Last Call Review of draft-ietf-nvo3-vmm-04
|Requested rev.||04 (document currently at 05)|
|Type||Last Call Review|
|Team||Transport Area Review Team (tsvart)|
|Requested by||Matthew Bocci|
|Authors||Linda Dunbar, Behcet Sarikaya, Bhumip Khasnabish, Tom Herbert, Saumya Dikshit|
|Draft last updated||2018-09-03|
Opsdir Last Call review of -03 by Mahesh Jethanandani
Rtgdir Last Call review of -03 by IJsbrand Wijnands (diff)
Tsvart Last Call review of -04 by Bob Briscoe (diff)
This document discusses TCP connection handling when a virtual machine moves. During the Ops Area review of this NVO3 draft, which has recently gone through WG last call, it was therefore proposed that this was reviewed from a TCP perspective.
|Reviewed rev.||04 (document currently at 05)|
|Review result||Not Ready|
I have been selected as the Transport Directorate reviewer for this draft. The Transport Directorate seeks to review all transport or transport-related drafts as they pass through IETF last call and IESG review, and sometimes on special request. The purpose of the review is to provide assistance to the Transport ADs. For more information about the Transport Directorate Reviews and the Transport Area Review Team, please see https://trac.ietf.org/trac/tsv/wiki/TSV-Directorate-Reviews In this case, very very few of the review comments relate to transport issues, although the greatest issue concerns a desire that the network could pause or stop connections during L3 VM Mobility, which is certainly a transport issue. ==Summary== The technical aspects of the draft concerning L2 VM mobility (within a subnet) seem sound. However, this is only part of the draft, which has the following issues: #. The introduction does not say what the purpose of publishing this draft is. It seems that, rather than describing a specific protocol or protocols, it intends to describe the overall system procedure that would typically be used in DCs for VM mobility. It is tagged as a BCP, but it does not say who needs this BCP, why it is useful for the IETF to publish this BCP, how wide the authors' knowledge is of current practice (given DCs are private), or why this is a BCP rather than a protocol spec. The draft starts out (S.3) as if it intends to say what a good VM Mobility protocol should or shouldn't do, but the rest of the document doesn't give any reasoning for these recommendations, it just asserts what appears to be one view of how a whole VM Mobility system works, sometimes referring to one example protocol RFC for a component part, but more often with no references or details. #. It does not seem as if the NVO WG has discussed the purpose of using normative text in this draft. See detailed comments. #. The draft silently slips back and forth between VM mobility and VM redundancy, without recognizing the differences. See detailed comments. #. Please adopt different terminology than "source NVE" and "destination NVE", which are really poor choices of terms for an intermediate node. See detailed comments. Why not use "old NVE" and "new NVE", which is what you mean? #. Applicability is fairly clearly outlined, but it is not clear whether hosts corresponding with the mobile VMs are part of the same controlled environment or on the uncontrolled public Internet. See detailed comments. #. Section 4.2.1 on L3 VM mobility reads like some potential half-thought-through ideas on how to solve L3 mobility, rather than current practice, let alone best current practice. Either current practice should be described instead, or the scope of the draft should be narrowed solely to L2 VM mobility. See detailed comments. # The VM's file system is described as state that moves with the VM (S.6), but VM mobility solutions often move the VM but stitch it back to its (unmoved) storage. Conversely, the storage can also move independent of the VM. #. The draft omits some of the security, transport and management aspects of VM mobility. See detailed comments. #. The draft reads as if different sections have been written by different authors and no-one has edited the whole to give it a coherent structure, or to ensure consistency (both technical and editorial) between the parts. See detailed comments. #. The quality of the English grammar does not allow a reviewer to concentrate on the technical aspects rather than the English. It would have been useful if one of the English-speaking co-authors had improved the English before submission for review. See detailed comments. ==Detailed Comments== ===#. Normative statements=== In the body of the document, there is just one occurrence of normative text (actually two "MUST"s, but both state a common requirement - just written separately for IPv4 and IPv6). This merely serves to imply that everything else the document says is less important or optional, which was probably not the intention. At the start there is a requirements section, which states what a VM Mobility protocol "SHOULD" or "SHOULD NOT" do. I think this is intended as a set of goals for the rest of the document. If so, these "SHOULDs" are not intended to apply to implementations, so they ought not to be capitalized. The first requirement, "Data center network SHOULD support virtual machine mobility in IPv6", is written as a requirement on all DC networks, not on implementations. I assume this was intended to read as "Data center network virtual machine mobility protocols SHOULD support IPv6". Even then, it doesn't really add anything to say VM mobility should support v6 and it should support v4. A L2 solution won't. While undoubtedly, a L3 solution will at least support one of them. I'm not sure that 'protocol' is the right word anyway; I think 'VM Mobility procedure' would be a better phrase, because it includes steps such as suspending the VM, which is more than a protocol. The requirement "Virtual machine mobility protocol MAY support host routes to accomplish virtualization", is not followed up at all in the rest of the draft. Even if this requirement stays, the last 3 words should be deleted. By the end of the draft, the solution falls far short of the most relevant "Requirements" anyway, so one assumes the title of the section ought to have been "Goals". Specifically, even in the simpler case of L2 VM mobility, S.4.1 says that triangular routing and tunnelling persist "until a neighbour cache entry times out". A cache timeout is about 10 orders of magnitude longer than the requirement to only persist "while handling packets in flight", which would be a few milliseconds at most (the time for packets to clear the network that were already launched into flight when the old VM stopped). Whatever, it would be preferable for the draft to give rationale for these requirements, rather than just assert them. This would help to shed light on the merits of the different trade offs that solutions choose. ===#. Mobility vs. Redundancy=== Redundancy and mobility have a lot of similarities, but they have different goals. With mobility, it is necessary to know the exact instant when one set of state is identical to the other so it can hand over. With redundancy, the aim is to keep two (or more) sets of state evolving through the same sequence of changes, but there is no need to know the point at which one is the same as the other was at a certain point. The draft slips from mobility to resilience in the following places: * S.2. Terminology: Warm VM Mobility is defined without any ending, as if it is permanent replication. * S.7. "Handling of Hot, Warm and Cold Virtual Machine Mobility" is actually all about redundancy, and doesn't address mobility explicitly. ===#. Terminology=== Packets run from the source at A to the destination at B via NVE1, then via NVE2. Please don't call NVE1 and NVE2 the source NVE and the destination NVE. In future, no-one will thank you for the apparent contradictions when they continually stumble over phrases like this one in S.4.1: "...send their packets to the source NVE". The term "packets in flight" is used incorrectly to refer to all the packets sent to the old NVE after the VM has moved, even if they were launched into flight long after the old VM stopped receiving packets. BTW, I think s/before/after/ in: "that have old ARP or neighbor cache entry before VM or task migration". I think: s/IP-based VM mobility/L3 VM mobility/ throughout, because "based" sounds (to me) like the mobility control protocol is over (i.e. based on) IP. ===#. Applicability=== In section 4.2 it says that the protocol mostly used as the IP based task migration protocol is ILA. This implies that all hosts corresponding with the mobile VMs are either part of the same controlled environment, or they are proxied via nodes that are part of the same controlled environment (I only have passing knowledge of ILA, but I understand that it depends on ILA routers on the path). If I am correct, this aspect of scope needs to be made clear from the start. Also under the heading of applicabiliy, the sentence "Since migrations should be relatively rare events" appears very late in the document (S.4.2.1). The assumed level of churn ought to be stated nearer the start. ===#. L3 Mobility=== L2 VM mobility is independent of the application, because resolution of L2 mappings is delegated to the stack. In contrast, L3 VM mobility is only feasible under certain conditions, because an application needs an IP address to open a socket (resolution of DNS names is not delegated to the stack, and apps can use IP addresses directly anyway). Examples of the 'certain conditions': a) /All/ applications used in the whole DC load balancing scheme contain IP address migration logic for /all/ their connections; b) VMs running solely applications that support IP address migration register this fact with the NVA, and it only select such VMs for mobility. c) An abstraction is layered over /all/ the IP addresses exposed to applications (at both ends) so that the IP addresses that applications use are solely identifiers (e.g. ILA, LISP, HIP), not also locators. The introduction says the draft is about VM mobility in a multi-tenant DC, so the DC admin will not know the range of applications being used. This excludes condition (a) above. When the draft says "...if all applications running are known to handle this gracefully...", it doesn't quantify just how restrictive this condition is, and it gives no explanation of how this knowledge might be 'known' or which function within the system 'knows' it. S.4.2.1 contains what seems like plenty of arm-waving. * "TCP connections could be automatically closed in the network stack during a migration event." o There is no TCP connection state in the network stack. o Even if the network starts to drop every packet, the TCP connection state persists in the end-points for a duration of the order of 30-90 minutes (OS-dependent) before TCP deems the connection is broken. o Other transport protocols have similar designs (including the app-layer of protocols over UDP). * "More involved approach to connection migration": o pausing the connection [does this refer to an actual feature of any L4 protocol?] o packaging connection state and sending to target [does this assume logic written into the application, or is this assuming the stack handles this and the app is restricted to using some form of separate identifier/locator addresses?] o instantiating connection state in the peer stack [ditto?]. There's some arm-waving in S.7 too: "Cold Virtual Machine mobility is facilitated by the VM initially sending an ARP or Neighbor Discovery message at the destination NVE but the source NVE not receiving any packets inflight." [How is it arranged for the source NVE not to receive any packets in flight?] And in S.7: "In hot standby option, regarding TCP connections, one option is to start with and maintain TCP connections to two different VMs at the same time." [This sounds like resilience logic has been written into the application, which would be a special case but not something VM mobility infrastructure could depend on.] ===#. Gaps=== #. Security Considerations: repeats issues in other drafts that are not specific to mobility, but it does not mention any security issues specifically due to VM mobility. It says that address spoofing may arise in a DC (sort-of implying it is worse than in non-DC environments, but not saying why). The handshake at the start of a connection (e.g. TCP, SCTP, QUIC) checks for source address spoofing. So L3 VM mobility would be more vulnerable to source address spoofing in cases where the mobile VM was the connection initiator and there was not a new handshake after the move. However, this draft does not contain any detailed mobility protocols, so it is not possible to identify any specific security flaws. #. Transport Issues: Effect of delay on the transport: Cold mobility introduces significant delay, and other forms less, but still some delay. It should be pointed out that some applications (e.g. real-time) will therefore not be useful if subjected to VM mobility. Similarly, even a short period of delay will drive most congestion controls to severely reduce throughput. These points might be self-evident, but perhaps they should be stated explicitly. BTW, in the L3 VM mobility case, the draft often refers to TCP connections, but the address bindings of any transport protocols would have to be migrated due to VM mobility (e.g. SCTP; sequences of datagrams over UDP; streams over UDP such as with RTP, QUIC). #. Management Issues: perhaps the draft ought to recommend statistics gathering (e.g. time taken, amount of duplicate data) to aid a DC's future decisions on the cost-benefit of moving a VM. The OPSDIR review says a BCP does not /have/ to describe management issues, but this document seems to describe a whole system procedure, not just a protocol, which then surely includes the management plane. ===#. Incoherent Structure=== S.4.1. happens to talk about VMs moving, while S.4.2. happens to talk about tasks moving, but this is not the distinguishing aspect of these two sections (anyway, S.2. says "the draft uses task and VM interchangeably"): * "4.1 VM Migration" is about "L2 VM Mobility" so this ought to be the section heading, * "4.2 Task Migration" is about "L3 VM Mobility" so this ought to be the section heading. It would also help not to switch from VM to task across these sections - it's just a distraction. S.4.1 needs better signposting of where each sub-case ends (Subsections might be useful to solve this): * IPv4 * end-user client * 2 paras starting "All NVEs communicating with this virtual machine..." [Not clear that the end-user case has ended and we have returned to the general IPv4 case?] * IPv6 [Strictly, it still hasn't said whether the end-user client case has ended.] [Also, it doesn't explain why there is no need for an end-user client case under IPv6?] Sections 5 & 6 seem to be about either L2 or L3 mobility, whereas Sections 7 & 8 seem to be restricted to L2. The draft vacillates over what to do with packets arriving at the old NVE in the L3 case (see also L3 mobility above): * S4.2 first says packets are dropped, possibly with an ICMP error message; o then later it says they are silently dropped; o then in the very next sentence it says either silently drop them or forward them to the new location * S.5 says they should not be lost, but instead delivered to the destination hypervisor o then it describes how they are tunnelled (which is not the same as "forwarding"). The order in which all the stages of mobilty are given is jumbled up across sections that also appear in arbitrary order: * S.5 prepares, establishes uses then stops a tunnel, but it doesn't say where the other stages fit between these steps o When tunneling packets, it talks about the *migrating* VM not the *migrated* VM, which implies tunnelling has started before the new VM is running. Does this imply there is a huge buffer? o It says "Stop Tunneling Packets - When source NVE stops receiving packets destined to..." but it is never clear when a source has stopped sending packets to a destination, unless it explicitly closes the connection (e.g. with a FIN in the case of TCP). Often there are long gaps between packets, because many flows are 'thin' (meaning the application frequently has nothing to send). These gaps can last for milliseconds, hours or even days without any implication that the connection has ended. * Then S.6. describes moving state, but doesn't say that this is not after the previous tunnelling steps (or where it fits within those steps). * Then S.7 describes hot, warm and cold mobility, but doesn't lay out the tunnelling or steps to move state in each case. * Then S.8 says it's about VM life-cycle, but just gives the very first 3 steps for allocation of resources to a VM, then abruptly ends, without even starting the VM, let alone getting to move it. S.5 exhibits another inconsistency by talking about the hypervisor, not the NVE. ==#. Nits== Nits with the English are too numerous to mention them all. Below are pointers to general problems as well as some individual instances. S.4 "Layer 2 and Layer 3 protocols are described next. In the following sections, we examine more advanced features." s/following/subsequent/ S.4.1 Expand WSC, MSC and NVA on first use. s/the VM moves in the same link/the VM moves in the same subnet/ "i.e. end-user clients ask for the same MAC address upon migration. [...] to ensure that the same IPv4 address is assigned to the VM." I think s/IPv4/MAC/ was intended? " All NVEs communicating with this virtual machine uses the old ARP entry. If any VM in those NVEs need to talk to the new VM in the destination NVE, it uses the old ARP entry." Repetition: these 2 sentences say the same. (The mistake is also repeated when these 2 sentences are repeated for IPv6). S.4.2.1 s/Push the new mapping to hosts./Push the new mapping to communicating hosts./ S.5. The IPv4/IPv6 pairs of paras for "tunnel estabilshment" and "tunneling packets" only differ in the words "IPv4"/"IPv6". So in each case a single para could be given for IP (irrespective of whether v4 or v6).