Performance Implications of Link-layer Characteristics (PILC) BOF - IETF 44 Reported by Mike Kallas (mailto:mkallas@nortelnetworks.com) Mailing list contact information: http://pilc.grc.nasa.gov/pilc MINUTES Chairs: Mark Allman (mailto:mallman@grc.nasa.gov) Spencer Dawkins (mailto:sdawkins@nortelnetworks.com) Aaron Falk (mailto:a.falk@ieee.org) Internet Draft: draft-montenegro-pilc-ltn-01.txt (currently individual submission) The second PILC BOF was held March 18, 1999 at IETF 44 in Minneapolis, Minnesota. Aaron Falk, BOF Co-chair, started off with some history of the PILC effort and experiences gained from the TCP over Satellite, and presented a draft charter for the proposed PILC working group. His slides are available at: http://pilc.grc.nasa.gov/pilc/. Mark Allman, BOF Co-chair, continued the discussion by summarizing the interests that have been expressed on the PILC mailing list, and leading the discussion of the draft charter. Mark's slides are also available from the PILC web page. Three major types of documents will be produced in PILC: "Advice for Internet Subnetwork Designers", recommendations about using (and not using) TCP Performance-Enhancing Proxies, and recommendations about mitigations for specific network characteristics. In addition, a pointer document for certain communities of interest may be produced, depending on the need for a specific network type to deal with more than one "unfortunate" link-layer characteristic and the potential for "unfortunate" interactions between mitigations. Many questions and comments were made about these "Pointer" documents. The conclusion was to do the work on the mitigations documents and then see whether the pointer documents are still required. The major threads on the PILC list to date have been on these topics: Links with high BER, Inconsistent BER, Links with occasional outages, and Low bandwidth links. (Although, there have been brief threads on other topics). A fairly contentious discussion on PEPs ensued. The consensus seemed to be for including a document discussing the advantages/disadvantages of using PEPs. Both Aaron and Mark suggested specifications that don't target a specific type of network (say, "TCP over wireless"), but target a specific network characteristic (say, "TCP over links with high BERs after link-layer correction has been applied"). The rationale was to make sure the TCP community doesn't ignore PILC specifications as being tied to a specific network type - Aaron expressed the opinion that this had happened in TCPSAT, and slowed down the process of getting consensus on TCPSAT Internet-Drafts. Phil Karn, who has been working on a draft of the "Advice for Internet Subnetwork Designers" document (draft-in-process at http://people.qualcomm.com/karn/pilc.txt) presented a proposed outline of the document, collected additions, and identified additional contributors. Phil's document table of contents at the beginning of the meeting was: MTU size, packet size Framing in the sub-net Connection oriented sub-net Link characteristics QoS delay / BER Reordering Mobility Multicasting Routing or use IP Security Discovery protocols broadcasting Congestion feedback signals These subjects were suggested as additions during the discussion: MAC algorithm, tuning Media access affects link characteristics Link layer should take advantage of burstiness, clumpiness of error helps TCP Total error statistics behavior (not just BER) Compression Asymmetry Buffering Fairness Vs performance Implications of protocol design on power consumption for mobile/wireless devices Significant interest and energy was expressed in support of the charter items discussed. The initial areas of concentration, based on mailing list traffic to date, will be "thin", or low-bitrate, networks, and lossy networks. Additional areas may be addressed as communities with interest and energy are identified. Steve Deering noted that interim meetings could easily be held to focus on a specific link characteristic. There was substantial discussion of a number of topics. Relevant comments are summarized below. Q: Is PILC only about TCP? What about RTP, multicast? A: No. We are also working on link layer issues today. Some characteristics originally proposed for PILC were IP layer issues, and UDP, RTP, and multicast are all acknowledged as possible areas of work for PILC. The energy that has materialized so far in PILC has been focused on link layer and TCP issues. Q: Do all wireless networks have same issues? Why is a pointer document needed? A: We can't anticipate all "wireless" characteristics, and there are different characteristics for different types of wireless networks. Each "mitigation" document will look at different link characteristics. The Pointer document for wireless needs to be flexible so that we can add more sections as needed. Different solutions work for different networks. Also note that "wireless" was used as shorthand for "wireless WAN" in the discussion. Q: Shouldn't we remove "wireless" from charter? A: We need to keep the word in the charter as a commitment that PILC will work to meet the requirements of the wireless community, but ("you're right,") wireless environments are different and changing. We will not group all wireless into one category, so different Pointer documents may be produced. We will focus discussions on link characteristics, and not on network types like wireless. We will use wireless as an example Q: We're concerned about usefulness of the Pointer documents. Should Pointer documents be IETF RFCs? Would a conference paper be better? A: Many groups are publishing similar pointer documents in many places (conference proceedings, industry forums). There is still a place for documentation of mitigations that involve TCP that have been reviewed within IETF. Q: Is designing a way for the applications to determine underlying network characteristics in scope? A: No, network knowledge of link topology is a different problem. and is also too big. We can discuss this on the list if you disagree. Q: How good does the link need to be? If you can not achieve recommendations, then what? A: Van Jacobson noted that error behavior is non-linear. There is a tradeoff between MAC design and error response. Error response is not linear and breakpoint is distinct and visible. Chairs asked for people with energy to work on problems not in the proposed charter. Hari Balkrishnan volunteered to work on documents related to asymmetric networks. It was also suggested that this WG should consider architecture.