Endsystem Support for Network Virtualization

1
Endsystem Support for Network Virtualization

• Fred Kuhns

2
Overview

• Context
• Endsystem networking model
• Protocol instances user or kernel space
• pros and cons
• explore user space protocols
• propose kernel level model

3
Context Virtual (Diversified) Networking
substrate link
substrate router
virtual router
virtual link
virtual end-system
4
Simulates Star Topology for Substrate Links

VLANX1
VLANX2
VLANXN

• Internetworking over a diversified network
• Ethernet example
• VLANs are used to provide the equivalent of a
  virtualized wire connecting an endsystem to a
  specific access router.
• All vnets on an endsystem share common VLAN
• Use priority queuing (802.1P/Q) to isolate vnet
  traffic.
• Use admission control (static or dynamic) to
  provide bandwidth guarantees to vnet traffic.
• Substrate layer on endsystems enforce per VLAN
  and per vnet bandwidth constraints

vNetX VR1

• Each host to substrate router connection is
  assigned a distinct VLAN. So N hosts implies N
  VLANs on ethernet.
• Alternative is to define one VLAN tree for each
  protocol suite (i.e. vnet).

5
vnetX traffic uses high priority queues

Ethernet Hub with High and Low Priority TX queues
6
Substrate Link as a VLAN Tree

VLANX

• One VLAN is used for all virtual net traffic
  to/from a substrate router.

7
Multiple Substrate Links

• Three VLANs are used for all virtual net traffic
  to/from a substrate router.
• Corresponds to 3 substrate links
• Low priority default for best-effort traffic
• Medium priority for virtual nets with soft
  performance requirements (average bandwidth)
• High priority for isochronous or low-delay,
  interactive applications

VLANdgram
VLANhigh
VLANmed
8
Multiple vNets per Host

ether addr/vlan
ether addr/vlan
ether addr/vlan
vlan 1
vlan 2
vlan 3

• Substrate link serves to connect an endsystem to
  a substrate router. Virtualization of a physical
  cable or wire. A packet enters one end, exists
  the other and is opaque within. Simplex or
  Duplex?
• Substrate interface (need better term?)
  endsystem abstraction representing a substrate
  link.
• Ethernet ltinterface, VLAN, destgt.
• Could be an IP tunnel
• Not required to be point-to-point.
• Virtual link represents the logical
  interconnection of adjacent network nodes for a
  given protocol suite.
• Point-to-point. Simplex or Duplex?
• Virtual interface endsystem abstraction
  representing one end of a virtual link. Substrate
  defines mechanism for multiplexing onto common
  substrate link. For example a virtual link
  identifier (VLI) in a substrate header. Simplex
  or Duplex?

ethernet LAN
filter on ethernet address and vlan
membership for substrate router
9
Multiple next hop VRs?
Host A on vnetX
vNetX VR2
vNetX VR3
VLANXA2
VLANXA3
ethernet switched LAN

• Not a fundamental part of the model but it is
  consistent with the current model used for TCP/IP
  in endsystem.
• Allows us to implement TCP/IP as a virtual net
  protocol and not change the basic model

VLANXA1
vNetX VR1
10
TCP/IP as an Example Protocol

IP Route Table
vint0 (eth0 VLANX) LL Info SR1 addr VLI
standard ethernet Interface
ethernet device
direct connect
VLANX
ethernet LAN
Substrate Interface Ethernet interface.
Destination address by ARP. Directly connected
destination IP address ARP enet
addr Gateway (Gateways IP ARP enet addr)
VLAN Virtual Interface Directly connected
Not used, model only for internetworking Gateway
VLI assigned by substrate.
ethernet dest. addr
Substrate Router SR1
IP
11
OS Kernel Block Diagram
User Space (Applications)
Socket Interface
ops
AST Processing
callback
routes
IP
SW int (AST)
task management
util
TCP
TC/ AST
qdisc
poll
scheduler
callout Q
hardware independent layer
clock handler
process accounting scheduling time management
Device independent I/O
ethernet
Interrupt Processing
hardware dependent layer
configuration registers, MMU (TLB, cache, VM)
bus and peripherals System Exception handlers
eth0
uart
timer
OS ISR demux
Hardware
HW interrupt/Exception
12
User or kernel Space protocols?

• Each has pros and cons
• User space protocols
• easier to implement and debug
• easier to introduce new protocols (not tightly
  dependent on socket layer knowing about the new
  protocol)
• easier to isolate and protect protocols and apps
  from each other (leverage process model)
• kernel level protocols
• easier to integrate into existing framework
  (simplifies support for system interface
  functions like select/poll)
• simplifies intra-protocol security and protection
  (since protocol runs within trusted kernel)
• simplifies kernel demultiplexing to correct
  protocol context (endpoint)
• increased efficiency

13
User Space Protocol Implementation

• Uncommon outside of high-performance community,
  they want zero-copy and specialized demux keys.
• Problems asynchronous processing, life cycle,
  authentication and demiultiplexing to endpoints
• latency in delivering packets (i.e. acks) to user
  space
• increased overhead in per packet processing
  before a drop/keep decision is made
• processing received acks
• timeouts and retransmissions
• establishing connections and security snooping,
  masquerading
• supporting select and poll
• protocols where connection may outlive process
  (TCPs TIMED_WAIT)
• global routing and address resolution tables
• global connection tables
• need to know what other ports are being used
  (locally)
• accepting/rejecting new connections

14
Assumptions

• Assumptions
• Applications using different VNs (or no VN) will
  need to communicate using the various IPC
  mechanisms
• We want to manage all aspects of Network I/O but
  not the use of other traditional resources
  (memory, files etc)
• CPU, memory and interface bandwidth controlled
  at the virtual net granularity
• intra-VN, implementers should have the mechanisms
  to support QoS and Security
• simple mechanism for adding new protocols/VNs

15
User Space Protocols

• Chandramohan A. Thekkath , Thu D. Nguyen , Evelyn
  Moy , Edward D. Lazowska, Implementing network
  protocols at user level, IEEE/ACM Transactions on
  Networking (TON), v.1 n.5, p.554-565, Oct. 1993
• Chris Maeda, Brian Bershad, Protocol Service
  Decomposition for High-Performance Networking,
  Proceedings of the 14th ACM Symposium on
  Operating Systems Principles. December 1993, pp.
  244-255.
• Aled Edwards , Steve Muir, Experiences
  implementing a high performance TCP in
  user-space, Proceedings of the conference on
  Applications, technologies, architectures, and
  protocols for computer communication, p.196-205,
  1995
• Kieran Mansley, Engineering a User-Level TCP for
  the CLAN Network, Proceedings of the ACM SIGCOMM
  workshop on Network-I/O convergence experience,
  lessons, implications, Pages 228 236, 2003

16
user-space protocols Global Issues

• Routing Direct packets to/from correct
  endpoint/interface
• How is traffic demultiplexed and sent to the
  correct endpoint/process?
• In-kernel filters
• Where are the routing tables and how are they
  maintained?
• route fixed when connection established or
  located in shared memory
• Control I use IPv4 as an example
• Address resolution protocols/tables?
• Other control protocols. For example ICMP, IGRP,
  others?
• Where are the routing protocols implemented?
• Management
• Must manage a protocols namespace (for example,
  port numbers in IPv4).
• Common programming technique, allow protocol
  instance to select local address part
• specify port 0 and addr 0 then implementation
  will assign correct values
• Passive connect model?
• In IPv4 a server listens on a port
  (hostportproto) for a connection request. To
  establish a connection a unique (to the
  endsystem) port number is assigned and new socket
  allocated.
• socket-oriented system calls must be supported.
  On UNIX must support non-blocking I/O with select
  and poll.
• Connection lifetime may outlast process.
• For example TCP TIME_WAIT or simply waiting for a
  final ack or resending if no ack received.
• Security we must provide sufficient mechanisms
  for protocol developers

17
User Space Configurations

• Given these global issues there are two likely
  configurations
• all traffic passes through common protocol daemon
  in user space
• control daemon implements basic set of control
  functions while user library implements majority
  of data path functions
• prior work has shown the latter approach to be
  superior.
• Having all traffic pass through a common protocol
  daemon gt at least one extra copy operation
  (kernel -gt daemon -gt user process)
• A better solution is for a daemon to insert
  relatively simple packet filters in kernel for
  established connections which directs packets
  to/filters packets from endpoints.

18
User-Space Passive Open
0. listen/accept (passive open)
vnetX control daemon (namespace, lifecycle,
connections)
4. new connection
data copy
socket layer
3. insert incoming and outgoing filters for vnetX
connection
1. connection request (in)
5. data, established connections
compare against connection specific outgoing
filter
2. ack (out)
vnet demux
connection filters
use VLI to access incoming filters and use to
demux to filter set and/or socket.
ethernet
19
User-Space Active Open
0. connect
vnetX control daemon (namespace, lifecycle,
connections)
4. new connection
data copy
socket layer
1. connection request (out)
3. insert incoming and outgoing filters for vnetX
connection
5. data, established connections
compare against connection specific outgoing
filter
2. ack (in)
vnet demux
connection filters
use VLI to access incoming filters and use to
demux to filter set and/or socket.
ethernet
20
User-Space Datagram (Connectionless)
daemon fills in local address and binds to
socket. No restrictions on destination
0. open(any)
vnetX control daemon (namespace, lifecycle,
connections)
data copy
2. new connection (local address)
socket layer
1. insert incoming and outgoing filters for vnetX
connection
3. data established connections
compare against connection specific outgoing
filter
vnet demux
connection filters
use VLI to access incoming filters and use to
demux to socket. In this case only the local part
is used.
ethernet
21
User-Space Datagram (Connectionless)
daemon fills in both local and destination
addresses. Destination restricted
0. open(local and remote addr)
vnetX control daemon (namespace, lifecycle,
connections)
2. new connection(local and remote)
data copy
socket layer
1. insert incoming and outgoing filters for vnetX
connection
3. data established connections
compare against connection specific outgoing
filter
vnet demux
connection filters
ethernet
use VLI to access incoming filters and use to
demux to socket.
22
User-Space App exits
TCP enters TIME_WAIT after close
vnetX control daemon (namespace, lifecycle,
connections)
socket layer
3. remove filters
1. connection close (out)
2. ack (in/out)
vnet demux
connection filters
ethernet
drop
23
Extensible protocol frameworks in the kernel

• Herbert Bos, Bart Samwel, Safe Kernel Programming
  in the OKE, Proceedings of the fifth IEEE
  Conference on Open Architectures and Network
  Programming, June 2002

24
OKE

• Context For performance reasons it is useful to
  permit third parties to load optimized modules
  into the kernel
• Problem Third party code is untrusted so loading
  into kernel will compromise system security and
  reliability. Could use safe execution environment
  like java but incurs expensive runtime checks.
• Solution create set of mechanisms and policies
  to permit non-root users to safely load untrusted
  application modules into kernel space with
  minimal impact on runtime performance.
• Safety use a trusted compile to enforce policies
  (constraints). The constraints are designed to
  ensure the untrusted module will not adversely
  affect the kernel (core and loadable modules) or
  unrelated processes.
• User privileges Vary enforced constraints based
  on user privileges (customizable language)
• Termination well defined termination boundaries
  to protect system state
• Enforcement Static and dynamic checks language
  extensions
• Ease of use Familiar development environment
  using Cyclone (type safe, C extension) and kernel
  module.
• Contribution definition of safe kernel
  programming environment that meets competing
  needs
• performance
• safety
• ease of use
• hosted in a commodity OS

25
Considerations

• Identified areas where modules may impact system
  behavior
• program correctness language restrictions for
  safety and enforce coding conventions
• Memory access static and dynamic enforcement of
  memory access rules
• Kernel module access static and dynamic
  enforcement of kernel module (interface) access
  restrictions
• Resource usage Bounded (deterministic or
  limited)

26
Pushing protocols into the Kernel

• Positives
• All the issues associated with user-space
  protocol simply go away. Global tables and
  lifetime of the kernel
• Performance, efficiency, existing code base
• Enhances intra-Protocol security
• Simplifies integration with existing network I/O
  subsystems and interfaces
• Negatives
• Isolation More difficult to isolate system from
  protocol instances. Inter-protocol isolation
  difficult.
• Security Proving trust/security more difficult
• Implementation and debugging more difficult in
  kernel

27
Kernel-Space Protocols
Rework!
Application(s)
/dev/protoX /dev/vnet
User Space (Applications)

udpport
tcpport
rawIP
vnetep
vnetep

TCP
vnet
UDP
RAW IP
TCPn
TCP2
TCP1

TCP/IP

IP
route to interface
routes
SW Interrupt
HW Interrupt
Hardware
HW interrupt/Exception