DECADE WG, IETF79 Beijing, November 2010 ======================================== (version 2 of minutes) DECADE WG Minutes Meeting: IETF 79, Thursday, November 11, 2010 Location: Beijing Shangri-La Hotel, Emerald Conference Room, 1320-1520 Chairs: Rich Woundy, Haibin Song Minutes: Christian Schmidt, edited by Haibin Song and Rich Woundy According to the blue sheets, there were 59 attendees. [These minutes represent a condensed summary of the meeting. Fully detailed discussions can be heard on the audio recording of the meeting, currently found at , starting at offset 00:13:42.] Agenda ------ 1. Agenda bash, chairs, 5 minutes, 1520-1525 2. Problem statement, Haibin Song, 5 minutes, 1525-1530 draft-ietf-decade-problem-statement-00 3. Architecture, Richard Alimi, 20 minutes, 1530-1550 draft-alimi-decade-arch-01 4. Requirements, David Bryan, 20 minutes, 1550-1610 draft-ietf-decade-reqs-00 5. Requirements for accessing data in network storage, Börje Ohlman, 15 minutes, 1610-1625 draft-ohlman-decade-add-use-cases-reqs-02 6. Survey, Akbar Rahman, 15 minutes, 1625-1640 draft-ietf-decade-survey-01 7. Integration examples, Lijiang Chen, 20 minutes, 1640-1700 draft-chen-decade-intgr-livestr-exmp-00 draft-huang-decade-integration-example-00 8. Data regeneration, Jin Zhao, 15 minutes 1700-1715 draft-wang-decade-data-regeneration-01 9. Conclusion and next steps, chairs, 5 minutes, 1715-1720 Haibin Song and Richard Woundy led the agenda bashing and milestone discussion. They noted that the DECADE WG moved from the Application Area to the Transport Area this week. The WG priorities are to prepare the Problem Statement and survey drafts for WG last call, and to accept the Architecture draft as a WG draft. There were no comments nor questions. [audio offset 00:17:30] Problem statement ----------------- Haibin Song presented the problem statement draft. About 30 people have read the problem statement draft. The chairs are looking for document reviews, to be sent to the mailing list. After the authors incorporate the feedback from the reviews in a new draft iteration, the chairs will take the document to WGLC. There are three volunteers to review the draft: David Bryan, Akbar Rahman, and Ning Zong. [audio offset 00:22:28] Architecture ------------ Richard Alimi presented the architecture draft. [audio offset 00:34:26] Richard Alimi discussed explicit control by applications, including control of the DECADE server. Börje Ohlman proposed not to use the term "DECADE server", and suggested that it would be better to talk about "DECADE storage". The service that DECADE offers is storage. Richard Alimi: We want to differentiate DECADE service from the underlaying DECADE storage (such as a SAN). Particular location should be addressable by the client. Börje Ohlman: The provider offers a storage service, not access to a particular physical server. Richard Alimi: That's a good comment, as long as the storage service does not span across the network, and specific locations can be chosen. [audio offset 00:43:15] Richard Alimi discussed options for data object naming. David McDysan: Is this what you call self-certifying names? Richard Alimi: Yes. David McDysan: That seems to meet a requirement that Börje raised in Maastricht (IETF78) that the same content should have the same identifier. Is the hash over the storage location or over the content? Richard Alimi: The hash is computed over the content. David McDysan: Collisions are possible with a hash, even if they are a low probability. How do we avoid collisions? Börje Ohlman: We discuss this in another draft that was provided to PPSP. You need a sufficiently large namespace to ensure there is a very low probability of collision. Haibin Song: What are the basic requirements for data object naming? Richard Alimi: Good question. An obvious requirement is uniqueness with a sufficient probability. Please contribute more requirements. Börje Ohlman: One way to ensure naming uniqueness is to create a two part label: one part with hash of the publisher, and the other part with the hash of the content. Richard Yang: How to divide objects? Client generated (e.g. block sizes) or server generated (e.g. hashed using a Rabin Signature Scheme)? Which scheme is better? Jan Seedorf: Since the main use case for Decade is P2P applications, clients should do this. [audio offset 00:50:56] Richard Alimi discussed the resource control model. Richard Woundy: We should take this discussion to the mailing list. [audio offset 00:53:16] About 10 people have read the draft. Hum: Should the WG adopt the architecture draft (draft-alimi-decade-arch-01)? Outcome: There was weak consensus, but none against. Richard Woundy: We will take the adoption question to the mailing list. The architecture draft appears to be going into the right direction. [audio offset 00:54:38] Requirements ------------ David Bryan presented the requirements draft. There have been significant changes in the draft since the last meeting. About 10 people have read the new version of the draft. [audio offset 00:57:40] Open Item 1: Is the break down of issues correct? Are people comfortable with this approach? Nobody complained. Open Item 2: What types of applications besides P2P are in scope (eg distributed VoD)? We are not looking to build a generic data store system. Richard Alimi: Focus on the properties of application to be supported, not the names of the applications. David Bryan: We will try to respin as a characterization of the data access patterns in supported applications. [audio offset 01:01:20] Open Item 3: How would we enable duplicate data removal within a server? David Bryan: Doing duplicate data detection within a single server is trivial. If the names are generated from the content (eg hashing), it may be possible across distributed servers as well. There are no requirements to remove duplicate data across servers. Richard Alimi: There is another type of duplication detection: Is a version of an object already available on the server? This is an implementation issue only, to avoid a second download. Open Item 4a: Discovery: How do we locate some Decade server, not necessarily the “best” server. Open Item 4b: Mobility: How do we locate some Decade server for mobility / roaming? Richard Woundy: Do not want to repeat experiences from ALTO discovery. Do not re-invent service location here. [audio offset 01:06:38] Open Item 5: Can we include a caution against putting the “smarts” of the applications into DECADE? David Bryan: We want to keep Decade as simple and as generic as possible. Wording for this would be appreciated. Börje Ohlman: Are other users allowed to access the content, when owning user is offline or not available? Default operation needs to be clarified. David Bryan: Do you have a use case in mind, for when storage should be deleted when the user goes offline? Börje Ohlman: In case of sensitive data. David Bryan: Then you should not use DECADE. Let's take the discussion to the list. [audio offset 01:11:25] Requirements for accessing data in network storage -------------------------------------------------- Börje Ohlman presented the draft on Requirements for accessing data in network storage. Additional requirements have been added. Text has been reworked to be more specific. [audio offset 01:14:01] Akbar Rahman: IPTV is often associated with IP multicast. What is the underlying distribution? Börje Ohlman: Not assuming IP multicast. David Bryan: We might mean VoD here. Börje Ohlman: Two items to point out. The provider has Decade storage in the network, and the provider also has an agreement with the end user to use the storage of the user. Ning Zong: IPTV clients have to buffer video content, so the server cannot tell the client not to store anything. Richard Alimi: I assume "the server on the way" refer to the server in the data path. The server does not sniff data automatically. Börje Ohlman: Looking for a control protocol to mark the data for storage. Richard Alimi: Would you use something like DPI - deep packet inspection? Or application layer issue? Börje Ohlman: Did not think about yet. [audio offset 01:19:31] Akbar Rahman: If it is a unique name, by definition it would be unique across different applications. What is significant about uniqueness across applications? Börje Ohlman: One could enable name uniqueness simply by adding the application label to the name, but we want to avoid that. David Bryan: We need to be careful with this requirement, though the spirit of the requirement is good. The only way to practically satisfy this requirement (with uniqueness) is to generate names based on the content. Börje Ohlman: We do not want to introduce another naming authority for Decade. David McDysan: The spirit of this requirement could benefit from a use case example. [audio offset 01:23:57] Richard Woundy: Why do we need real-time support, if we only support VoD? David Bryan: It is a requirement to support pipelining. It is not about real-time transport. [audio offset 01:27:12] Richard Woundy: Thank you for working on this. We are going into the right direction. Where to go from here? - Should we include these requirements into a single WG requirements draft? - Should we maintain a separate draft? David Bryan: I would be happy to merge this. Decision about the requirements would be done on the list. Several requirements have to be discussed. There might be charter implications. Börje Ohlman: Do we want to include any use case discussion in the problem statement draft? Richard Alimi: We could describe the applications we want to support, e.g. Internet TV usage. Haibin Song: It would be OK to add certain use cases to the problem statement. Yunfei Zhang: You did not mean IPTV, right? Börje Ohlman: We are not very specific in this area. David Bryan: We should not talk about names, but the usage patterns and applicability statements. Börje Ohlman: Agreed. Richard Alimi: It is already in the problem statement and if not, it should be there. David McDysan: Maybe some of the material moves to the requirements and problem statement drafts, and maybe the remaining material stays in a separate draft as a future applicability statement. [audio offset 01:35:48] Survey ------ Akbar Rahman presented the DECADE Survey draft. [audio offset 01:40:56] Akbar Rahman: Named Data Networking is an interesting idea related to decade. Ning Zong : Which parts are related to decade? Akbar Rahman: Some key points are that there is a content store in the router, the content is retrieved using a unique name, and the content can be stored long-term. Ning Zong : That is quite unlike decade. Akbar Rahman: Agreed. But this is a survey, not a description of decade itself. [audio offset 01:49:09] About 6 people have read the current version of the draft. Richard Alimi: Could we use the same review process as for the problem statement? The chairs are looking for document reviews, to be sent to the mailing list. After the authors incorporate the feedback from the reviews in a new draft iteration, the chairs will take the document to WGLC. There are three volunteers to review the draft: Yunfei Zhang, David Bryan and Ove Standberg. [audio offset 01:51:52] Integration examples -------------------- Lijiang Chen presented the integration examples draft. (Presentation time was compressed to 10 minutes.) Richard Woundy: We don't have any milestones for integration examples, but it is in the decade charter. [audio offset 02:01:20] Richard Yang: There is a typo on slide 12. For startup delay, it should say "reduced to 1/3 when no storage". Richard Woundy: The improvement in startup delay is because more data is available to the swarm at the beginning of the content distribution process. The improvement doesn't help as much as the swarm matures. Yunfei Zhang: How many simultaneous connections can you obtain for one peer on average? Lijiang Chen: P2P applications use multiple connections, say 30-50. But there may not be as many connections per P2P client with the storage server. If a connection sustains a high transfer rate, then you don't need as many connections for good performance. About 6 people have read the draft. Richard Woundy: Would anybody object to take this Integration example draft as a WG draft? Song Habin: Should we merge the 2 Integration example drafts before adoption? Richard Woundy: Do we need a consolidated Integration example draft or not? Let us discuss this offline and take it to the list. [audio offset 02:05:27] Data regeneration ----------------- Xin Wang presented the Router Supported Data Regeneration draft. (Presentation time was compressed to 5 minutes.) Richard Woundy: Let us take it to the mailing list for applicability, and discuss how routers might interact with storage systems. This presentation concluded the working group session. [audio offset 02:15:37]