comments. As > the designer of the GreenTE model, I feel obliged to clarify some of > the confusions. Please see in-lines below. > > Thanks, > Mingui Zhang > Huawei Technologies
Mingui, Thank you for the clarifications. A few comments are inline. Curtis > >-----Original Message----- > >From: [email protected] [mailto:[email protected]] On Behalf Of > >Curtis Villamizar > >Sent: Thursday, February 07, 2013 4:05 AM > >To: Balaji venkat Venkataswami > >Cc: Shankar Raman M J; [email protected] > >Subject: Re: Power aware networks : Comments requested from routing > >community > > > > > >In message > ><[email protected] > >mail.com> > >Balaji venkat Venkataswami writes: > > > >> As you might recall one such research paper on GreenTE by Beichuan > >> Zhang rang a lot of bells in the IETF by collecting the ANRP prize. So > >> please dont misstate facts as you might know them and shove your ideas > >> of how networks are built down our throats. > > > >That was a research paper published in IEEE International Conference > >on Network Protocols (ICNP). > > > >Power savings potential (a computation, not a measurement) were as > >high as 20-40%. That is good but ... > > > >Line cards considered were old OC3, OC48, and one port OC192 cards. > >The paper ignores the reality that shutting off one port on a > >multiport card makes little difference, certainly far less than a > >reduction of 1/N. The power consumption of transport was not > >included. Impact on fault restoration was not considered. Lots of > >other flaws, but still good work and a potential 20-40% is nice, even > >if factoring in reality might get only half that or less. > > If you read the model, you will find a variable B_m. It indicates the > base power of the line card. This base power is high and can only be > saved when all ports on a multiport card are shutting off. So the > model did consider the reality that "shutting one port saves far less > than a reduction of 1/N". Having said that, I agree that your reality > is based on most current practice. But vendors can purposely provide > line cards with single port. Then the card can be shut off when the > port is shut off. Also, don't forget about composite links which are > powered physically separately. GreenTE is now able to handle composite > links. We have a demo website to show this: http://greente.65536.cn/. You are correct. B_m does consider baseline power. Line cards with single ports are impractical. The fastest interfaces are 100 Gb/s. Packet processors ranging up to 500 Gb/s (line side plus fabric side, so max 2x 100G) exist today (though they are few and very recent) and 100G chips exist today (100G line side, plus > 100G fabric side). Some chassis today make use of a 500+ Gb/s card slot. Composite Link or any other means of shutting down specific ports may help. Shutting down ULH transport capacity on a DWDM is more problematic than shutting down a VSR Ethernet. Even so, when the forwarding chip supports more than one port, as is common, only the MSA can be shut off. In such a case, shutting down the receive side saves nothing, so either side could initiate a bring up. For lower speed interfaces, today's forwarding chips support on the order of 12-24 x 10 GbE and much larger numbers of GbE. > Let me add another positive reality: when link cards are shutting off > in a chassis, blowers in the same chassis may be shutting off or > operate at a low speed. This helps the whole chassis save more power. Yes. Even running an average packet size of larger than minimum packet size should reduce power draw (less PPS) unless the vendor is cheating a bit in the max power draw figure. The fan speed will not be at max except with fully utilized cards in all slots and AC fail. > >The point though is that a 1 port OC192 at 174 watts (10 Gb/s) is no > >match for a 500 Gb/s card at roughly 650 watts. That is more than > >40%. It is more than a decimal order of magnitude greater. It is not > >comparing apples to oranges, one is an IP/MPLS router with full > >routing tables and short haul interfaces, the other does OTN switching > >drives either short haul or long haul transport interfaces, and could > >do IP/MPLS, but with only IGP routes and MPLS LSPs at that power > >figure. So maybe the way the network is built needs to change. > > > > What do you mean about the number "40%"? Anyway, line cards consume a > large percentage of energy of the whole Chassis and this percentage is > increasing fast. This judgment always holds. > > GreenTE only cares about line cards. So 20-40% saving actually means > 20-40% of the total power consumed by all line-cards. I think the range of potential savings in the figures indicates a maximum of 20-40% saving network wide. > >Anyway, if *I* were to continue this discussion, then *I* should get > >off the IETF lists and go write an academic paper. [Or better yet, > >work for a service provider or equipment vendor and work toward > >building better equipment and networks.] > > > >The point is that efficiency has gone up by orders of magnitude as we > >went from OC3 to OC12 to OC48 to OC192 (circa 1995-2000) and then > >multiport OC192 and OC768 and 10GbE (circa 2005-2010) to 100GbE and > >ODU4 (somewhat recent) to multiport 100GbE and ODU4 (even more > >recent). Power per bit drops about one to two decimal order of > >magnitude per decade and for any given a specific snapshot in the > >technology timeline, there are means to reduce power that get far more > >than the 20-40% that GreenTE potentially offers. > > Besides the power saved from the efficiency of the hardware, GreenTE > can add another 20-40% saving, right? :) Yes. I am not arguing against continued work on GreenTE. I am saying that it is still in the research stage and there are bigger gains elsewhere. > >Of course a new generation of equipment means CapEx and GreenTE is > >perceived as just routing and therefore free, so therefore the > >interest. > > > >Another point is that no one ever tried this on a real network. With > >RSVP-TE no protocol changes would be necessary, just offline > >computation, a configured preferred explicit-path, and some interface > >shutdowns. I don't know that anyone has even measured the reduction > >in power of an idled or administratively shut down interface or line > >card. There was nothing in the paper on this. > > > >> > If you want to try to advance a research paper with your theories on > >> > power reduction, please choose an appropriate venue such as a refereed > >> > technical journal. > > > >I repeat my main point. [And hope to be among the reviewers] > > > >Also IMHO GreenTE as described by Zhang is more likely to be feasible > >with MPLS and RSVP-TE than with OSPF or BGP or PIM, etc. But then, > >the feasibility and potential benefits of GreenTE using IGP, BGP, and > >PIM is what your academic paper can try to defend. > > When a link is shut off, no control messages (e.g., HELLOs) can be > delivered on it. Link state routing protocols (e.g., ISIS, OSPF), is > required to be aware of this. The network should never be partitioned. If partitioned due to network fault, leaving the recieve side up with minimal logic enabled allows either side to enable reactivation. Detection of a partition could also trigger bringing up idled capacity. > [end of my clarifications] Thanks again for the clarifications. GreenTE is a worthwhile research topic to pursue. My primary point is that it still remain a topic of research and attempts at IETF standardization are premature. _______________________________________________ rtgwg mailing list [email protected] https://www.ietf.org/mailman/listinfo/rtgwg
