12
HUAWEI TECHNOLOGIES CO., LTD. Fast Configuration Change Impact Analysis for Network Overlay DCNs Lizhao You, Hao Tang, Jiahua Zhang, Xiao Li Huawei Technologies [email protected] APNET’20, August 3-4, 2020

Fast Configuration Change Impact Analysis for Network ...conferences.sigcomm.org/events/apnet2020/material/PPT/lz_apnet.pdf · HUAWEI TECHNOLOGIES CO., LTD. Fast Configuration Change

  • Upload
    others

  • View
    4

  • Download
    0

Embed Size (px)

Citation preview

  • HUAWEI TECHNOLOGIES CO., LTD.

    Fast Configuration Change Impact Analysis for Network Overlay DCNs

    Lizhao You, Hao Tang, Jiahua Zhang, Xiao Li

    Huawei Technologies

    [email protected]

    APNET’20, August 3-4, 2020

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Network Overlay DCNs

    Page 2

    ToR Leaf

    Border Leaf

    & Spine

    PE Router

    Server

    FW,

    LB,

    B C D

    ToR switches: VxLAN tunnel endpoints and

    distributed gateways

    Layer 2: bridge-domain interface (BDIF)

    Layer 3: virtual BDIF (VBDIF)

    Distributed protocol: BGP EVPN

    Policies: ACLs, micro-segmentation (MCS)

    To deploy/update services, users design the VPC

    network, and specify policies through the controller

    Delta configurations are generated, and pushed to

    the fabric

    A

    DCN Fabric

    Delta Configurations

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Problem: All-Pair Reachability Changes

    Page 3

    All-Pair Reachability Matrix

    A

    B C

    Unexpected

    Expected

    All-Pair Reachability Matrix

    A

    B C

    DCN Fabric

    ToR Leaf

    Border Leaf

    & Spine

    PE Router

    Server

    B C D

    Delta Configurations

    A

    Delta

    Changes

    Re-computing full matrix is not scalable

    Incremental: find possibly changed pairs, and re-

    compute their reachability only

    Existing incremental approaches do not address the

    new challenges introduced by network overlay DCNs

    FW,

    LB,

    Before Deployment

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Challenges #I: Model Expressiveness

    Page 4

    Existing BDD-based EC approaches (APV[ICNP’13, APKeep[NSDI’20])

    and TBV-based EC approaches (VeriFlow[NSDI’13], DeltaNet[NSDI’17])

    • Current EC computation does not support tunneling

    • Extensive packet rewrites Performance degradation

    • The TBV model relies on IP rules instead of Group ID rules

    • Conversion Excessive IP rules

    • Invalid: some SRCs may not reach DSTs

    Feature 1: tunneling with overlay/underlay traffic

    Feature 2: packet rewrites for multiple tunnels

    Feature 3: MCS, group definitions are distributed

    A: 1.1.1.0/24, 1.1.2.0/24 Group ID: 10

    B: 1.1.3.0/24 Group ID: 10

    C: 1.1.4.0/24 Group ID: 20

    Policy on groups:

    SRC ID 10, DST ID 20, SRC PORT 80, DENY

    VRF2VRF1

    A B C D

    Cross-VRF static route: two tunnels

    VPC1/VRF1 VPC2/VRF2

    Border Leaf

    ToR Leaf

    Tunnel Header

    Packet Rewrite

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Challenges #2: Completeness of Indexing Methods

    Page 5

    A B

    B1

    D

    Add a static route

    L1 L2 L3 L4

    C

    Indexing method (e.g., TenantGuard [NDSS’17])

    Compute a reachable path

    Associate the visited device with the reachable pair

    B1 {(A,B), (A,C)}

    Any changes on device B1

    Re-compute (A,B) and (A,C)

    Device-level association is coarse-grained and inefficient

    • E.g., E only changes (A,C), no need to re-compute (A,B)

    • Interface-level association may be preferable

    Can not find new reachable pairs

    • E.g., B2 adds a static route for A D, and

    B has no association with (A,D) initially

    EB2

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Our Solutions

    Page 6

    BDD Predicate Model Fine-Grained Indexing Table

    We follow the control-plane verification approach:

    protocol simulation reachability analysis

    1 2

    3

    Waypoint to find relevant (new) reachable pairs

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Port-Predicate Model

    Page 7

    Forwarding ports of devices

    Symbolic Packet & Boolean Formula (Predicate)

    𝑥𝑖

    Outer IP HDs

    GroupID VNI Inner IP HDs

    VxLAN header

    64bits 16bits 24bits 64bits

    ∀𝑥, 𝑥v𝑛𝑖 == 𝑓𝑎𝑙𝑠𝑒

    Underlay Packet

    𝑓 𝑥 = 𝑓(𝑥1, … , 𝑥168)

    i-th bit: 1

    i-th bit: 0 𝑥𝑖

    i-th bit: 0 or 1 𝑥𝑖 ∪ 𝑥𝑖 = 𝑡𝑟𝑢𝑒 universal quantifierOverlay Packet

    else

    Underlay or Overlay: general if

    ∃𝑥, 𝑥v𝑛𝑖 ∩ 𝑥𝑣𝑛𝑖𝑛𝑒𝑤

    existence quantifier == erase

    Rewrite: erase-and-set

    Each IP/Prefix

    𝑔ACL𝐵𝐷𝐼𝐹1 𝑥 ∩ 𝑔𝐹𝐼𝐵(𝑥) ∩ 𝑔ACL

    𝑁𝑉𝐸 𝑥

    Header Space 𝑔 𝑥 = 𝑔(𝑓1 𝑥 , … , 𝑓𝑘 𝑥 )

    All Boolean formulas are represented by BDD

    ifthen

    Predicate quantifiers improve our model expressiveness

    (union, intersection, difference)

    Aggregation of allowed space of ACLs and FIB

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    All-Pair Reachability Matrix

    Page 8

    Start End

    DFS for each starting endpoint

    A

    B

    If1If1

    If2

    If2

    A B

    All-Pair Reachability Matrix

    A

    B C

    Forwarding Graph

    If1=(L1,NVE), If2=(L2,NVE)

    Indexing Table

    EndpointsA

    B

    If1

    If2

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Change-Impact Analysis

    Page 9

    ForwardBackward

    Step 2: Waypoint ComputationStep 1: Changed Interface Detection

    Changed

    Interfaces

    DEL

    Reachable Pairs

    Reachable

    Pairs

    Step 3: Comparison

    MOD ADD

    DEL MOD ADD

    Indexing

    DEL MOD ADD

    Comparison

    Old forwarding graph New forwarding graph

  • APNET’20, August 3-4, 2020

    35pt

    32pt

    ) :18pt

    Preliminary Evaluation

    Page 10

    Implementation

    Based on Batfish*

    Support Huawei CLI/YANG

    BGP EVPN protocol

    JavaBDD

    Synthesized Datasets (similar to real networks)

    Base dataset (20Leafs ~ 100Leafs)

    Inter-VPC traffic

    Intra-VPC (inter-subnet) traffic

    1 Leaf has 20 endpoints (subnets)

    Metric: new graph modeling, detection, waypoint computation and comparison time

    Different service update cases

    Case A/B: ADD SUBNET/ADD VPC

    Case C: ADD cross-VRF Static Routes

    Case D: MOD MCS

    Cases A-C: increase

    is due to modeling,

    detection, comparison

    Case D: increase is

    mostly due to waypoint.

    To be optimized.

    Less than 25s

    for 2k EPs (4 mill. Pairs)

    Case C

    6X

    * https://github.com/batfish/batfish

    2X

  • APNET’20, August 3-4, 2020

    Conclusion

    Page 11

    We are the first to design and demonstrate an incremental configuration verifier

    for network overlay DCNs.

    We design a fast incremental verification algorithm that leverages fine-grained

    indexing and waypoint computation methods to find all-pair reachability changes.

    In the future, we will further explore new forwarding features: policy-based

    routing, firewall zone policies, NAT policies, etc.

  • Copyright©2011 Huawei Technologies Co., Ltd. All Rights Reserved.The information in this document may contain predictive statements including, without limitation, statements regarding thefuture financial and operating results, future product portfolio, new technology, etc. There are a number of factors thatcould cause actual results and developments to differ materially from those expressed or implied in the predictive statements.Therefore, such information is provided for reference purpose only and constitutes neither an offer nor an acceptance. Huaweimay change the information at any time without notice.

    Thank youwww.huawei.com